Whamcloud - gitweb
LU-9679 osc: centralize handling of PTLRPCD_SET
[fs/lustre-release.git] / lustre / ptlrpc / client.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) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  */
32
33 /** Implementation of client-side PortalRPC interfaces */
34
35 #define DEBUG_SUBSYSTEM S_RPC
36
37 #include <linux/delay.h>
38 #include <linux/random.h>
39
40 #include <lnet/lib-lnet.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
47
48 #include "ptlrpc_internal.h"
49
50 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
51                                       struct page *page, int pageoffset,
52                                       int len)
53 {
54         __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
55 }
56
57 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
58                                         struct page *page, int pageoffset,
59                                         int len)
60 {
61         __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
62 }
63
64 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
65 {
66         int i;
67
68         for (i = 0; i < desc->bd_iov_count ; i++)
69                 put_page(BD_GET_KIOV(desc, i).kiov_page);
70 }
71
72 static int ptlrpc_prep_bulk_frag_pages(struct ptlrpc_bulk_desc *desc,
73                                        void *frag, int len)
74 {
75         unsigned int offset = (unsigned long)frag & ~PAGE_MASK;
76
77         ENTRY;
78         while (len > 0) {
79                 int page_len = min_t(unsigned int, PAGE_SIZE - offset,
80                                      len);
81                 unsigned long vaddr = (unsigned long)frag;
82
83                 ptlrpc_prep_bulk_page_nopin(desc,
84                                             lnet_kvaddr_to_page(vaddr),
85                                             offset, page_len);
86                 offset = 0;
87                 len -= page_len;
88                 frag += page_len;
89         }
90
91         RETURN(desc->bd_nob);
92 }
93
94 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
95         .add_kiov_frag  = ptlrpc_prep_bulk_page_pin,
96         .release_frags  = ptlrpc_release_bulk_page_pin,
97 };
98 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
99
100 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
101         .add_kiov_frag  = ptlrpc_prep_bulk_page_nopin,
102         .release_frags  = ptlrpc_release_bulk_noop,
103         .add_iov_frag   = ptlrpc_prep_bulk_frag_pages,
104 };
105 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
106
107 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
108         .add_iov_frag = ptlrpc_prep_bulk_frag,
109 };
110 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
111
112 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
113 static int ptlrpcd_check_work(struct ptlrpc_request *req);
114 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
115
116 /**
117  * Initialize passed in client structure \a cl.
118  */
119 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
120                         struct ptlrpc_client *cl)
121 {
122         cl->cli_request_portal = req_portal;
123         cl->cli_reply_portal   = rep_portal;
124         cl->cli_name           = name;
125 }
126 EXPORT_SYMBOL(ptlrpc_init_client);
127
128 /**
129  * Return PortalRPC connection for remore uud \a uuid
130  */
131 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
132                                                     lnet_nid_t nid4refnet)
133 {
134         struct ptlrpc_connection *c;
135         lnet_nid_t self;
136         struct lnet_process_id peer;
137         int err;
138
139         /*
140          * ptlrpc_uuid_to_peer() initializes its 2nd parameter
141          * before accessing its values.
142          */
143         /* coverity[uninit_use_in_call] */
144         peer.nid = nid4refnet;
145         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
146         if (err != 0) {
147                 CNETERR("cannot find peer %s!\n", uuid->uuid);
148                 return NULL;
149         }
150
151         c = ptlrpc_connection_get(peer, self, uuid);
152         if (c) {
153                 memcpy(c->c_remote_uuid.uuid,
154                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
155         }
156
157         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
158
159         return c;
160 }
161
162 /**
163  * Allocate and initialize new bulk descriptor on the sender.
164  * Returns pointer to the descriptor or NULL on error.
165  */
166 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
167                                          unsigned int max_brw,
168                                          enum ptlrpc_bulk_op_type type,
169                                          unsigned int portal,
170                                          const struct ptlrpc_bulk_frag_ops *ops)
171 {
172         struct ptlrpc_bulk_desc *desc;
173         int i;
174
175         /* ensure that only one of KIOV or IOVEC is set but not both */
176         LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
177                  ops->add_kiov_frag != NULL) ||
178                 (ptlrpc_is_bulk_desc_kvec(type) &&
179                  ops->add_iov_frag != NULL));
180
181         OBD_ALLOC_PTR(desc);
182         if (!desc)
183                 return NULL;
184         if (type & PTLRPC_BULK_BUF_KIOV) {
185                 OBD_ALLOC_LARGE(GET_KIOV(desc),
186                                 nfrags * sizeof(*GET_KIOV(desc)));
187                 if (!GET_KIOV(desc))
188                         goto out;
189         } else {
190                 OBD_ALLOC_LARGE(GET_KVEC(desc),
191                                 nfrags * sizeof(*GET_KVEC(desc)));
192                 if (!GET_KVEC(desc))
193                         goto out;
194         }
195
196         spin_lock_init(&desc->bd_lock);
197         init_waitqueue_head(&desc->bd_waitq);
198         desc->bd_max_iov = nfrags;
199         desc->bd_iov_count = 0;
200         desc->bd_portal = portal;
201         desc->bd_type = type;
202         desc->bd_md_count = 0;
203         desc->bd_frag_ops = ops;
204         LASSERT(max_brw > 0);
205         desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
206         /*
207          * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
208          * node. Negotiated ocd_brw_size will always be <= this number.
209          */
210         for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
211                 LNetInvalidateMDHandle(&desc->bd_mds[i]);
212
213         return desc;
214 out:
215         OBD_FREE_PTR(desc);
216         return NULL;
217 }
218
219 /**
220  * Prepare bulk descriptor for specified outgoing request \a req that
221  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
222  * the bulk to be sent. Used on client-side.
223  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
224  * error.
225  */
226 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
227                                               unsigned int nfrags,
228                                               unsigned int max_brw,
229                                               unsigned int type,
230                                               unsigned int portal,
231                                               const struct ptlrpc_bulk_frag_ops
232                                                 *ops)
233 {
234         struct obd_import *imp = req->rq_import;
235         struct ptlrpc_bulk_desc *desc;
236
237         ENTRY;
238         LASSERT(ptlrpc_is_bulk_op_passive(type));
239
240         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
241         if (!desc)
242                 RETURN(NULL);
243
244         desc->bd_import = class_import_get(imp);
245         desc->bd_req = req;
246
247         desc->bd_cbid.cbid_fn  = client_bulk_callback;
248         desc->bd_cbid.cbid_arg = desc;
249
250         /* This makes req own desc, and free it when she frees herself */
251         req->rq_bulk = desc;
252
253         return desc;
254 }
255 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
256
257 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
258                              struct page *page, int pageoffset, int len,
259                              int pin)
260 {
261         lnet_kiov_t *kiov;
262
263         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
264         LASSERT(page != NULL);
265         LASSERT(pageoffset >= 0);
266         LASSERT(len > 0);
267         LASSERT(pageoffset + len <= PAGE_SIZE);
268         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
269
270         kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
271
272         desc->bd_nob += len;
273
274         if (pin)
275                 get_page(page);
276
277         kiov->kiov_page = page;
278         kiov->kiov_offset = pageoffset;
279         kiov->kiov_len = len;
280
281         desc->bd_iov_count++;
282 }
283 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
284
285 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
286                           void *frag, int len)
287 {
288         struct kvec *iovec;
289
290         ENTRY;
291
292         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
293         LASSERT(frag != NULL);
294         LASSERT(len > 0);
295         LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
296
297         iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
298
299         desc->bd_nob += len;
300
301         iovec->iov_base = frag;
302         iovec->iov_len = len;
303
304         desc->bd_iov_count++;
305
306         RETURN(desc->bd_nob);
307 }
308 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
309
310 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
311 {
312         ENTRY;
313
314         LASSERT(desc != NULL);
315         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
316         LASSERT(desc->bd_md_count == 0);         /* network hands off */
317         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
318         LASSERT(desc->bd_frag_ops != NULL);
319
320         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
321                 sptlrpc_enc_pool_put_pages(desc);
322
323         if (desc->bd_export)
324                 class_export_put(desc->bd_export);
325         else
326                 class_import_put(desc->bd_import);
327
328         if (desc->bd_frag_ops->release_frags != NULL)
329                 desc->bd_frag_ops->release_frags(desc);
330
331         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
332                 OBD_FREE_LARGE(GET_KIOV(desc),
333                                desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
334         else
335                 OBD_FREE_LARGE(GET_KVEC(desc),
336                                desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
337         OBD_FREE_PTR(desc);
338         EXIT;
339 }
340 EXPORT_SYMBOL(ptlrpc_free_bulk);
341
342 /**
343  * Set server timelimit for this req, i.e. how long are we willing to wait
344  * for reply before timing out this request.
345  */
346 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
347 {
348         __u32 serv_est;
349         int idx;
350         struct imp_at *at;
351
352         LASSERT(req->rq_import);
353
354         if (AT_OFF) {
355                 /* non-AT settings */
356                 /**
357                  * \a imp_server_timeout means this is reverse import and
358                  * we send (currently only) ASTs to the client and cannot afford
359                  * to wait too long for the reply, otherwise the other client
360                  * (because of which we are sending this request) would
361                  * timeout waiting for us
362                  */
363                 req->rq_timeout = req->rq_import->imp_server_timeout ?
364                                   obd_timeout / 2 : obd_timeout;
365         } else {
366                 at = &req->rq_import->imp_at;
367                 idx = import_at_get_index(req->rq_import,
368                                           req->rq_request_portal);
369                 serv_est = at_get(&at->iat_service_estimate[idx]);
370                 req->rq_timeout = at_est2timeout(serv_est);
371         }
372         /*
373          * We could get even fancier here, using history to predict increased
374          * loading...
375          */
376
377         /*
378          * Let the server know what this RPC timeout is by putting it in the
379          * reqmsg
380          */
381         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
382 }
383 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
384
385 /* Adjust max service estimate based on server value */
386 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
387                                   unsigned int serv_est)
388 {
389         int idx;
390         unsigned int oldse;
391         struct imp_at *at;
392
393         LASSERT(req->rq_import);
394         at = &req->rq_import->imp_at;
395
396         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
397         /*
398          * max service estimates are tracked on the server side,
399          * so just keep minimal history here
400          */
401         oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
402         if (oldse != 0)
403                 CDEBUG(D_ADAPTTO,
404                        "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
405                        req->rq_import->imp_obd->obd_name,
406                        req->rq_request_portal,
407                        oldse, at_get(&at->iat_service_estimate[idx]));
408 }
409
410 /* Expected network latency per remote node (secs) */
411 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
412 {
413         return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
414 }
415
416 /* Adjust expected network latency */
417 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
418                                unsigned int service_time)
419 {
420         unsigned int nl, oldnl;
421         struct imp_at *at;
422         time64_t now = ktime_get_real_seconds();
423
424         LASSERT(req->rq_import);
425
426         if (service_time > now - req->rq_sent + 3) {
427                 /*
428                  * b=16408, however, this can also happen if early reply
429                  * is lost and client RPC is expired and resent, early reply
430                  * or reply of original RPC can still be fit in reply buffer
431                  * of resent RPC, now client is measuring time from the
432                  * resent time, but server sent back service time of original
433                  * RPC.
434                  */
435                 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
436                        D_ADAPTTO : D_WARNING,
437                        "Reported service time %u > total measured time %lld\n",
438                        service_time, now - req->rq_sent);
439                 return;
440         }
441
442         /* Network latency is total time less server processing time */
443         nl = max_t(int, now - req->rq_sent -
444                         service_time, 0) + 1; /* st rounding */
445         at = &req->rq_import->imp_at;
446
447         oldnl = at_measured(&at->iat_net_latency, nl);
448         if (oldnl != 0)
449                 CDEBUG(D_ADAPTTO,
450                        "The network latency for %s (nid %s) has changed from %d to %d\n",
451                        req->rq_import->imp_obd->obd_name,
452                        obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
453                        oldnl, at_get(&at->iat_net_latency));
454 }
455
456 static int unpack_reply(struct ptlrpc_request *req)
457 {
458         int rc;
459
460         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
461                 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
462                 if (rc) {
463                         DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
464                                   rc);
465                         return -EPROTO;
466                 }
467         }
468
469         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
470         if (rc) {
471                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
472                           rc);
473                 return -EPROTO;
474         }
475         return 0;
476 }
477
478 /**
479  * Handle an early reply message, called with the rq_lock held.
480  * If anything goes wrong just ignore it - same as if it never happened
481  */
482 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
483 __must_hold(&req->rq_lock)
484 {
485         struct ptlrpc_request *early_req;
486         time64_t olddl;
487         int rc;
488
489         ENTRY;
490         req->rq_early = 0;
491         spin_unlock(&req->rq_lock);
492
493         rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
494         if (rc) {
495                 spin_lock(&req->rq_lock);
496                 RETURN(rc);
497         }
498
499         rc = unpack_reply(early_req);
500         if (rc != 0) {
501                 sptlrpc_cli_finish_early_reply(early_req);
502                 spin_lock(&req->rq_lock);
503                 RETURN(rc);
504         }
505
506         /*
507          * Use new timeout value just to adjust the local value for this
508          * request, don't include it into at_history. It is unclear yet why
509          * service time increased and should it be counted or skipped, e.g.
510          * that can be recovery case or some error or server, the real reply
511          * will add all new data if it is worth to add.
512          */
513         req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
514         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
515
516         /* Network latency can be adjusted, it is pure network delays */
517         ptlrpc_at_adj_net_latency(req,
518                                   lustre_msg_get_service_time(early_req->rq_repmsg));
519
520         sptlrpc_cli_finish_early_reply(early_req);
521
522         spin_lock(&req->rq_lock);
523         olddl = req->rq_deadline;
524         /*
525          * server assumes it now has rq_timeout from when the request
526          * arrived, so the client should give it at least that long.
527          * since we don't know the arrival time we'll use the original
528          * sent time
529          */
530         req->rq_deadline = req->rq_sent + req->rq_timeout +
531                            ptlrpc_at_get_net_latency(req);
532
533         /* The below message is checked in replay-single.sh test_65{a,b} */
534         /* The below message is checked in sanity-{gss,krb5} test_8 */
535         DEBUG_REQ(D_ADAPTTO, req,
536                   "Early reply #%d, new deadline in %llds (%llds)",
537                   req->rq_early_count,
538                   req->rq_deadline - ktime_get_real_seconds(),
539                   req->rq_deadline - olddl);
540
541         RETURN(rc);
542 }
543
544 static struct kmem_cache *request_cache;
545
546 int ptlrpc_request_cache_init(void)
547 {
548         request_cache = kmem_cache_create("ptlrpc_cache",
549                                           sizeof(struct ptlrpc_request),
550                                           0, SLAB_HWCACHE_ALIGN, NULL);
551         return request_cache ? 0 : -ENOMEM;
552 }
553
554 void ptlrpc_request_cache_fini(void)
555 {
556         kmem_cache_destroy(request_cache);
557 }
558
559 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
560 {
561         struct ptlrpc_request *req;
562
563         OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
564         return req;
565 }
566
567 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
568 {
569         OBD_SLAB_FREE_PTR(req, request_cache);
570 }
571
572 /**
573  * Wind down request pool \a pool.
574  * Frees all requests from the pool too
575  */
576 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
577 {
578         struct list_head *l, *tmp;
579         struct ptlrpc_request *req;
580
581         LASSERT(pool != NULL);
582
583         spin_lock(&pool->prp_lock);
584         list_for_each_safe(l, tmp, &pool->prp_req_list) {
585                 req = list_entry(l, struct ptlrpc_request, rq_list);
586                 list_del(&req->rq_list);
587                 LASSERT(req->rq_reqbuf);
588                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
589                 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
590                 ptlrpc_request_cache_free(req);
591         }
592         spin_unlock(&pool->prp_lock);
593         OBD_FREE(pool, sizeof(*pool));
594 }
595 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
596
597 /**
598  * Allocates, initializes and adds \a num_rq requests to the pool \a pool
599  */
600 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
601 {
602         int i;
603         int size = 1;
604
605         while (size < pool->prp_rq_size)
606                 size <<= 1;
607
608         LASSERTF(list_empty(&pool->prp_req_list) ||
609                  size == pool->prp_rq_size,
610                  "Trying to change pool size with nonempty pool from %d to %d bytes\n",
611                  pool->prp_rq_size, size);
612
613         pool->prp_rq_size = size;
614         for (i = 0; i < num_rq; i++) {
615                 struct ptlrpc_request *req;
616                 struct lustre_msg *msg;
617
618                 req = ptlrpc_request_cache_alloc(GFP_NOFS);
619                 if (!req)
620                         return i;
621                 OBD_ALLOC_LARGE(msg, size);
622                 if (!msg) {
623                         ptlrpc_request_cache_free(req);
624                         return i;
625                 }
626                 req->rq_reqbuf = msg;
627                 req->rq_reqbuf_len = size;
628                 req->rq_pool = pool;
629                 spin_lock(&pool->prp_lock);
630                 list_add_tail(&req->rq_list, &pool->prp_req_list);
631                 spin_unlock(&pool->prp_lock);
632         }
633         return num_rq;
634 }
635 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
636
637 /**
638  * Create and initialize new request pool with given attributes:
639  * \a num_rq - initial number of requests to create for the pool
640  * \a msgsize - maximum message size possible for requests in thid pool
641  * \a populate_pool - function to be called when more requests need to be added
642  *                    to the pool
643  * Returns pointer to newly created pool or NULL on error.
644  */
645 struct ptlrpc_request_pool *
646 ptlrpc_init_rq_pool(int num_rq, int msgsize,
647                     int (*populate_pool)(struct ptlrpc_request_pool *, int))
648 {
649         struct ptlrpc_request_pool *pool;
650
651         OBD_ALLOC_PTR(pool);
652         if (!pool)
653                 return NULL;
654
655         /*
656          * Request next power of two for the allocation, because internally
657          * kernel would do exactly this
658          */
659         spin_lock_init(&pool->prp_lock);
660         INIT_LIST_HEAD(&pool->prp_req_list);
661         pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
662         pool->prp_populate = populate_pool;
663
664         populate_pool(pool, num_rq);
665
666         return pool;
667 }
668 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
669
670 /**
671  * Fetches one request from pool \a pool
672  */
673 static struct ptlrpc_request *
674 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
675 {
676         struct ptlrpc_request *request;
677         struct lustre_msg *reqbuf;
678
679         if (!pool)
680                 return NULL;
681
682         spin_lock(&pool->prp_lock);
683
684         /*
685          * See if we have anything in a pool, and bail out if nothing,
686          * in writeout path, where this matters, this is safe to do, because
687          * nothing is lost in this case, and when some in-flight requests
688          * complete, this code will be called again.
689          */
690         if (unlikely(list_empty(&pool->prp_req_list))) {
691                 spin_unlock(&pool->prp_lock);
692                 return NULL;
693         }
694
695         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
696                              rq_list);
697         list_del_init(&request->rq_list);
698         spin_unlock(&pool->prp_lock);
699
700         LASSERT(request->rq_reqbuf);
701         LASSERT(request->rq_pool);
702
703         reqbuf = request->rq_reqbuf;
704         memset(request, 0, sizeof(*request));
705         request->rq_reqbuf = reqbuf;
706         request->rq_reqbuf_len = pool->prp_rq_size;
707         request->rq_pool = pool;
708
709         return request;
710 }
711
712 /**
713  * Returns freed \a request to pool.
714  */
715 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
716 {
717         struct ptlrpc_request_pool *pool = request->rq_pool;
718
719         spin_lock(&pool->prp_lock);
720         LASSERT(list_empty(&request->rq_list));
721         LASSERT(!request->rq_receiving_reply);
722         list_add_tail(&request->rq_list, &pool->prp_req_list);
723         spin_unlock(&pool->prp_lock);
724 }
725
726 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
727 {
728         struct obd_import *imp = req->rq_import;
729         struct list_head *tmp;
730         struct ptlrpc_request *iter;
731
732         assert_spin_locked(&imp->imp_lock);
733         LASSERT(list_empty(&req->rq_unreplied_list));
734
735         /* unreplied list is sorted by xid in ascending order */
736         list_for_each_prev(tmp, &imp->imp_unreplied_list) {
737                 iter = list_entry(tmp, struct ptlrpc_request,
738                                   rq_unreplied_list);
739
740                 LASSERT(req->rq_xid != iter->rq_xid);
741                 if (req->rq_xid < iter->rq_xid)
742                         continue;
743                 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
744                 return;
745         }
746         list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
747 }
748
749 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
750 {
751         req->rq_xid = ptlrpc_next_xid();
752         ptlrpc_add_unreplied(req);
753 }
754
755 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
756 {
757         spin_lock(&req->rq_import->imp_lock);
758         ptlrpc_assign_next_xid_nolock(req);
759         spin_unlock(&req->rq_import->imp_lock);
760 }
761
762 static atomic64_t ptlrpc_last_xid;
763
764 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
765 {
766         spin_lock(&req->rq_import->imp_lock);
767         list_del_init(&req->rq_unreplied_list);
768         ptlrpc_assign_next_xid_nolock(req);
769         spin_unlock(&req->rq_import->imp_lock);
770         DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
771 }
772
773 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
774 {
775         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
776         __u32 opc;
777         __u16 tag;
778
779         opc = lustre_msg_get_opc(req->rq_reqmsg);
780         tag = obd_get_mod_rpc_slot(cli, opc);
781         lustre_msg_set_tag(req->rq_reqmsg, tag);
782         ptlrpc_reassign_next_xid(req);
783 }
784 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
785
786 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
787 {
788         __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
789
790         if (tag != 0) {
791                 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
792                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
793
794                 obd_put_mod_rpc_slot(cli, opc, tag);
795         }
796 }
797 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
798
799 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
800                              __u32 version, int opcode, char **bufs,
801                              struct ptlrpc_cli_ctx *ctx)
802 {
803         int count;
804         struct obd_import *imp;
805         __u32 *lengths;
806         int rc;
807
808         ENTRY;
809
810         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
811         imp = request->rq_import;
812         lengths = request->rq_pill.rc_area[RCL_CLIENT];
813
814         if (ctx) {
815                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
816         } else {
817                 rc = sptlrpc_req_get_ctx(request);
818                 if (rc)
819                         GOTO(out_free, rc);
820         }
821         sptlrpc_req_set_flavor(request, opcode);
822
823         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
824                                  lengths, bufs);
825         if (rc)
826                 GOTO(out_ctx, rc);
827
828         lustre_msg_add_version(request->rq_reqmsg, version);
829         request->rq_send_state = LUSTRE_IMP_FULL;
830         request->rq_type = PTL_RPC_MSG_REQUEST;
831
832         request->rq_req_cbid.cbid_fn  = request_out_callback;
833         request->rq_req_cbid.cbid_arg = request;
834
835         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
836         request->rq_reply_cbid.cbid_arg = request;
837
838         request->rq_reply_deadline = 0;
839         request->rq_bulk_deadline = 0;
840         request->rq_req_deadline = 0;
841         request->rq_phase = RQ_PHASE_NEW;
842         request->rq_next_phase = RQ_PHASE_UNDEFINED;
843
844         request->rq_request_portal = imp->imp_client->cli_request_portal;
845         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
846
847         ptlrpc_at_set_req_timeout(request);
848
849         lustre_msg_set_opc(request->rq_reqmsg, opcode);
850
851         /* Let's setup deadline for req/reply/bulk unlink for opcode. */
852         if (cfs_fail_val == opcode) {
853                 time64_t *fail_t = NULL, *fail2_t = NULL;
854
855                 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
856                         fail_t = &request->rq_bulk_deadline;
857                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
858                         fail_t = &request->rq_reply_deadline;
859                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
860                         fail_t = &request->rq_req_deadline;
861                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
862                         fail_t = &request->rq_reply_deadline;
863                         fail2_t = &request->rq_bulk_deadline;
864                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
865                         time64_t now = ktime_get_real_seconds();
866                         u64 xid = ((u64)now >> 4) << 24;
867
868                         atomic64_set(&ptlrpc_last_xid, xid);
869                 }
870
871                 if (fail_t) {
872                         *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
873
874                         if (fail2_t)
875                                 *fail2_t = ktime_get_real_seconds() +
876                                            LONG_UNLINK;
877
878                         /*
879                          * The RPC is infected, let the test to change the
880                          * fail_loc
881                          */
882                         msleep(4 * MSEC_PER_SEC);
883                 }
884         }
885         ptlrpc_assign_next_xid(request);
886
887         RETURN(0);
888
889 out_ctx:
890         LASSERT(!request->rq_pool);
891         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
892 out_free:
893         class_import_put(imp);
894
895         return rc;
896 }
897 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
898
899 /**
900  * Pack request buffers for network transfer, performing necessary encryption
901  * steps if necessary.
902  */
903 int ptlrpc_request_pack(struct ptlrpc_request *request,
904                         __u32 version, int opcode)
905 {
906         return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
907 }
908 EXPORT_SYMBOL(ptlrpc_request_pack);
909
910 /**
911  * Helper function to allocate new request on import \a imp
912  * and possibly using existing request from pool \a pool if provided.
913  * Returns allocated request structure with import field filled or
914  * NULL on error.
915  */
916 static inline
917 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
918                                               struct ptlrpc_request_pool *pool)
919 {
920         struct ptlrpc_request *request = NULL;
921
922         request = ptlrpc_request_cache_alloc(GFP_NOFS);
923
924         if (!request && pool)
925                 request = ptlrpc_prep_req_from_pool(pool);
926
927         if (request) {
928                 ptlrpc_cli_req_init(request);
929
930                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
931                 LASSERT(imp != LP_POISON);
932                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
933                          imp->imp_client);
934                 LASSERT(imp->imp_client != LP_POISON);
935
936                 request->rq_import = class_import_get(imp);
937         } else {
938                 CERROR("request allocation out of memory\n");
939         }
940
941         return request;
942 }
943
944 /**
945  * Helper function for creating a request.
946  * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
947  * buffer structures according to capsule template \a format.
948  * Returns allocated request structure pointer or NULL on error.
949  */
950 static struct ptlrpc_request *
951 ptlrpc_request_alloc_internal(struct obd_import *imp,
952                               struct ptlrpc_request_pool *pool,
953                               const struct req_format *format)
954 {
955         struct ptlrpc_request *request;
956
957         request = __ptlrpc_request_alloc(imp, pool);
958         if (!request)
959                 return NULL;
960
961         /*
962          * initiate connection if needed when the import has been
963          * referenced by the new request to avoid races with disconnect
964          */
965         if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
966                 int rc;
967
968                 CDEBUG_LIMIT(imp->imp_idle_debug,
969                              "%s: reconnect after %llds idle\n",
970                              imp->imp_obd->obd_name, ktime_get_real_seconds() -
971                                                      imp->imp_last_reply_time);
972                 spin_lock(&imp->imp_lock);
973                 if (imp->imp_state == LUSTRE_IMP_IDLE) {
974                         imp->imp_generation++;
975                         imp->imp_initiated_at = imp->imp_generation;
976                         imp->imp_state = LUSTRE_IMP_NEW;
977
978                         /* connect_import_locked releases imp_lock */
979                         rc = ptlrpc_connect_import_locked(imp);
980                         if (rc < 0) {
981                                 ptlrpc_request_free(request);
982                                 return NULL;
983                         }
984                         ptlrpc_pinger_add_import(imp);
985                 } else {
986                         spin_unlock(&imp->imp_lock);
987                 }
988         }
989
990         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
991         req_capsule_set(&request->rq_pill, format);
992         return request;
993 }
994
995 /**
996  * Allocate new request structure for import \a imp and initialize its
997  * buffer structure according to capsule template \a format.
998  */
999 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
1000                                             const struct req_format *format)
1001 {
1002         return ptlrpc_request_alloc_internal(imp, NULL, format);
1003 }
1004 EXPORT_SYMBOL(ptlrpc_request_alloc);
1005
1006 /**
1007  * Allocate new request structure for import \a imp from pool \a pool and
1008  * initialize its buffer structure according to capsule template \a format.
1009  */
1010 struct ptlrpc_request *
1011 ptlrpc_request_alloc_pool(struct obd_import *imp,
1012                           struct ptlrpc_request_pool *pool,
1013                           const struct req_format *format)
1014 {
1015         return ptlrpc_request_alloc_internal(imp, pool, format);
1016 }
1017 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
1018
1019 /**
1020  * For requests not from pool, free memory of the request structure.
1021  * For requests obtained from a pool earlier, return request back to pool.
1022  */
1023 void ptlrpc_request_free(struct ptlrpc_request *request)
1024 {
1025         if (request->rq_pool)
1026                 __ptlrpc_free_req_to_pool(request);
1027         else
1028                 ptlrpc_request_cache_free(request);
1029 }
1030 EXPORT_SYMBOL(ptlrpc_request_free);
1031
1032 /**
1033  * Allocate new request for operatione \a opcode and immediatelly pack it for
1034  * network transfer.
1035  * Only used for simple requests like OBD_PING where the only important
1036  * part of the request is operation itself.
1037  * Returns allocated request or NULL on error.
1038  */
1039 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1040                                                  const struct req_format *format,
1041                                                  __u32 version, int opcode)
1042 {
1043         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1044         int rc;
1045
1046         if (req) {
1047                 rc = ptlrpc_request_pack(req, version, opcode);
1048                 if (rc) {
1049                         ptlrpc_request_free(req);
1050                         req = NULL;
1051                 }
1052         }
1053         return req;
1054 }
1055 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1056
1057 /**
1058  * Allocate and initialize new request set structure on the current CPT.
1059  * Returns a pointer to the newly allocated set structure or NULL on error.
1060  */
1061 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1062 {
1063         struct ptlrpc_request_set *set;
1064         int cpt;
1065
1066         ENTRY;
1067         cpt = cfs_cpt_current(cfs_cpt_tab, 0);
1068         OBD_CPT_ALLOC(set, cfs_cpt_tab, cpt, sizeof(*set));
1069         if (!set)
1070                 RETURN(NULL);
1071         atomic_set(&set->set_refcount, 1);
1072         INIT_LIST_HEAD(&set->set_requests);
1073         init_waitqueue_head(&set->set_waitq);
1074         atomic_set(&set->set_new_count, 0);
1075         atomic_set(&set->set_remaining, 0);
1076         spin_lock_init(&set->set_new_req_lock);
1077         INIT_LIST_HEAD(&set->set_new_requests);
1078         set->set_max_inflight = UINT_MAX;
1079         set->set_producer     = NULL;
1080         set->set_producer_arg = NULL;
1081         set->set_rc           = 0;
1082
1083         RETURN(set);
1084 }
1085 EXPORT_SYMBOL(ptlrpc_prep_set);
1086
1087 /**
1088  * Allocate and initialize new request set structure with flow control
1089  * extension. This extension allows to control the number of requests in-flight
1090  * for the whole set. A callback function to generate requests must be provided
1091  * and the request set will keep the number of requests sent over the wire to
1092  * @max_inflight.
1093  * Returns a pointer to the newly allocated set structure or NULL on error.
1094  */
1095 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1096                                              void *arg)
1097
1098 {
1099         struct ptlrpc_request_set *set;
1100
1101         set = ptlrpc_prep_set();
1102         if (!set)
1103                 RETURN(NULL);
1104
1105         set->set_max_inflight  = max;
1106         set->set_producer      = func;
1107         set->set_producer_arg  = arg;
1108
1109         RETURN(set);
1110 }
1111
1112 /**
1113  * Wind down and free request set structure previously allocated with
1114  * ptlrpc_prep_set.
1115  * Ensures that all requests on the set have completed and removes
1116  * all requests from the request list in a set.
1117  * If any unsent request happen to be on the list, pretends that they got
1118  * an error in flight and calls their completion handler.
1119  */
1120 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1121 {
1122         struct list_head *tmp;
1123         struct list_head *next;
1124         int expected_phase;
1125         int n = 0;
1126
1127         ENTRY;
1128
1129         /* Requests on the set should either all be completed, or all be new */
1130         expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1131                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1132         list_for_each(tmp, &set->set_requests) {
1133                 struct ptlrpc_request *req =
1134                         list_entry(tmp, struct ptlrpc_request,
1135                                    rq_set_chain);
1136
1137                 LASSERT(req->rq_phase == expected_phase);
1138                 n++;
1139         }
1140
1141         LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1142                  atomic_read(&set->set_remaining) == n, "%d / %d\n",
1143                  atomic_read(&set->set_remaining), n);
1144
1145         list_for_each_safe(tmp, next, &set->set_requests) {
1146                 struct ptlrpc_request *req =
1147                         list_entry(tmp, struct ptlrpc_request,
1148                                    rq_set_chain);
1149                 list_del_init(&req->rq_set_chain);
1150
1151                 LASSERT(req->rq_phase == expected_phase);
1152
1153                 if (req->rq_phase == RQ_PHASE_NEW) {
1154                         ptlrpc_req_interpret(NULL, req, -EBADR);
1155                         atomic_dec(&set->set_remaining);
1156                 }
1157
1158                 spin_lock(&req->rq_lock);
1159                 req->rq_set = NULL;
1160                 req->rq_invalid_rqset = 0;
1161                 spin_unlock(&req->rq_lock);
1162
1163                 ptlrpc_req_finished(req);
1164         }
1165
1166         LASSERT(atomic_read(&set->set_remaining) == 0);
1167
1168         ptlrpc_reqset_put(set);
1169         EXIT;
1170 }
1171 EXPORT_SYMBOL(ptlrpc_set_destroy);
1172
1173 /**
1174  * Add a new request to the general purpose request set.
1175  * Assumes request reference from the caller.
1176  */
1177 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1178                         struct ptlrpc_request *req)
1179 {
1180         if (set == PTLRPCD_SET) {
1181                 ptlrpcd_add_req(req);
1182                 return;
1183         }
1184
1185         LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1186         LASSERT(list_empty(&req->rq_set_chain));
1187
1188         if (req->rq_allow_intr)
1189                 set->set_allow_intr = 1;
1190
1191         /* The set takes over the caller's request reference */
1192         list_add_tail(&req->rq_set_chain, &set->set_requests);
1193         req->rq_set = set;
1194         atomic_inc(&set->set_remaining);
1195         req->rq_queued_time = ktime_get_seconds();
1196
1197         if (req->rq_reqmsg)
1198                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1199
1200         if (set->set_producer)
1201                 /*
1202                  * If the request set has a producer callback, the RPC must be
1203                  * sent straight away
1204                  */
1205                 ptlrpc_send_new_req(req);
1206 }
1207 EXPORT_SYMBOL(ptlrpc_set_add_req);
1208
1209 /**
1210  * Add a request to a request with dedicated server thread
1211  * and wake the thread to make any necessary processing.
1212  * Currently only used for ptlrpcd.
1213  */
1214 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1215                             struct ptlrpc_request *req)
1216 {
1217         struct ptlrpc_request_set *set = pc->pc_set;
1218         int count, i;
1219
1220         LASSERT(req->rq_set == NULL);
1221         LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1222
1223         spin_lock(&set->set_new_req_lock);
1224         /*
1225          * The set takes over the caller's request reference.
1226          */
1227         req->rq_set = set;
1228         req->rq_queued_time = ktime_get_seconds();
1229         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1230         count = atomic_inc_return(&set->set_new_count);
1231         spin_unlock(&set->set_new_req_lock);
1232
1233         /* Only need to call wakeup once for the first entry. */
1234         if (count == 1) {
1235                 wake_up(&set->set_waitq);
1236
1237                 /*
1238                  * XXX: It maybe unnecessary to wakeup all the partners. But to
1239                  *      guarantee the async RPC can be processed ASAP, we have
1240                  *      no other better choice. It maybe fixed in future.
1241                  */
1242                 for (i = 0; i < pc->pc_npartners; i++)
1243                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1244         }
1245 }
1246
1247 /**
1248  * Based on the current state of the import, determine if the request
1249  * can be sent, is an error, or should be delayed.
1250  *
1251  * Returns true if this request should be delayed. If false, and
1252  * *status is set, then the request can not be sent and *status is the
1253  * error code.  If false and status is 0, then request can be sent.
1254  *
1255  * The imp->imp_lock must be held.
1256  */
1257 static int ptlrpc_import_delay_req(struct obd_import *imp,
1258                                    struct ptlrpc_request *req, int *status)
1259 {
1260         int delay = 0;
1261
1262         ENTRY;
1263         LASSERT(status);
1264         *status = 0;
1265
1266         if (req->rq_ctx_init || req->rq_ctx_fini) {
1267                 /* always allow ctx init/fini rpc go through */
1268         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1269                 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1270                 *status = -EIO;
1271         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1272                 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1273
1274                 /*
1275                  * pings or MDS-equivalent STATFS may safely
1276                  * race with umount
1277                  */
1278                 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1279                           D_HA : D_ERROR, req, "IMP_CLOSED");
1280                 *status = -EIO;
1281         } else if (ptlrpc_send_limit_expired(req)) {
1282                 /* probably doesn't need to be a D_ERROR afterinitial testing */
1283                 DEBUG_REQ(D_HA, req, "send limit expired");
1284                 *status = -ETIMEDOUT;
1285         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1286                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
1287                 ;/* allow CONNECT even if import is invalid */
1288                 if (atomic_read(&imp->imp_inval_count) != 0) {
1289                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1290                         *status = -EIO;
1291                 }
1292         } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1293                 if (!imp->imp_deactive)
1294                         DEBUG_REQ(D_NET, req, "IMP_INVALID");
1295                 *status = -ESHUTDOWN; /* b=12940 */
1296         } else if (req->rq_import_generation != imp->imp_generation) {
1297                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1298                 *status = -EIO;
1299         } else if (req->rq_send_state != imp->imp_state) {
1300                 /* invalidate in progress - any requests should be drop */
1301                 if (atomic_read(&imp->imp_inval_count) != 0) {
1302                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1303                         *status = -EIO;
1304                 } else if (req->rq_no_delay &&
1305                            imp->imp_generation != imp->imp_initiated_at) {
1306                         /* ignore nodelay for requests initiating connections */
1307                         *status = -EWOULDBLOCK;
1308                 } else if (req->rq_allow_replay &&
1309                            (imp->imp_state == LUSTRE_IMP_REPLAY ||
1310                             imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1311                             imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1312                             imp->imp_state == LUSTRE_IMP_RECOVER)) {
1313                         DEBUG_REQ(D_HA, req, "allow during recovery");
1314                 } else {
1315                         delay = 1;
1316                 }
1317         }
1318
1319         RETURN(delay);
1320 }
1321
1322 /**
1323  * Decide if the error message should be printed to the console or not.
1324  * Makes its decision based on request type, status, and failure frequency.
1325  *
1326  * \param[in] req  request that failed and may need a console message
1327  *
1328  * \retval false if no message should be printed
1329  * \retval true  if console message should be printed
1330  */
1331 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1332 {
1333         LASSERT(req->rq_reqmsg != NULL);
1334
1335         /* Suppress particular reconnect errors which are to be expected. */
1336         if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1337                 /* Suppress timed out reconnect requests */
1338                 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1339                     req->rq_timedout)
1340                         return false;
1341
1342                 /*
1343                  * Suppress most unavailable/again reconnect requests, but
1344                  * print occasionally so it is clear client is trying to
1345                  * connect to a server where no target is running.
1346                  */
1347                 if ((err == -ENODEV || err == -EAGAIN) &&
1348                     req->rq_import->imp_conn_cnt % 30 != 20)
1349                         return false;
1350         }
1351
1352         if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1353                 /* -EAGAIN is normal when using POSIX flocks */
1354                 return false;
1355
1356         if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1357             (req->rq_xid & 0xf) != 10)
1358                 /* Suppress most ping requests, they may fail occasionally */
1359                 return false;
1360
1361         return true;
1362 }
1363
1364 /**
1365  * Check request processing status.
1366  * Returns the status.
1367  */
1368 static int ptlrpc_check_status(struct ptlrpc_request *req)
1369 {
1370         int rc;
1371
1372         ENTRY;
1373         rc = lustre_msg_get_status(req->rq_repmsg);
1374         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1375                 struct obd_import *imp = req->rq_import;
1376                 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1377                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1378
1379                 if (ptlrpc_console_allow(req, opc, rc))
1380                         LCONSOLE_ERROR_MSG(0x11,
1381                                            "%s: operation %s to node %s failed: rc = %d\n",
1382                                            imp->imp_obd->obd_name,
1383                                            ll_opcode2str(opc),
1384                                            libcfs_nid2str(nid), rc);
1385                 RETURN(rc < 0 ? rc : -EINVAL);
1386         }
1387
1388         if (rc)
1389                 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1390
1391         RETURN(rc);
1392 }
1393
1394 /**
1395  * save pre-versions of objects into request for replay.
1396  * Versions are obtained from server reply.
1397  * used for VBR.
1398  */
1399 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1400 {
1401         struct lustre_msg *repmsg = req->rq_repmsg;
1402         struct lustre_msg *reqmsg = req->rq_reqmsg;
1403         __u64 *versions = lustre_msg_get_versions(repmsg);
1404
1405         ENTRY;
1406         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1407                 return;
1408
1409         LASSERT(versions);
1410         lustre_msg_set_versions(reqmsg, versions);
1411         CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1412                versions[0], versions[1]);
1413
1414         EXIT;
1415 }
1416
1417 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1418 {
1419         struct ptlrpc_request *req;
1420
1421         assert_spin_locked(&imp->imp_lock);
1422         if (list_empty(&imp->imp_unreplied_list))
1423                 return 0;
1424
1425         req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1426                          rq_unreplied_list);
1427         LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1428
1429         if (imp->imp_known_replied_xid < req->rq_xid - 1)
1430                 imp->imp_known_replied_xid = req->rq_xid - 1;
1431
1432         return req->rq_xid - 1;
1433 }
1434
1435 /**
1436  * Callback function called when client receives RPC reply for \a req.
1437  * Returns 0 on success or error code.
1438  * The return alue would be assigned to req->rq_status by the caller
1439  * as request processing status.
1440  * This function also decides if the request needs to be saved for later replay.
1441  */
1442 static int after_reply(struct ptlrpc_request *req)
1443 {
1444         struct obd_import *imp = req->rq_import;
1445         struct obd_device *obd = req->rq_import->imp_obd;
1446         ktime_t work_start;
1447         u64 committed;
1448         s64 timediff;
1449         int rc;
1450
1451         ENTRY;
1452         LASSERT(obd != NULL);
1453         /* repbuf must be unlinked */
1454         LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1455
1456         if (req->rq_reply_truncated) {
1457                 if (ptlrpc_no_resend(req)) {
1458                         DEBUG_REQ(D_ERROR, req,
1459                                   "reply buffer overflow, expected=%d, actual size=%d",
1460                                   req->rq_nob_received, req->rq_repbuf_len);
1461                         RETURN(-EOVERFLOW);
1462                 }
1463
1464                 sptlrpc_cli_free_repbuf(req);
1465                 /*
1466                  * Pass the required reply buffer size (include
1467                  * space for early reply).
1468                  * NB: no need to roundup because alloc_repbuf
1469                  * will roundup it
1470                  */
1471                 req->rq_replen = req->rq_nob_received;
1472                 req->rq_nob_received = 0;
1473                 spin_lock(&req->rq_lock);
1474                 req->rq_resend       = 1;
1475                 spin_unlock(&req->rq_lock);
1476                 RETURN(0);
1477         }
1478
1479         work_start = ktime_get_real();
1480         timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1481
1482         /*
1483          * NB Until this point, the whole of the incoming message,
1484          * including buflens, status etc is in the sender's byte order.
1485          */
1486         rc = sptlrpc_cli_unwrap_reply(req);
1487         if (rc) {
1488                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1489                 RETURN(rc);
1490         }
1491
1492         /*
1493          * Security layer unwrap might ask resend this request.
1494          */
1495         if (req->rq_resend)
1496                 RETURN(0);
1497
1498         rc = unpack_reply(req);
1499         if (rc)
1500                 RETURN(rc);
1501
1502         /* retry indefinitely on EINPROGRESS */
1503         if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1504             ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1505                 time64_t now = ktime_get_real_seconds();
1506
1507                 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1508                           D_RPCTRACE, req, "resending request on EINPROGRESS");
1509                 spin_lock(&req->rq_lock);
1510                 req->rq_resend = 1;
1511                 spin_unlock(&req->rq_lock);
1512                 req->rq_nr_resend++;
1513
1514                 /* Readjust the timeout for current conditions */
1515                 ptlrpc_at_set_req_timeout(req);
1516                 /*
1517                  * delay resend to give a chance to the server to get ready.
1518                  * The delay is increased by 1s on every resend and is capped to
1519                  * the current request timeout (i.e. obd_timeout if AT is off,
1520                  * or AT service time x 125% + 5s, see at_est2timeout)
1521                  */
1522                 if (req->rq_nr_resend > req->rq_timeout)
1523                         req->rq_sent = now + req->rq_timeout;
1524                 else
1525                         req->rq_sent = now + req->rq_nr_resend;
1526
1527                 /* Resend for EINPROGRESS will use a new XID */
1528                 spin_lock(&imp->imp_lock);
1529                 list_del_init(&req->rq_unreplied_list);
1530                 spin_unlock(&imp->imp_lock);
1531
1532                 RETURN(0);
1533         }
1534
1535         if (obd->obd_svc_stats) {
1536                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1537                                     timediff);
1538                 ptlrpc_lprocfs_rpc_sent(req, timediff);
1539         }
1540
1541         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1542             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1543                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1544                           lustre_msg_get_type(req->rq_repmsg));
1545                 RETURN(-EPROTO);
1546         }
1547
1548         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1549                 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1550         ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1551         ptlrpc_at_adj_net_latency(req,
1552                                   lustre_msg_get_service_time(req->rq_repmsg));
1553
1554         rc = ptlrpc_check_status(req);
1555
1556         if (rc) {
1557                 /*
1558                  * Either we've been evicted, or the server has failed for
1559                  * some reason. Try to reconnect, and if that fails, punt to
1560                  * the upcall.
1561                  */
1562                 if (ptlrpc_recoverable_error(rc)) {
1563                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
1564                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1565                                 RETURN(rc);
1566                         }
1567                         ptlrpc_request_handle_notconn(req);
1568                         RETURN(rc);
1569                 }
1570         } else {
1571                 /*
1572                  * Let's look if server sent slv. Do it only for RPC with
1573                  * rc == 0.
1574                  */
1575                 ldlm_cli_update_pool(req);
1576         }
1577
1578         /*
1579          * Store transno in reqmsg for replay.
1580          */
1581         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1582                 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1583                 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1584         }
1585
1586         if (imp->imp_replayable) {
1587                 spin_lock(&imp->imp_lock);
1588                 /*
1589                  * No point in adding already-committed requests to the replay
1590                  * list, we will just remove them immediately. b=9829
1591                  */
1592                 if (req->rq_transno != 0 &&
1593                     (req->rq_transno >
1594                      lustre_msg_get_last_committed(req->rq_repmsg) ||
1595                      req->rq_replay)) {
1596                         /** version recovery */
1597                         ptlrpc_save_versions(req);
1598                         ptlrpc_retain_replayable_request(req, imp);
1599                 } else if (req->rq_commit_cb &&
1600                            list_empty(&req->rq_replay_list)) {
1601                         /*
1602                          * NB: don't call rq_commit_cb if it's already on
1603                          * rq_replay_list, ptlrpc_free_committed() will call
1604                          * it later, see LU-3618 for details
1605                          */
1606                         spin_unlock(&imp->imp_lock);
1607                         req->rq_commit_cb(req);
1608                         spin_lock(&imp->imp_lock);
1609                 }
1610
1611                 /*
1612                  * Replay-enabled imports return commit-status information.
1613                  */
1614                 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1615                 if (likely(committed > imp->imp_peer_committed_transno))
1616                         imp->imp_peer_committed_transno = committed;
1617
1618                 ptlrpc_free_committed(imp);
1619
1620                 if (!list_empty(&imp->imp_replay_list)) {
1621                         struct ptlrpc_request *last;
1622
1623                         last = list_entry(imp->imp_replay_list.prev,
1624                                           struct ptlrpc_request,
1625                                           rq_replay_list);
1626                         /*
1627                          * Requests with rq_replay stay on the list even if no
1628                          * commit is expected.
1629                          */
1630                         if (last->rq_transno > imp->imp_peer_committed_transno)
1631                                 ptlrpc_pinger_commit_expected(imp);
1632                 }
1633
1634                 spin_unlock(&imp->imp_lock);
1635         }
1636
1637         RETURN(rc);
1638 }
1639
1640 /**
1641  * Helper function to send request \a req over the network for the first time
1642  * Also adjusts request phase.
1643  * Returns 0 on success or error code.
1644  */
1645 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1646 {
1647         struct obd_import *imp = req->rq_import;
1648         __u64 min_xid = 0;
1649         int rc;
1650
1651         ENTRY;
1652         LASSERT(req->rq_phase == RQ_PHASE_NEW);
1653
1654         /* do not try to go further if there is not enough memory in enc_pool */
1655         if (req->rq_sent && req->rq_bulk)
1656                 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1657                     pool_is_at_full_capacity())
1658                         RETURN(-ENOMEM);
1659
1660         if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1661             (!req->rq_generation_set ||
1662              req->rq_import_generation == imp->imp_generation))
1663                 RETURN(0);
1664
1665         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1666
1667         spin_lock(&imp->imp_lock);
1668
1669         LASSERT(req->rq_xid != 0);
1670         LASSERT(!list_empty(&req->rq_unreplied_list));
1671
1672         if (!req->rq_generation_set)
1673                 req->rq_import_generation = imp->imp_generation;
1674
1675         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1676                 spin_lock(&req->rq_lock);
1677                 req->rq_waiting = 1;
1678                 spin_unlock(&req->rq_lock);
1679
1680                 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1681                           ptlrpc_import_state_name(req->rq_send_state),
1682                           ptlrpc_import_state_name(imp->imp_state));
1683                 LASSERT(list_empty(&req->rq_list));
1684                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1685                 atomic_inc(&req->rq_import->imp_inflight);
1686                 spin_unlock(&imp->imp_lock);
1687                 RETURN(0);
1688         }
1689
1690         if (rc != 0) {
1691                 spin_unlock(&imp->imp_lock);
1692                 req->rq_status = rc;
1693                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1694                 RETURN(rc);
1695         }
1696
1697         LASSERT(list_empty(&req->rq_list));
1698         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1699         atomic_inc(&req->rq_import->imp_inflight);
1700
1701         /*
1702          * find the known replied XID from the unreplied list, CONNECT
1703          * and DISCONNECT requests are skipped to make the sanity check
1704          * on server side happy. see process_req_last_xid().
1705          *
1706          * For CONNECT: Because replay requests have lower XID, it'll
1707          * break the sanity check if CONNECT bump the exp_last_xid on
1708          * server.
1709          *
1710          * For DISCONNECT: Since client will abort inflight RPC before
1711          * sending DISCONNECT, DISCONNECT may carry an XID which higher
1712          * than the inflight RPC.
1713          */
1714         if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1715                 min_xid = ptlrpc_known_replied_xid(imp);
1716         spin_unlock(&imp->imp_lock);
1717
1718         lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1719
1720         lustre_msg_set_status(req->rq_reqmsg, current_pid());
1721
1722         rc = sptlrpc_req_refresh_ctx(req, -1);
1723         if (rc) {
1724                 if (req->rq_err) {
1725                         req->rq_status = rc;
1726                         RETURN(1);
1727                 } else {
1728                         spin_lock(&req->rq_lock);
1729                         req->rq_wait_ctx = 1;
1730                         spin_unlock(&req->rq_lock);
1731                         RETURN(0);
1732                 }
1733         }
1734
1735         CDEBUG(D_RPCTRACE,
1736                "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1737                req, current_comm(),
1738                imp->imp_obd->obd_uuid.uuid,
1739                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1740                obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1741                lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1742
1743         rc = ptl_send_rpc(req, 0);
1744         if (rc == -ENOMEM) {
1745                 spin_lock(&imp->imp_lock);
1746                 if (!list_empty(&req->rq_list)) {
1747                         list_del_init(&req->rq_list);
1748                         if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1749                                 wake_up(&req->rq_import->imp_recovery_waitq);
1750                 }
1751                 spin_unlock(&imp->imp_lock);
1752                 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1753                 RETURN(rc);
1754         }
1755         if (rc) {
1756                 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1757                           rc);
1758                 spin_lock(&req->rq_lock);
1759                 req->rq_net_err = 1;
1760                 spin_unlock(&req->rq_lock);
1761                 RETURN(rc);
1762         }
1763         RETURN(0);
1764 }
1765
1766 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1767 {
1768         int remaining, rc;
1769
1770         ENTRY;
1771         LASSERT(set->set_producer != NULL);
1772
1773         remaining = atomic_read(&set->set_remaining);
1774
1775         /*
1776          * populate the ->set_requests list with requests until we
1777          * reach the maximum number of RPCs in flight for this set
1778          */
1779         while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1780                 rc = set->set_producer(set, set->set_producer_arg);
1781                 if (rc == -ENOENT) {
1782                         /* no more RPC to produce */
1783                         set->set_producer     = NULL;
1784                         set->set_producer_arg = NULL;
1785                         RETURN(0);
1786                 }
1787         }
1788
1789         RETURN((atomic_read(&set->set_remaining) - remaining));
1790 }
1791
1792 /**
1793  * this sends any unsent RPCs in \a set and returns 1 if all are sent
1794  * and no more replies are expected.
1795  * (it is possible to get less replies than requests sent e.g. due to timed out
1796  * requests or requests that we had trouble to send out)
1797  *
1798  * NOTE: This function contains a potential schedule point (cond_resched()).
1799  */
1800 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1801 {
1802         struct list_head *tmp, *next;
1803         LIST_HEAD(comp_reqs);
1804         int force_timer_recalc = 0;
1805
1806         ENTRY;
1807         if (atomic_read(&set->set_remaining) == 0)
1808                 RETURN(1);
1809
1810         list_for_each_safe(tmp, next, &set->set_requests) {
1811                 struct ptlrpc_request *req =
1812                         list_entry(tmp, struct ptlrpc_request,
1813                                    rq_set_chain);
1814                 struct obd_import *imp = req->rq_import;
1815                 int unregistered = 0;
1816                 int async = 1;
1817                 int rc = 0;
1818
1819                 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1820                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1821                         continue;
1822                 }
1823
1824                 /*
1825                  * This schedule point is mainly for the ptlrpcd caller of this
1826                  * function.  Most ptlrpc sets are not long-lived and unbounded
1827                  * in length, but at the least the set used by the ptlrpcd is.
1828                  * Since the processing time is unbounded, we need to insert an
1829                  * explicit schedule point to make the thread well-behaved.
1830                  */
1831                 cond_resched();
1832
1833                 /*
1834                  * If the caller requires to allow to be interpreted by force
1835                  * and it has really been interpreted, then move the request
1836                  * to RQ_PHASE_INTERPRET phase in spite of what the current
1837                  * phase is.
1838                  */
1839                 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1840                         req->rq_status = -EINTR;
1841                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1842
1843                         /*
1844                          * Since it is interpreted and we have to wait for
1845                          * the reply to be unlinked, then use sync mode.
1846                          */
1847                         async = 0;
1848
1849                         GOTO(interpret, req->rq_status);
1850                 }
1851
1852                 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1853                         force_timer_recalc = 1;
1854
1855                 /* delayed send - skip */
1856                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1857                         continue;
1858
1859                 /* delayed resend - skip */
1860                 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1861                     req->rq_sent > ktime_get_real_seconds())
1862                         continue;
1863
1864                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1865                       req->rq_phase == RQ_PHASE_BULK ||
1866                       req->rq_phase == RQ_PHASE_INTERPRET ||
1867                       req->rq_phase == RQ_PHASE_UNREG_RPC ||
1868                       req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1869                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1870                         LBUG();
1871                 }
1872
1873                 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1874                     req->rq_phase == RQ_PHASE_UNREG_BULK) {
1875                         LASSERT(req->rq_next_phase != req->rq_phase);
1876                         LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1877
1878                         if (req->rq_req_deadline &&
1879                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1880                                 req->rq_req_deadline = 0;
1881                         if (req->rq_reply_deadline &&
1882                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1883                                 req->rq_reply_deadline = 0;
1884                         if (req->rq_bulk_deadline &&
1885                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1886                                 req->rq_bulk_deadline = 0;
1887
1888                         /*
1889                          * Skip processing until reply is unlinked. We
1890                          * can't return to pool before that and we can't
1891                          * call interpret before that. We need to make
1892                          * sure that all rdma transfers finished and will
1893                          * not corrupt any data.
1894                          */
1895                         if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1896                             ptlrpc_client_recv_or_unlink(req))
1897                                 continue;
1898                         if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1899                             ptlrpc_client_bulk_active(req))
1900                                 continue;
1901
1902                         /*
1903                          * Turn fail_loc off to prevent it from looping
1904                          * forever.
1905                          */
1906                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1907                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1908                                                      OBD_FAIL_ONCE);
1909                         }
1910                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1911                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1912                                                      OBD_FAIL_ONCE);
1913                         }
1914
1915                         /*
1916                          * Move to next phase if reply was successfully
1917                          * unlinked.
1918                          */
1919                         ptlrpc_rqphase_move(req, req->rq_next_phase);
1920                 }
1921
1922                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1923                         GOTO(interpret, req->rq_status);
1924
1925                 /*
1926                  * Note that this also will start async reply unlink.
1927                  */
1928                 if (req->rq_net_err && !req->rq_timedout) {
1929                         ptlrpc_expire_one_request(req, 1);
1930
1931                         /*
1932                          * Check if we still need to wait for unlink.
1933                          */
1934                         if (ptlrpc_client_recv_or_unlink(req) ||
1935                             ptlrpc_client_bulk_active(req))
1936                                 continue;
1937                         /* If there is no need to resend, fail it now. */
1938                         if (req->rq_no_resend) {
1939                                 if (req->rq_status == 0)
1940                                         req->rq_status = -EIO;
1941                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1942                                 GOTO(interpret, req->rq_status);
1943                         } else {
1944                                 continue;
1945                         }
1946                 }
1947
1948                 if (req->rq_err) {
1949                         spin_lock(&req->rq_lock);
1950                         req->rq_replied = 0;
1951                         spin_unlock(&req->rq_lock);
1952                         if (req->rq_status == 0)
1953                                 req->rq_status = -EIO;
1954                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1955                         GOTO(interpret, req->rq_status);
1956                 }
1957
1958                 /*
1959                  * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1960                  * so it sets rq_intr regardless of individual rpc
1961                  * timeouts. The synchronous IO waiting path sets
1962                  * rq_intr irrespective of whether ptlrpcd
1963                  * has seen a timeout.  Our policy is to only interpret
1964                  * interrupted rpcs after they have timed out, so we
1965                  * need to enforce that here.
1966                  */
1967
1968                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1969                                      req->rq_wait_ctx)) {
1970                         req->rq_status = -EINTR;
1971                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1972                         GOTO(interpret, req->rq_status);
1973                 }
1974
1975                 if (req->rq_phase == RQ_PHASE_RPC) {
1976                         if (req->rq_timedout || req->rq_resend ||
1977                             req->rq_waiting || req->rq_wait_ctx) {
1978                                 int status;
1979
1980                                 if (!ptlrpc_unregister_reply(req, 1)) {
1981                                         ptlrpc_unregister_bulk(req, 1);
1982                                         continue;
1983                                 }
1984
1985                                 spin_lock(&imp->imp_lock);
1986                                 if (ptlrpc_import_delay_req(imp, req,
1987                                                             &status)) {
1988                                         /*
1989                                          * put on delay list - only if we wait
1990                                          * recovery finished - before send
1991                                          */
1992                                         list_move_tail(&req->rq_list,
1993                                                        &imp->imp_delayed_list);
1994                                         spin_unlock(&imp->imp_lock);
1995                                         continue;
1996                                 }
1997
1998                                 if (status != 0)  {
1999                                         req->rq_status = status;
2000                                         ptlrpc_rqphase_move(req,
2001                                                             RQ_PHASE_INTERPRET);
2002                                         spin_unlock(&imp->imp_lock);
2003                                         GOTO(interpret, req->rq_status);
2004                                 }
2005                                 /* ignore on just initiated connections */
2006                                 if (ptlrpc_no_resend(req) &&
2007                                     !req->rq_wait_ctx &&
2008                                     imp->imp_generation !=
2009                                     imp->imp_initiated_at) {
2010                                         req->rq_status = -ENOTCONN;
2011                                         ptlrpc_rqphase_move(req,
2012                                                             RQ_PHASE_INTERPRET);
2013                                         spin_unlock(&imp->imp_lock);
2014                                         GOTO(interpret, req->rq_status);
2015                                 }
2016
2017                                 list_move_tail(&req->rq_list,
2018                                                &imp->imp_sending_list);
2019
2020                                 spin_unlock(&imp->imp_lock);
2021
2022                                 spin_lock(&req->rq_lock);
2023                                 req->rq_waiting = 0;
2024                                 spin_unlock(&req->rq_lock);
2025
2026                                 if (req->rq_timedout || req->rq_resend) {
2027                                         /*
2028                                          * This is re-sending anyways,
2029                                          * let's mark req as resend.
2030                                          */
2031                                         spin_lock(&req->rq_lock);
2032                                         req->rq_resend = 1;
2033                                         spin_unlock(&req->rq_lock);
2034                                 }
2035                                 /*
2036                                  * rq_wait_ctx is only touched by ptlrpcd,
2037                                  * so no lock is needed here.
2038                                  */
2039                                 status = sptlrpc_req_refresh_ctx(req, -1);
2040                                 if (status) {
2041                                         if (req->rq_err) {
2042                                                 req->rq_status = status;
2043                                                 spin_lock(&req->rq_lock);
2044                                                 req->rq_wait_ctx = 0;
2045                                                 spin_unlock(&req->rq_lock);
2046                                                 force_timer_recalc = 1;
2047                                         } else {
2048                                                 spin_lock(&req->rq_lock);
2049                                                 req->rq_wait_ctx = 1;
2050                                                 spin_unlock(&req->rq_lock);
2051                                         }
2052
2053                                         continue;
2054                                 } else {
2055                                         spin_lock(&req->rq_lock);
2056                                         req->rq_wait_ctx = 0;
2057                                         spin_unlock(&req->rq_lock);
2058                                 }
2059
2060                                 /*
2061                                  * In any case, the previous bulk should be
2062                                  * cleaned up to prepare for the new sending
2063                                  */
2064                                 if (req->rq_bulk &&
2065                                     !ptlrpc_unregister_bulk(req, 1))
2066                                         continue;
2067
2068                                 rc = ptl_send_rpc(req, 0);
2069                                 if (rc == -ENOMEM) {
2070                                         spin_lock(&imp->imp_lock);
2071                                         if (!list_empty(&req->rq_list))
2072                                                 list_del_init(&req->rq_list);
2073                                         spin_unlock(&imp->imp_lock);
2074                                         ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2075                                         continue;
2076                                 }
2077                                 if (rc) {
2078                                         DEBUG_REQ(D_HA, req,
2079                                                   "send failed: rc = %d", rc);
2080                                         force_timer_recalc = 1;
2081                                         spin_lock(&req->rq_lock);
2082                                         req->rq_net_err = 1;
2083                                         spin_unlock(&req->rq_lock);
2084                                         continue;
2085                                 }
2086                                 /* need to reset the timeout */
2087                                 force_timer_recalc = 1;
2088                         }
2089
2090                         spin_lock(&req->rq_lock);
2091
2092                         if (ptlrpc_client_early(req)) {
2093                                 ptlrpc_at_recv_early_reply(req);
2094                                 spin_unlock(&req->rq_lock);
2095                                 continue;
2096                         }
2097
2098                         /* Still waiting for a reply? */
2099                         if (ptlrpc_client_recv(req)) {
2100                                 spin_unlock(&req->rq_lock);
2101                                 continue;
2102                         }
2103
2104                         /* Did we actually receive a reply? */
2105                         if (!ptlrpc_client_replied(req)) {
2106                                 spin_unlock(&req->rq_lock);
2107                                 continue;
2108                         }
2109
2110                         spin_unlock(&req->rq_lock);
2111
2112                         /*
2113                          * unlink from net because we are going to
2114                          * swab in-place of reply buffer
2115                          */
2116                         unregistered = ptlrpc_unregister_reply(req, 1);
2117                         if (!unregistered)
2118                                 continue;
2119
2120                         req->rq_status = after_reply(req);
2121                         if (req->rq_resend)
2122                                 continue;
2123
2124                         /*
2125                          * If there is no bulk associated with this request,
2126                          * then we're done and should let the interpreter
2127                          * process the reply. Similarly if the RPC returned
2128                          * an error, and therefore the bulk will never arrive.
2129                          */
2130                         if (!req->rq_bulk || req->rq_status < 0) {
2131                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2132                                 GOTO(interpret, req->rq_status);
2133                         }
2134
2135                         ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2136                 }
2137
2138                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2139                 if (ptlrpc_client_bulk_active(req))
2140                         continue;
2141
2142                 if (req->rq_bulk->bd_failure) {
2143                         /*
2144                          * The RPC reply arrived OK, but the bulk screwed
2145                          * up!  Dead weird since the server told us the RPC
2146                          * was good after getting the REPLY for her GET or
2147                          * the ACK for her PUT.
2148                          */
2149                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2150                         req->rq_status = -EIO;
2151                 }
2152
2153                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2154
2155 interpret:
2156                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2157
2158                 /*
2159                  * This moves to "unregistering" phase we need to wait for
2160                  * reply unlink.
2161                  */
2162                 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2163                         /* start async bulk unlink too */
2164                         ptlrpc_unregister_bulk(req, 1);
2165                         continue;
2166                 }
2167
2168                 if (!ptlrpc_unregister_bulk(req, async))
2169                         continue;
2170
2171                 /*
2172                  * When calling interpret receiving already should be
2173                  * finished.
2174                  */
2175                 LASSERT(!req->rq_receiving_reply);
2176
2177                 ptlrpc_req_interpret(env, req, req->rq_status);
2178
2179                 if (ptlrpcd_check_work(req)) {
2180                         atomic_dec(&set->set_remaining);
2181                         continue;
2182                 }
2183                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2184
2185                 if (req->rq_reqmsg)
2186                         CDEBUG(D_RPCTRACE,
2187                                "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2188                                req, current_comm(),
2189                                imp->imp_obd->obd_uuid.uuid,
2190                                lustre_msg_get_status(req->rq_reqmsg),
2191                                req->rq_xid,
2192                                obd_import_nid2str(imp),
2193                                lustre_msg_get_opc(req->rq_reqmsg),
2194                                lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2195
2196                 spin_lock(&imp->imp_lock);
2197                 /*
2198                  * Request already may be not on sending or delaying list. This
2199                  * may happen in the case of marking it erroneous for the case
2200                  * ptlrpc_import_delay_req(req, status) find it impossible to
2201                  * allow sending this rpc and returns *status != 0.
2202                  */
2203                 if (!list_empty(&req->rq_list)) {
2204                         list_del_init(&req->rq_list);
2205                         if (atomic_dec_and_test(&imp->imp_inflight))
2206                                 wake_up(&imp->imp_recovery_waitq);
2207                 }
2208                 list_del_init(&req->rq_unreplied_list);
2209                 spin_unlock(&imp->imp_lock);
2210
2211                 atomic_dec(&set->set_remaining);
2212                 wake_up(&imp->imp_recovery_waitq);
2213
2214                 if (set->set_producer) {
2215                         /* produce a new request if possible */
2216                         if (ptlrpc_set_producer(set) > 0)
2217                                 force_timer_recalc = 1;
2218
2219                         /*
2220                          * free the request that has just been completed
2221                          * in order not to pollute set->set_requests
2222                          */
2223                         list_del_init(&req->rq_set_chain);
2224                         spin_lock(&req->rq_lock);
2225                         req->rq_set = NULL;
2226                         req->rq_invalid_rqset = 0;
2227                         spin_unlock(&req->rq_lock);
2228
2229                         /* record rq_status to compute the final status later */
2230                         if (req->rq_status != 0)
2231                                 set->set_rc = req->rq_status;
2232                         ptlrpc_req_finished(req);
2233                 } else {
2234                         list_move_tail(&req->rq_set_chain, &comp_reqs);
2235                 }
2236         }
2237
2238         /*
2239          * move completed request at the head of list so it's easier for
2240          * caller to find them
2241          */
2242         list_splice(&comp_reqs, &set->set_requests);
2243
2244         /* If we hit an error, we want to recover promptly. */
2245         RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2246 }
2247 EXPORT_SYMBOL(ptlrpc_check_set);
2248
2249 /**
2250  * Time out request \a req. is \a async_unlink is set, that means do not wait
2251  * until LNet actually confirms network buffer unlinking.
2252  * Return 1 if we should give up further retrying attempts or 0 otherwise.
2253  */
2254 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2255 {
2256         struct obd_import *imp = req->rq_import;
2257         unsigned int debug_mask = D_RPCTRACE;
2258         int rc = 0;
2259
2260         ENTRY;
2261         spin_lock(&req->rq_lock);
2262         req->rq_timedout = 1;
2263         spin_unlock(&req->rq_lock);
2264
2265         if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2266                                  lustre_msg_get_status(req->rq_reqmsg)))
2267                 debug_mask = D_WARNING;
2268         DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2269                   req->rq_net_err ? "failed due to network error" :
2270                      ((req->rq_real_sent == 0 ||
2271                        req->rq_real_sent < req->rq_sent ||
2272                        req->rq_real_sent >= req->rq_deadline) ?
2273                       "timed out for sent delay" : "timed out for slow reply"),
2274                   (s64)req->rq_sent, (s64)req->rq_real_sent);
2275
2276         if (imp && obd_debug_peer_on_timeout)
2277                 LNetDebugPeer(imp->imp_connection->c_peer);
2278
2279         ptlrpc_unregister_reply(req, async_unlink);
2280         ptlrpc_unregister_bulk(req, async_unlink);
2281
2282         if (obd_dump_on_timeout)
2283                 libcfs_debug_dumplog();
2284
2285         if (!imp) {
2286                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2287                 RETURN(1);
2288         }
2289
2290         atomic_inc(&imp->imp_timeouts);
2291
2292         /* The DLM server doesn't want recovery run on its imports. */
2293         if (imp->imp_dlm_fake)
2294                 RETURN(1);
2295
2296         /*
2297          * If this request is for recovery or other primordial tasks,
2298          * then error it out here.
2299          */
2300         if (req->rq_ctx_init || req->rq_ctx_fini ||
2301             req->rq_send_state != LUSTRE_IMP_FULL ||
2302             imp->imp_obd->obd_no_recov) {
2303                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2304                           ptlrpc_import_state_name(req->rq_send_state),
2305                           ptlrpc_import_state_name(imp->imp_state));
2306                 spin_lock(&req->rq_lock);
2307                 req->rq_status = -ETIMEDOUT;
2308                 req->rq_err = 1;
2309                 spin_unlock(&req->rq_lock);
2310                 RETURN(1);
2311         }
2312
2313         /*
2314          * if a request can't be resent we can't wait for an answer after
2315          * the timeout
2316          */
2317         if (ptlrpc_no_resend(req)) {
2318                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2319                 rc = 1;
2320         }
2321
2322         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2323
2324         RETURN(rc);
2325 }
2326
2327 /**
2328  * Time out all uncompleted requests in request set pointed by \a data
2329  * Callback used when waiting on sets with l_wait_event.
2330  * Always returns 1.
2331  */
2332 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2333 {
2334         struct list_head *tmp;
2335         time64_t now = ktime_get_real_seconds();
2336
2337         ENTRY;
2338         LASSERT(set != NULL);
2339
2340         /*
2341          * A timeout expired. See which reqs it applies to...
2342          */
2343         list_for_each(tmp, &set->set_requests) {
2344                 struct ptlrpc_request *req =
2345                         list_entry(tmp, struct ptlrpc_request,
2346                                    rq_set_chain);
2347
2348                 /* don't expire request waiting for context */
2349                 if (req->rq_wait_ctx)
2350                         continue;
2351
2352                 /* Request in-flight? */
2353                 if (!((req->rq_phase == RQ_PHASE_RPC &&
2354                        !req->rq_waiting && !req->rq_resend) ||
2355                       (req->rq_phase == RQ_PHASE_BULK)))
2356                         continue;
2357
2358                 if (req->rq_timedout ||     /* already dealt with */
2359                     req->rq_deadline > now) /* not expired */
2360                         continue;
2361
2362                 /*
2363                  * Deal with this guy. Do it asynchronously to not block
2364                  * ptlrpcd thread.
2365                  */
2366                 ptlrpc_expire_one_request(req, 1);
2367         }
2368 }
2369
2370 /**
2371  * Interrupts (sets interrupted flag) all uncompleted requests in
2372  * a set \a data. Callback for l_wait_event for interruptible waits.
2373  */
2374 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2375 {
2376         struct list_head *tmp;
2377
2378         LASSERT(set != NULL);
2379         CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2380
2381         list_for_each(tmp, &set->set_requests) {
2382                 struct ptlrpc_request *req =
2383                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2384
2385                 if (req->rq_intr)
2386                         continue;
2387
2388                 if (req->rq_phase != RQ_PHASE_RPC &&
2389                     req->rq_phase != RQ_PHASE_UNREG_RPC &&
2390                     !req->rq_allow_intr)
2391                         continue;
2392
2393                 spin_lock(&req->rq_lock);
2394                 req->rq_intr = 1;
2395                 spin_unlock(&req->rq_lock);
2396         }
2397 }
2398
2399 /**
2400  * Get the smallest timeout in the set; this does NOT set a timeout.
2401  */
2402 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2403 {
2404         struct list_head *tmp;
2405         time64_t now = ktime_get_real_seconds();
2406         int timeout = 0;
2407         struct ptlrpc_request *req;
2408         time64_t deadline;
2409
2410         ENTRY;
2411         list_for_each(tmp, &set->set_requests) {
2412                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2413
2414                 /* Request in-flight? */
2415                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2416                       (req->rq_phase == RQ_PHASE_BULK) ||
2417                       (req->rq_phase == RQ_PHASE_NEW)))
2418                         continue;
2419
2420                 /* Already timed out. */
2421                 if (req->rq_timedout)
2422                         continue;
2423
2424                 /* Waiting for ctx. */
2425                 if (req->rq_wait_ctx)
2426                         continue;
2427
2428                 if (req->rq_phase == RQ_PHASE_NEW)
2429                         deadline = req->rq_sent;
2430                 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2431                         deadline = req->rq_sent;
2432                 else
2433                         deadline = req->rq_sent + req->rq_timeout;
2434
2435                 if (deadline <= now)    /* actually expired already */
2436                         timeout = 1;    /* ASAP */
2437                 else if (timeout == 0 || timeout > deadline - now)
2438                         timeout = deadline - now;
2439         }
2440         RETURN(timeout);
2441 }
2442
2443 /**
2444  * Send all unset request from the set and then wait untill all
2445  * requests in the set complete (either get a reply, timeout, get an
2446  * error or otherwise be interrupted).
2447  * Returns 0 on success or error code otherwise.
2448  */
2449 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2450 {
2451         struct list_head *tmp;
2452         struct ptlrpc_request *req;
2453         time64_t timeout;
2454         int rc;
2455
2456         ENTRY;
2457         if (set->set_producer)
2458                 (void)ptlrpc_set_producer(set);
2459         else
2460                 list_for_each(tmp, &set->set_requests) {
2461                         req = list_entry(tmp, struct ptlrpc_request,
2462                                          rq_set_chain);
2463                         if (req->rq_phase == RQ_PHASE_NEW)
2464                                 (void)ptlrpc_send_new_req(req);
2465                 }
2466
2467         if (list_empty(&set->set_requests))
2468                 RETURN(0);
2469
2470         do {
2471                 timeout = ptlrpc_set_next_timeout(set);
2472
2473                 /*
2474                  * wait until all complete, interrupted, or an in-flight
2475                  * req times out
2476                  */
2477                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2478                        set, timeout);
2479
2480                 if ((timeout == 0 && !signal_pending(current)) ||
2481                     set->set_allow_intr) {
2482                         /*
2483                          * No requests are in-flight (ether timed out
2484                          * or delayed), so we can allow interrupts.
2485                          * We still want to block for a limited time,
2486                          * so we allow interrupts during the timeout.
2487                          */
2488                         rc = l_wait_event_abortable_timeout(
2489                                 set->set_waitq,
2490                                 ptlrpc_check_set(NULL, set),
2491                                 cfs_time_seconds(timeout ? timeout : 1));
2492                         if (rc == 0) {
2493                                 rc = -ETIMEDOUT;
2494                                 ptlrpc_expired_set(set);
2495                         } else if (rc < 0) {
2496                                 rc = -EINTR;
2497                                 ptlrpc_interrupted_set(set);
2498                         } else {
2499                                 rc = 0;
2500                         }
2501                 } else {
2502                         /*
2503                          * At least one request is in flight, so no
2504                          * interrupts are allowed. Wait until all
2505                          * complete, or an in-flight req times out.
2506                          */
2507                         rc = wait_event_idle_timeout(
2508                                 set->set_waitq,
2509                                 ptlrpc_check_set(NULL, set),
2510                                 cfs_time_seconds(timeout ? timeout : 1));
2511                         if (rc == 0) {
2512                                 ptlrpc_expired_set(set);
2513                                 rc = -ETIMEDOUT;
2514                         } else {
2515                                 rc = 0;
2516                         }
2517
2518                         /*
2519                          * LU-769 - if we ignored the signal because
2520                          * it was already pending when we started, we
2521                          * need to handle it now or we risk it being
2522                          * ignored forever
2523                          */
2524                         if (rc == -ETIMEDOUT &&
2525                             signal_pending(current)) {
2526                                 sigset_t blocked_sigs =
2527                                         cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2528
2529                                 /*
2530                                  * In fact we only interrupt for the
2531                                  * "fatal" signals like SIGINT or
2532                                  * SIGKILL. We still ignore less
2533                                  * important signals since ptlrpc set
2534                                  * is not easily reentrant from
2535                                  * userspace again
2536                                  */
2537                                 if (signal_pending(current))
2538                                         ptlrpc_interrupted_set(set);
2539                                 cfs_restore_sigs(blocked_sigs);
2540                         }
2541                 }
2542
2543                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2544
2545                 /*
2546                  * -EINTR => all requests have been flagged rq_intr so next
2547                  * check completes.
2548                  * -ETIMEDOUT => someone timed out.  When all reqs have
2549                  * timed out, signals are enabled allowing completion with
2550                  * EINTR.
2551                  * I don't really care if we go once more round the loop in
2552                  * the error cases -eeb.
2553                  */
2554                 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2555                         list_for_each(tmp, &set->set_requests) {
2556                                 req = list_entry(tmp, struct ptlrpc_request,
2557                                                  rq_set_chain);
2558                                 spin_lock(&req->rq_lock);
2559                                 req->rq_invalid_rqset = 1;
2560                                 spin_unlock(&req->rq_lock);
2561                         }
2562                 }
2563         } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2564
2565         LASSERT(atomic_read(&set->set_remaining) == 0);
2566
2567         rc = set->set_rc; /* rq_status of already freed requests if any */
2568         list_for_each(tmp, &set->set_requests) {
2569                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2570
2571                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2572                 if (req->rq_status != 0)
2573                         rc = req->rq_status;
2574         }
2575
2576         RETURN(rc);
2577 }
2578 EXPORT_SYMBOL(ptlrpc_set_wait);
2579
2580 /**
2581  * Helper fuction for request freeing.
2582  * Called when request count reached zero and request needs to be freed.
2583  * Removes request from all sorts of sending/replay lists it might be on,
2584  * frees network buffers if any are present.
2585  * If \a locked is set, that means caller is already holding import imp_lock
2586  * and so we no longer need to reobtain it (for certain lists manipulations)
2587  */
2588 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2589 {
2590         ENTRY;
2591
2592         if (!request)
2593                 RETURN_EXIT;
2594
2595         LASSERT(!request->rq_srv_req);
2596         LASSERT(request->rq_export == NULL);
2597         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2598         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2599         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2600         LASSERTF(!request->rq_replay, "req %p\n", request);
2601
2602         req_capsule_fini(&request->rq_pill);
2603
2604         /*
2605          * We must take it off the imp_replay_list first.  Otherwise, we'll set
2606          * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2607          */
2608         if (request->rq_import) {
2609                 if (!locked)
2610                         spin_lock(&request->rq_import->imp_lock);
2611                 list_del_init(&request->rq_replay_list);
2612                 list_del_init(&request->rq_unreplied_list);
2613                 if (!locked)
2614                         spin_unlock(&request->rq_import->imp_lock);
2615         }
2616         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2617
2618         if (atomic_read(&request->rq_refcount) != 0) {
2619                 DEBUG_REQ(D_ERROR, request,
2620                           "freeing request with nonzero refcount");
2621                 LBUG();
2622         }
2623
2624         if (request->rq_repbuf)
2625                 sptlrpc_cli_free_repbuf(request);
2626
2627         if (request->rq_import) {
2628                 class_import_put(request->rq_import);
2629                 request->rq_import = NULL;
2630         }
2631         if (request->rq_bulk)
2632                 ptlrpc_free_bulk(request->rq_bulk);
2633
2634         if (request->rq_reqbuf || request->rq_clrbuf)
2635                 sptlrpc_cli_free_reqbuf(request);
2636
2637         if (request->rq_cli_ctx)
2638                 sptlrpc_req_put_ctx(request, !locked);
2639
2640         if (request->rq_pool)
2641                 __ptlrpc_free_req_to_pool(request);
2642         else
2643                 ptlrpc_request_cache_free(request);
2644         EXIT;
2645 }
2646
2647 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2648 /**
2649  * Drop one request reference. Must be called with import imp_lock held.
2650  * When reference count drops to zero, request is freed.
2651  */
2652 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2653 {
2654         assert_spin_locked(&request->rq_import->imp_lock);
2655         (void)__ptlrpc_req_finished(request, 1);
2656 }
2657
2658 /**
2659  * Helper function
2660  * Drops one reference count for request \a request.
2661  * \a locked set indicates that caller holds import imp_lock.
2662  * Frees the request whe reference count reaches zero.
2663  *
2664  * \retval 1    the request is freed
2665  * \retval 0    some others still hold references on the request
2666  */
2667 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2668 {
2669         int count;
2670
2671         ENTRY;
2672         if (!request)
2673                 RETURN(1);
2674
2675         LASSERT(request != LP_POISON);
2676         LASSERT(request->rq_reqmsg != LP_POISON);
2677
2678         DEBUG_REQ(D_INFO, request, "refcount now %u",
2679                   atomic_read(&request->rq_refcount) - 1);
2680
2681         spin_lock(&request->rq_lock);
2682         count = atomic_dec_return(&request->rq_refcount);
2683         LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2684
2685         /*
2686          * For open RPC, the client does not know the EA size (LOV, ACL, and
2687          * so on) before replied, then the client has to reserve very large
2688          * reply buffer. Such buffer will not be released until the RPC freed.
2689          * Since The open RPC is replayable, we need to keep it in the replay
2690          * list until close. If there are a lot of files opened concurrently,
2691          * then the client may be OOM.
2692          *
2693          * If fact, it is unnecessary to keep reply buffer for open replay,
2694          * related EAs have already been saved via mdc_save_lovea() before
2695          * coming here. So it is safe to free the reply buffer some earlier
2696          * before releasing the RPC to avoid client OOM. LU-9514
2697          */
2698         if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2699                 spin_lock(&request->rq_early_free_lock);
2700                 sptlrpc_cli_free_repbuf(request);
2701                 request->rq_repbuf = NULL;
2702                 request->rq_repbuf_len = 0;
2703                 request->rq_repdata = NULL;
2704                 request->rq_reqdata_len = 0;
2705                 spin_unlock(&request->rq_early_free_lock);
2706         }
2707         spin_unlock(&request->rq_lock);
2708
2709         if (!count)
2710                 __ptlrpc_free_req(request, locked);
2711
2712         RETURN(!count);
2713 }
2714
2715 /**
2716  * Drops one reference count for a request.
2717  */
2718 void ptlrpc_req_finished(struct ptlrpc_request *request)
2719 {
2720         __ptlrpc_req_finished(request, 0);
2721 }
2722 EXPORT_SYMBOL(ptlrpc_req_finished);
2723
2724 /**
2725  * Returns xid of a \a request
2726  */
2727 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2728 {
2729         return request->rq_xid;
2730 }
2731 EXPORT_SYMBOL(ptlrpc_req_xid);
2732
2733 /**
2734  * Disengage the client's reply buffer from the network
2735  * NB does _NOT_ unregister any client-side bulk.
2736  * IDEMPOTENT, but _not_ safe against concurrent callers.
2737  * The request owner (i.e. the thread doing the I/O) must call...
2738  * Returns 0 on success or 1 if unregistering cannot be made.
2739  */
2740 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2741 {
2742         /*
2743          * Might sleep.
2744          */
2745         LASSERT(!in_interrupt());
2746
2747         /* Let's setup deadline for reply unlink. */
2748         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2749             async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2750                 request->rq_reply_deadline = ktime_get_real_seconds() +
2751                                              LONG_UNLINK;
2752
2753         /*
2754          * Nothing left to do.
2755          */
2756         if (!ptlrpc_client_recv_or_unlink(request))
2757                 RETURN(1);
2758
2759         LNetMDUnlink(request->rq_reply_md_h);
2760
2761         /*
2762          * Let's check it once again.
2763          */
2764         if (!ptlrpc_client_recv_or_unlink(request))
2765                 RETURN(1);
2766
2767         /* Move to "Unregistering" phase as reply was not unlinked yet. */
2768         ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2769
2770         /*
2771          * Do not wait for unlink to finish.
2772          */
2773         if (async)
2774                 RETURN(0);
2775
2776         /*
2777          * We have to l_wait_event() whatever the result, to give liblustre
2778          * a chance to run reply_in_callback(), and to make sure we've
2779          * unlinked before returning a req to the pool.
2780          */
2781         for (;;) {
2782                 wait_queue_head_t *wq = (request->rq_set) ?
2783                                         &request->rq_set->set_waitq :
2784                                         &request->rq_reply_waitq;
2785                 int seconds = LONG_UNLINK;
2786                 /*
2787                  * Network access will complete in finite time but the HUGE
2788                  * timeout lets us CWARN for visibility of sluggish NALs
2789                  */
2790                 while (seconds > 0 &&
2791                        wait_event_idle_timeout(
2792                                *wq,
2793                                !ptlrpc_client_recv_or_unlink(request),
2794                                cfs_time_seconds(1)) == 0)
2795                         seconds -= 1;
2796                 if (seconds > 0) {
2797                         ptlrpc_rqphase_move(request, request->rq_next_phase);
2798                         RETURN(1);
2799                 }
2800
2801                 DEBUG_REQ(D_WARNING, request,
2802                           "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2803                           request->rq_receiving_reply,
2804                           request->rq_req_unlinked,
2805                           request->rq_reply_unlinked);
2806         }
2807         RETURN(0);
2808 }
2809
2810 static void ptlrpc_free_request(struct ptlrpc_request *req)
2811 {
2812         spin_lock(&req->rq_lock);
2813         req->rq_replay = 0;
2814         spin_unlock(&req->rq_lock);
2815
2816         if (req->rq_commit_cb)
2817                 req->rq_commit_cb(req);
2818         list_del_init(&req->rq_replay_list);
2819
2820         __ptlrpc_req_finished(req, 1);
2821 }
2822
2823 /**
2824  * the request is committed and dropped from the replay list of its import
2825  */
2826 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2827 {
2828         struct obd_import *imp = req->rq_import;
2829
2830         spin_lock(&imp->imp_lock);
2831         if (list_empty(&req->rq_replay_list)) {
2832                 spin_unlock(&imp->imp_lock);
2833                 return;
2834         }
2835
2836         if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2837                 if (imp->imp_replay_cursor == &req->rq_replay_list)
2838                         imp->imp_replay_cursor = req->rq_replay_list.next;
2839                 ptlrpc_free_request(req);
2840         }
2841
2842         spin_unlock(&imp->imp_lock);
2843 }
2844 EXPORT_SYMBOL(ptlrpc_request_committed);
2845
2846 /**
2847  * Iterates through replay_list on import and prunes
2848  * all requests have transno smaller than last_committed for the
2849  * import and don't have rq_replay set.
2850  * Since requests are sorted in transno order, stops when meetign first
2851  * transno bigger than last_committed.
2852  * caller must hold imp->imp_lock
2853  */
2854 void ptlrpc_free_committed(struct obd_import *imp)
2855 {
2856         struct ptlrpc_request *req, *saved;
2857         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2858         bool skip_committed_list = true;
2859
2860         ENTRY;
2861         LASSERT(imp != NULL);
2862         assert_spin_locked(&imp->imp_lock);
2863
2864         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2865             imp->imp_generation == imp->imp_last_generation_checked) {
2866                 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2867                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2868                 RETURN_EXIT;
2869         }
2870         CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2871                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2872                imp->imp_generation);
2873
2874         if (imp->imp_generation != imp->imp_last_generation_checked ||
2875             imp->imp_last_transno_checked == 0)
2876                 skip_committed_list = false;
2877
2878         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2879         imp->imp_last_generation_checked = imp->imp_generation;
2880
2881         list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2882                                  rq_replay_list) {
2883                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
2884                 LASSERT(req != last_req);
2885                 last_req = req;
2886
2887                 if (req->rq_transno == 0) {
2888                         DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2889                         LBUG();
2890                 }
2891                 if (req->rq_import_generation < imp->imp_generation) {
2892                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2893                         GOTO(free_req, 0);
2894                 }
2895
2896                 /* not yet committed */
2897                 if (req->rq_transno > imp->imp_peer_committed_transno) {
2898                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2899                         break;
2900                 }
2901
2902                 if (req->rq_replay) {
2903                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2904                         list_move_tail(&req->rq_replay_list,
2905                                        &imp->imp_committed_list);
2906                         continue;
2907                 }
2908
2909                 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2910                           imp->imp_peer_committed_transno);
2911 free_req:
2912                 ptlrpc_free_request(req);
2913         }
2914
2915         if (skip_committed_list)
2916                 GOTO(out, 0);
2917
2918         list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2919                                  rq_replay_list) {
2920                 LASSERT(req->rq_transno != 0);
2921                 if (req->rq_import_generation < imp->imp_generation ||
2922                     !req->rq_replay) {
2923                         DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2924                                   req->rq_import_generation <
2925                                   imp->imp_generation ? "stale" : "closed");
2926
2927                         if (imp->imp_replay_cursor == &req->rq_replay_list)
2928                                 imp->imp_replay_cursor =
2929                                         req->rq_replay_list.next;
2930
2931                         ptlrpc_free_request(req);
2932                 }
2933         }
2934 out:
2935         EXIT;
2936 }
2937
2938 void ptlrpc_cleanup_client(struct obd_import *imp)
2939 {
2940         ENTRY;
2941         EXIT;
2942 }
2943
2944 /**
2945  * Schedule previously sent request for resend.
2946  * For bulk requests we assign new xid (to avoid problems with
2947  * lost replies and therefore several transfers landing into same buffer
2948  * from different sending attempts).
2949  */
2950 void ptlrpc_resend_req(struct ptlrpc_request *req)
2951 {
2952         DEBUG_REQ(D_HA, req, "going to resend");
2953         spin_lock(&req->rq_lock);
2954
2955         /*
2956          * Request got reply but linked to the import list still.
2957          * Let ptlrpc_check_set() process it.
2958          */
2959         if (ptlrpc_client_replied(req)) {
2960                 spin_unlock(&req->rq_lock);
2961                 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2962                 return;
2963         }
2964
2965         req->rq_status = -EAGAIN;
2966
2967         req->rq_resend = 1;
2968         req->rq_net_err = 0;
2969         req->rq_timedout = 0;
2970
2971         ptlrpc_client_wake_req(req);
2972         spin_unlock(&req->rq_lock);
2973 }
2974
2975 /* XXX: this function and rq_status are currently unused */
2976 void ptlrpc_restart_req(struct ptlrpc_request *req)
2977 {
2978         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2979         req->rq_status = -ERESTARTSYS;
2980
2981         spin_lock(&req->rq_lock);
2982         req->rq_restart = 1;
2983         req->rq_timedout = 0;
2984         ptlrpc_client_wake_req(req);
2985         spin_unlock(&req->rq_lock);
2986 }
2987
2988 /**
2989  * Grab additional reference on a request \a req
2990  */
2991 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2992 {
2993         ENTRY;
2994         atomic_inc(&req->rq_refcount);
2995         RETURN(req);
2996 }
2997 EXPORT_SYMBOL(ptlrpc_request_addref);
2998
2999 /**
3000  * Add a request to import replay_list.
3001  * Must be called under imp_lock
3002  */
3003 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3004                                       struct obd_import *imp)
3005 {
3006         struct list_head *tmp;
3007
3008         assert_spin_locked(&imp->imp_lock);
3009
3010         if (req->rq_transno == 0) {
3011                 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3012                 LBUG();
3013         }
3014
3015         /*
3016          * clear this for new requests that were resent as well
3017          * as resent replayed requests.
3018          */
3019         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3020
3021         /* don't re-add requests that have been replayed */
3022         if (!list_empty(&req->rq_replay_list))
3023                 return;
3024
3025         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3026
3027         spin_lock(&req->rq_lock);
3028         req->rq_resend = 0;
3029         spin_unlock(&req->rq_lock);
3030
3031         LASSERT(imp->imp_replayable);
3032         /* Balanced in ptlrpc_free_committed, usually. */
3033         ptlrpc_request_addref(req);
3034         list_for_each_prev(tmp, &imp->imp_replay_list) {
3035                 struct ptlrpc_request *iter = list_entry(tmp,
3036                                                          struct ptlrpc_request,
3037                                                          rq_replay_list);
3038
3039                 /*
3040                  * We may have duplicate transnos if we create and then
3041                  * open a file, or for closes retained if to match creating
3042                  * opens, so use req->rq_xid as a secondary key.
3043                  * (See bugs 684, 685, and 428.)
3044                  * XXX no longer needed, but all opens need transnos!
3045                  */
3046                 if (iter->rq_transno > req->rq_transno)
3047                         continue;
3048
3049                 if (iter->rq_transno == req->rq_transno) {
3050                         LASSERT(iter->rq_xid != req->rq_xid);
3051                         if (iter->rq_xid > req->rq_xid)
3052                                 continue;
3053                 }
3054
3055                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3056                 return;
3057         }
3058
3059         list_add(&req->rq_replay_list, &imp->imp_replay_list);
3060 }
3061
3062 /**
3063  * Send request and wait until it completes.
3064  * Returns request processing status.
3065  */
3066 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3067 {
3068         struct ptlrpc_request_set *set;
3069         int rc;
3070
3071         ENTRY;
3072         LASSERT(req->rq_set == NULL);
3073         LASSERT(!req->rq_receiving_reply);
3074
3075         set = ptlrpc_prep_set();
3076         if (!set) {
3077                 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3078                 RETURN(-ENOMEM);
3079         }
3080
3081         /* for distributed debugging */
3082         lustre_msg_set_status(req->rq_reqmsg, current_pid());
3083
3084         /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3085         ptlrpc_request_addref(req);
3086         ptlrpc_set_add_req(set, req);
3087         rc = ptlrpc_set_wait(NULL, set);
3088         ptlrpc_set_destroy(set);
3089
3090         RETURN(rc);
3091 }
3092 EXPORT_SYMBOL(ptlrpc_queue_wait);
3093
3094 /**
3095  * Callback used for replayed requests reply processing.
3096  * In case of successful reply calls registered request replay callback.
3097  * In case of error restart replay process.
3098  */
3099 static int ptlrpc_replay_interpret(const struct lu_env *env,
3100                                    struct ptlrpc_request *req,
3101                                    void *args, int rc)
3102 {
3103         struct ptlrpc_replay_async_args *aa = args;
3104         struct obd_import *imp = req->rq_import;
3105
3106         ENTRY;
3107         atomic_dec(&imp->imp_replay_inflight);
3108
3109         /*
3110          * Note: if it is bulk replay (MDS-MDS replay), then even if
3111          * server got the request, but bulk transfer timeout, let's
3112          * replay the bulk req again
3113          */
3114         if (!ptlrpc_client_replied(req) ||
3115             (req->rq_bulk &&
3116              lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3117                 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3118                 GOTO(out, rc = -ETIMEDOUT);
3119         }
3120
3121         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3122             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3123             lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3124                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3125
3126         /** VBR: check version failure */
3127         if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3128                 /** replay was failed due to version mismatch */
3129                 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3130                 spin_lock(&imp->imp_lock);
3131                 imp->imp_vbr_failed = 1;
3132                 spin_unlock(&imp->imp_lock);
3133                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3134         } else {
3135                 /** The transno had better not change over replay. */
3136                 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3137                          lustre_msg_get_transno(req->rq_repmsg) ||
3138                          lustre_msg_get_transno(req->rq_repmsg) == 0,
3139                          "%#llx/%#llx\n",
3140                          lustre_msg_get_transno(req->rq_reqmsg),
3141                          lustre_msg_get_transno(req->rq_repmsg));
3142         }
3143
3144         spin_lock(&imp->imp_lock);
3145         imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3146         spin_unlock(&imp->imp_lock);
3147         LASSERT(imp->imp_last_replay_transno);
3148
3149         /* transaction number shouldn't be bigger than the latest replayed */
3150         if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3151                 DEBUG_REQ(D_ERROR, req,
3152                           "Reported transno=%llu is bigger than replayed=%llu",
3153                           req->rq_transno,
3154                           lustre_msg_get_transno(req->rq_reqmsg));
3155                 GOTO(out, rc = -EINVAL);
3156         }
3157
3158         DEBUG_REQ(D_HA, req, "got reply");
3159
3160         /* let the callback do fixups, possibly including in the request */
3161         if (req->rq_replay_cb)
3162                 req->rq_replay_cb(req);
3163
3164         if (ptlrpc_client_replied(req) &&
3165             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3166                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3167                           lustre_msg_get_status(req->rq_repmsg),
3168                           aa->praa_old_status);
3169
3170                 /*
3171                  * Note: If the replay fails for MDT-MDT recovery, let's
3172                  * abort all of the following requests in the replay
3173                  * and sending list, because MDT-MDT update requests
3174                  * are dependent on each other, see LU-7039
3175                  */
3176                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3177                         struct ptlrpc_request *free_req;
3178                         struct ptlrpc_request *tmp;
3179
3180                         spin_lock(&imp->imp_lock);
3181                         list_for_each_entry_safe(free_req, tmp,
3182                                                  &imp->imp_replay_list,
3183                                                  rq_replay_list) {
3184                                 ptlrpc_free_request(free_req);
3185                         }
3186
3187                         list_for_each_entry_safe(free_req, tmp,
3188                                                  &imp->imp_committed_list,
3189                                                  rq_replay_list) {
3190                                 ptlrpc_free_request(free_req);
3191                         }
3192
3193                         list_for_each_entry_safe(free_req, tmp,
3194                                                  &imp->imp_delayed_list,
3195                                                  rq_list) {
3196                                 spin_lock(&free_req->rq_lock);
3197                                 free_req->rq_err = 1;
3198                                 free_req->rq_status = -EIO;
3199                                 ptlrpc_client_wake_req(free_req);
3200                                 spin_unlock(&free_req->rq_lock);
3201                         }
3202
3203                         list_for_each_entry_safe(free_req, tmp,
3204                                                  &imp->imp_sending_list,