Whamcloud - gitweb
cf2631a08af72fdefe15095184d1c3d8c5cd8d9d
[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 <obd_support.h>
41 #include <obd_class.h>
42 #include <lustre_lib.h>
43 #include <lustre_ha.h>
44 #include <lustre_import.h>
45 #include <lustre_req_layout.h>
46
47 #include "ptlrpc_internal.h"
48
49 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
50                                       struct page *page, int pageoffset,
51                                       int len)
52 {
53         __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
54 }
55
56 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
57                                         struct page *page, int pageoffset,
58                                         int len)
59 {
60         __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
61 }
62
63 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
64 {
65         int i;
66
67         for (i = 0; i < desc->bd_iov_count ; i++)
68                 put_page(BD_GET_KIOV(desc, i).kiov_page);
69 }
70
71 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
72         .add_kiov_frag  = ptlrpc_prep_bulk_page_pin,
73         .release_frags  = ptlrpc_release_bulk_page_pin,
74 };
75 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
76
77 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
78         .add_kiov_frag  = ptlrpc_prep_bulk_page_nopin,
79         .release_frags  = ptlrpc_release_bulk_noop,
80 };
81 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
82
83 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
84         .add_iov_frag = ptlrpc_prep_bulk_frag,
85 };
86 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
87
88 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
89 static int ptlrpcd_check_work(struct ptlrpc_request *req);
90 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
91
92 /**
93  * Initialize passed in client structure \a cl.
94  */
95 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
96                         struct ptlrpc_client *cl)
97 {
98         cl->cli_request_portal = req_portal;
99         cl->cli_reply_portal   = rep_portal;
100         cl->cli_name           = name;
101 }
102 EXPORT_SYMBOL(ptlrpc_init_client);
103
104 /**
105  * Return PortalRPC connection for remore uud \a uuid
106  */
107 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
108                                                     lnet_nid_t nid4refnet)
109 {
110         struct ptlrpc_connection *c;
111         lnet_nid_t self;
112         struct lnet_process_id peer;
113         int err;
114
115         /*
116          * ptlrpc_uuid_to_peer() initializes its 2nd parameter
117          * before accessing its values.
118          */
119         /* coverity[uninit_use_in_call] */
120         peer.nid = nid4refnet;
121         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
122         if (err != 0) {
123                 CNETERR("cannot find peer %s!\n", uuid->uuid);
124                 return NULL;
125         }
126
127         c = ptlrpc_connection_get(peer, self, uuid);
128         if (c) {
129                 memcpy(c->c_remote_uuid.uuid,
130                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
131         }
132
133         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
134
135         return c;
136 }
137
138 /**
139  * Allocate and initialize new bulk descriptor on the sender.
140  * Returns pointer to the descriptor or NULL on error.
141  */
142 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
143                                          unsigned int max_brw,
144                                          enum ptlrpc_bulk_op_type type,
145                                          unsigned int portal,
146                                          const struct ptlrpc_bulk_frag_ops *ops)
147 {
148         struct ptlrpc_bulk_desc *desc;
149         int i;
150
151         /* ensure that only one of KIOV or IOVEC is set but not both */
152         LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
153                  ops->add_kiov_frag != NULL) ||
154                 (ptlrpc_is_bulk_desc_kvec(type) &&
155                  ops->add_iov_frag != NULL));
156
157         OBD_ALLOC_PTR(desc);
158         if (!desc)
159                 return NULL;
160         if (type & PTLRPC_BULK_BUF_KIOV) {
161                 OBD_ALLOC_LARGE(GET_KIOV(desc),
162                                 nfrags * sizeof(*GET_KIOV(desc)));
163                 if (!GET_KIOV(desc))
164                         goto out;
165         } else {
166                 OBD_ALLOC_LARGE(GET_KVEC(desc),
167                                 nfrags * sizeof(*GET_KVEC(desc)));
168                 if (!GET_KVEC(desc))
169                         goto out;
170         }
171
172         spin_lock_init(&desc->bd_lock);
173         init_waitqueue_head(&desc->bd_waitq);
174         desc->bd_max_iov = nfrags;
175         desc->bd_iov_count = 0;
176         desc->bd_portal = portal;
177         desc->bd_type = type;
178         desc->bd_md_count = 0;
179         desc->bd_frag_ops = ops;
180         LASSERT(max_brw > 0);
181         desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
182         /*
183          * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
184          * node. Negotiated ocd_brw_size will always be <= this number.
185          */
186         for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
187                 LNetInvalidateMDHandle(&desc->bd_mds[i]);
188
189         return desc;
190 out:
191         OBD_FREE_PTR(desc);
192         return NULL;
193 }
194
195 /**
196  * Prepare bulk descriptor for specified outgoing request \a req that
197  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
198  * the bulk to be sent. Used on client-side.
199  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
200  * error.
201  */
202 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
203                                               unsigned int nfrags,
204                                               unsigned int max_brw,
205                                               unsigned int type,
206                                               unsigned int portal,
207                                               const struct ptlrpc_bulk_frag_ops
208                                                 *ops)
209 {
210         struct obd_import *imp = req->rq_import;
211         struct ptlrpc_bulk_desc *desc;
212
213         ENTRY;
214         LASSERT(ptlrpc_is_bulk_op_passive(type));
215
216         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
217         if (!desc)
218                 RETURN(NULL);
219
220         desc->bd_import = class_import_get(imp);
221         desc->bd_req = req;
222
223         desc->bd_cbid.cbid_fn  = client_bulk_callback;
224         desc->bd_cbid.cbid_arg = desc;
225
226         /* This makes req own desc, and free it when she frees herself */
227         req->rq_bulk = desc;
228
229         return desc;
230 }
231 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
232
233 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
234                              struct page *page, int pageoffset, int len,
235                              int pin)
236 {
237         lnet_kiov_t *kiov;
238
239         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
240         LASSERT(page != NULL);
241         LASSERT(pageoffset >= 0);
242         LASSERT(len > 0);
243         LASSERT(pageoffset + len <= PAGE_SIZE);
244         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
245
246         kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
247
248         desc->bd_nob += len;
249
250         if (pin)
251                 get_page(page);
252
253         kiov->kiov_page = page;
254         kiov->kiov_offset = pageoffset;
255         kiov->kiov_len = len;
256
257         desc->bd_iov_count++;
258 }
259 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
260
261 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
262                           void *frag, int len)
263 {
264         struct kvec *iovec;
265
266         ENTRY;
267
268         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
269         LASSERT(frag != NULL);
270         LASSERT(len > 0);
271         LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
272
273         iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
274
275         desc->bd_nob += len;
276
277         iovec->iov_base = frag;
278         iovec->iov_len = len;
279
280         desc->bd_iov_count++;
281
282         RETURN(desc->bd_nob);
283 }
284 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
285
286 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
287 {
288         ENTRY;
289
290         LASSERT(desc != NULL);
291         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
292         LASSERT(desc->bd_md_count == 0);         /* network hands off */
293         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
294         LASSERT(desc->bd_frag_ops != NULL);
295
296         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
297                 sptlrpc_enc_pool_put_pages(desc);
298
299         if (desc->bd_export)
300                 class_export_put(desc->bd_export);
301         else
302                 class_import_put(desc->bd_import);
303
304         if (desc->bd_frag_ops->release_frags != NULL)
305                 desc->bd_frag_ops->release_frags(desc);
306
307         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
308                 OBD_FREE_LARGE(GET_KIOV(desc),
309                                desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
310         else
311                 OBD_FREE_LARGE(GET_KVEC(desc),
312                                desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
313         OBD_FREE_PTR(desc);
314         EXIT;
315 }
316 EXPORT_SYMBOL(ptlrpc_free_bulk);
317
318 /**
319  * Set server timelimit for this req, i.e. how long are we willing to wait
320  * for reply before timing out this request.
321  */
322 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
323 {
324         __u32 serv_est;
325         int idx;
326         struct imp_at *at;
327
328         LASSERT(req->rq_import);
329
330         if (AT_OFF) {
331                 /* non-AT settings */
332                 /**
333                  * \a imp_server_timeout means this is reverse import and
334                  * we send (currently only) ASTs to the client and cannot afford
335                  * to wait too long for the reply, otherwise the other client
336                  * (because of which we are sending this request) would
337                  * timeout waiting for us
338                  */
339                 req->rq_timeout = req->rq_import->imp_server_timeout ?
340                                   obd_timeout / 2 : obd_timeout;
341         } else {
342                 at = &req->rq_import->imp_at;
343                 idx = import_at_get_index(req->rq_import,
344                                           req->rq_request_portal);
345                 serv_est = at_get(&at->iat_service_estimate[idx]);
346                 req->rq_timeout = at_est2timeout(serv_est);
347         }
348         /*
349          * We could get even fancier here, using history to predict increased
350          * loading...
351          */
352
353         /*
354          * Let the server know what this RPC timeout is by putting it in the
355          * reqmsg
356          */
357         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
358 }
359 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
360
361 /* Adjust max service estimate based on server value */
362 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
363                                   unsigned int serv_est)
364 {
365         int idx;
366         unsigned int oldse;
367         struct imp_at *at;
368
369         LASSERT(req->rq_import);
370         at = &req->rq_import->imp_at;
371
372         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
373         /*
374          * max service estimates are tracked on the server side,
375          * so just keep minimal history here
376          */
377         oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
378         if (oldse != 0)
379                 CDEBUG(D_ADAPTTO,
380                        "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
381                        req->rq_import->imp_obd->obd_name,
382                        req->rq_request_portal,
383                        oldse, at_get(&at->iat_service_estimate[idx]));
384 }
385
386 /* Expected network latency per remote node (secs) */
387 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
388 {
389         return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
390 }
391
392 /* Adjust expected network latency */
393 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
394                                unsigned int service_time)
395 {
396         unsigned int nl, oldnl;
397         struct imp_at *at;
398         time64_t now = ktime_get_real_seconds();
399
400         LASSERT(req->rq_import);
401
402         if (service_time > now - req->rq_sent + 3) {
403                 /*
404                  * b=16408, however, this can also happen if early reply
405                  * is lost and client RPC is expired and resent, early reply
406                  * or reply of original RPC can still be fit in reply buffer
407                  * of resent RPC, now client is measuring time from the
408                  * resent time, but server sent back service time of original
409                  * RPC.
410                  */
411                 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
412                        D_ADAPTTO : D_WARNING,
413                        "Reported service time %u > total measured time %lld\n",
414                        service_time, now - req->rq_sent);
415                 return;
416         }
417
418         /* Network latency is total time less server processing time */
419         nl = max_t(int, now - req->rq_sent -
420                         service_time, 0) + 1; /* st rounding */
421         at = &req->rq_import->imp_at;
422
423         oldnl = at_measured(&at->iat_net_latency, nl);
424         if (oldnl != 0)
425                 CDEBUG(D_ADAPTTO,
426                        "The network latency for %s (nid %s) has changed from %d to %d\n",
427                        req->rq_import->imp_obd->obd_name,
428                        obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
429                        oldnl, at_get(&at->iat_net_latency));
430 }
431
432 static int unpack_reply(struct ptlrpc_request *req)
433 {
434         int rc;
435
436         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
437                 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
438                 if (rc) {
439                         DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
440                                   rc);
441                         return -EPROTO;
442                 }
443         }
444
445         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
446         if (rc) {
447                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
448                           rc);
449                 return -EPROTO;
450         }
451         return 0;
452 }
453
454 /**
455  * Handle an early reply message, called with the rq_lock held.
456  * If anything goes wrong just ignore it - same as if it never happened
457  */
458 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
459 __must_hold(&req->rq_lock)
460 {
461         struct ptlrpc_request *early_req;
462         time64_t olddl;
463         int rc;
464
465         ENTRY;
466         req->rq_early = 0;
467         spin_unlock(&req->rq_lock);
468
469         rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
470         if (rc) {
471                 spin_lock(&req->rq_lock);
472                 RETURN(rc);
473         }
474
475         rc = unpack_reply(early_req);
476         if (rc != 0) {
477                 sptlrpc_cli_finish_early_reply(early_req);
478                 spin_lock(&req->rq_lock);
479                 RETURN(rc);
480         }
481
482         /*
483          * Use new timeout value just to adjust the local value for this
484          * request, don't include it into at_history. It is unclear yet why
485          * service time increased and should it be counted or skipped, e.g.
486          * that can be recovery case or some error or server, the real reply
487          * will add all new data if it is worth to add.
488          */
489         req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
490         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
491
492         /* Network latency can be adjusted, it is pure network delays */
493         ptlrpc_at_adj_net_latency(req,
494                                   lustre_msg_get_service_time(early_req->rq_repmsg));
495
496         sptlrpc_cli_finish_early_reply(early_req);
497
498         spin_lock(&req->rq_lock);
499         olddl = req->rq_deadline;
500         /*
501          * server assumes it now has rq_timeout from when the request
502          * arrived, so the client should give it at least that long.
503          * since we don't know the arrival time we'll use the original
504          * sent time
505          */
506         req->rq_deadline = req->rq_sent + req->rq_timeout +
507                            ptlrpc_at_get_net_latency(req);
508
509         /* The below message is checked in replay-single.sh test_65{a,b} */
510         /* The below message is checked in sanity-{gss,krb5} test_8 */
511         DEBUG_REQ(D_ADAPTTO, req,
512                   "Early reply #%d, new deadline in %llds (%llds)",
513                   req->rq_early_count,
514                   req->rq_deadline - ktime_get_real_seconds(),
515                   req->rq_deadline - olddl);
516
517         RETURN(rc);
518 }
519
520 static struct kmem_cache *request_cache;
521
522 int ptlrpc_request_cache_init(void)
523 {
524         request_cache = kmem_cache_create("ptlrpc_cache",
525                                           sizeof(struct ptlrpc_request),
526                                           0, SLAB_HWCACHE_ALIGN, NULL);
527         return request_cache ? 0 : -ENOMEM;
528 }
529
530 void ptlrpc_request_cache_fini(void)
531 {
532         kmem_cache_destroy(request_cache);
533 }
534
535 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
536 {
537         struct ptlrpc_request *req;
538
539         OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
540         return req;
541 }
542
543 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
544 {
545         OBD_SLAB_FREE_PTR(req, request_cache);
546 }
547
548 /**
549  * Wind down request pool \a pool.
550  * Frees all requests from the pool too
551  */
552 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
553 {
554         struct list_head *l, *tmp;
555         struct ptlrpc_request *req;
556
557         LASSERT(pool != NULL);
558
559         spin_lock(&pool->prp_lock);
560         list_for_each_safe(l, tmp, &pool->prp_req_list) {
561                 req = list_entry(l, struct ptlrpc_request, rq_list);
562                 list_del(&req->rq_list);
563                 LASSERT(req->rq_reqbuf);
564                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
565                 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
566                 ptlrpc_request_cache_free(req);
567         }
568         spin_unlock(&pool->prp_lock);
569         OBD_FREE(pool, sizeof(*pool));
570 }
571 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
572
573 /**
574  * Allocates, initializes and adds \a num_rq requests to the pool \a pool
575  */
576 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
577 {
578         int i;
579         int size = 1;
580
581         while (size < pool->prp_rq_size)
582                 size <<= 1;
583
584         LASSERTF(list_empty(&pool->prp_req_list) ||
585                  size == pool->prp_rq_size,
586                  "Trying to change pool size with nonempty pool from %d to %d bytes\n",
587                  pool->prp_rq_size, size);
588
589         pool->prp_rq_size = size;
590         for (i = 0; i < num_rq; i++) {
591                 struct ptlrpc_request *req;
592                 struct lustre_msg *msg;
593
594                 req = ptlrpc_request_cache_alloc(GFP_NOFS);
595                 if (!req)
596                         return i;
597                 OBD_ALLOC_LARGE(msg, size);
598                 if (!msg) {
599                         ptlrpc_request_cache_free(req);
600                         return i;
601                 }
602                 req->rq_reqbuf = msg;
603                 req->rq_reqbuf_len = size;
604                 req->rq_pool = pool;
605                 spin_lock(&pool->prp_lock);
606                 list_add_tail(&req->rq_list, &pool->prp_req_list);
607                 spin_unlock(&pool->prp_lock);
608         }
609         return num_rq;
610 }
611 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
612
613 /**
614  * Create and initialize new request pool with given attributes:
615  * \a num_rq - initial number of requests to create for the pool
616  * \a msgsize - maximum message size possible for requests in thid pool
617  * \a populate_pool - function to be called when more requests need to be added
618  *                    to the pool
619  * Returns pointer to newly created pool or NULL on error.
620  */
621 struct ptlrpc_request_pool *
622 ptlrpc_init_rq_pool(int num_rq, int msgsize,
623                     int (*populate_pool)(struct ptlrpc_request_pool *, int))
624 {
625         struct ptlrpc_request_pool *pool;
626
627         OBD_ALLOC_PTR(pool);
628         if (!pool)
629                 return NULL;
630
631         /*
632          * Request next power of two for the allocation, because internally
633          * kernel would do exactly this
634          */
635         spin_lock_init(&pool->prp_lock);
636         INIT_LIST_HEAD(&pool->prp_req_list);
637         pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
638         pool->prp_populate = populate_pool;
639
640         populate_pool(pool, num_rq);
641
642         return pool;
643 }
644 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
645
646 /**
647  * Fetches one request from pool \a pool
648  */
649 static struct ptlrpc_request *
650 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
651 {
652         struct ptlrpc_request *request;
653         struct lustre_msg *reqbuf;
654
655         if (!pool)
656                 return NULL;
657
658         spin_lock(&pool->prp_lock);
659
660         /*
661          * See if we have anything in a pool, and bail out if nothing,
662          * in writeout path, where this matters, this is safe to do, because
663          * nothing is lost in this case, and when some in-flight requests
664          * complete, this code will be called again.
665          */
666         if (unlikely(list_empty(&pool->prp_req_list))) {
667                 spin_unlock(&pool->prp_lock);
668                 return NULL;
669         }
670
671         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
672                              rq_list);
673         list_del_init(&request->rq_list);
674         spin_unlock(&pool->prp_lock);
675
676         LASSERT(request->rq_reqbuf);
677         LASSERT(request->rq_pool);
678
679         reqbuf = request->rq_reqbuf;
680         memset(request, 0, sizeof(*request));
681         request->rq_reqbuf = reqbuf;
682         request->rq_reqbuf_len = pool->prp_rq_size;
683         request->rq_pool = pool;
684
685         return request;
686 }
687
688 /**
689  * Returns freed \a request to pool.
690  */
691 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
692 {
693         struct ptlrpc_request_pool *pool = request->rq_pool;
694
695         spin_lock(&pool->prp_lock);
696         LASSERT(list_empty(&request->rq_list));
697         LASSERT(!request->rq_receiving_reply);
698         list_add_tail(&request->rq_list, &pool->prp_req_list);
699         spin_unlock(&pool->prp_lock);
700 }
701
702 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
703 {
704         struct obd_import *imp = req->rq_import;
705         struct list_head *tmp;
706         struct ptlrpc_request *iter;
707
708         assert_spin_locked(&imp->imp_lock);
709         LASSERT(list_empty(&req->rq_unreplied_list));
710
711         /* unreplied list is sorted by xid in ascending order */
712         list_for_each_prev(tmp, &imp->imp_unreplied_list) {
713                 iter = list_entry(tmp, struct ptlrpc_request,
714                                   rq_unreplied_list);
715
716                 LASSERT(req->rq_xid != iter->rq_xid);
717                 if (req->rq_xid < iter->rq_xid)
718                         continue;
719                 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
720                 return;
721         }
722         list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
723 }
724
725 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
726 {
727         req->rq_xid = ptlrpc_next_xid();
728         ptlrpc_add_unreplied(req);
729 }
730
731 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
732 {
733         spin_lock(&req->rq_import->imp_lock);
734         ptlrpc_assign_next_xid_nolock(req);
735         spin_unlock(&req->rq_import->imp_lock);
736 }
737
738 static atomic64_t ptlrpc_last_xid;
739
740 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
741 {
742         spin_lock(&req->rq_import->imp_lock);
743         list_del_init(&req->rq_unreplied_list);
744         ptlrpc_assign_next_xid_nolock(req);
745         spin_unlock(&req->rq_import->imp_lock);
746         DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
747 }
748
749 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
750 {
751         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
752         __u32 opc;
753         __u16 tag;
754
755         opc = lustre_msg_get_opc(req->rq_reqmsg);
756         tag = obd_get_mod_rpc_slot(cli, opc);
757         lustre_msg_set_tag(req->rq_reqmsg, tag);
758         ptlrpc_reassign_next_xid(req);
759 }
760 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
761
762 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
763 {
764         __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
765
766         if (tag != 0) {
767                 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
768                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
769
770                 obd_put_mod_rpc_slot(cli, opc, tag);
771         }
772 }
773 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
774
775 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
776                              __u32 version, int opcode, char **bufs,
777                              struct ptlrpc_cli_ctx *ctx)
778 {
779         int count;
780         struct obd_import *imp;
781         __u32 *lengths;
782         int rc;
783
784         ENTRY;
785
786         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
787         imp = request->rq_import;
788         lengths = request->rq_pill.rc_area[RCL_CLIENT];
789
790         if (ctx) {
791                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
792         } else {
793                 rc = sptlrpc_req_get_ctx(request);
794                 if (rc)
795                         GOTO(out_free, rc);
796         }
797         sptlrpc_req_set_flavor(request, opcode);
798
799         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
800                                  lengths, bufs);
801         if (rc)
802                 GOTO(out_ctx, rc);
803
804         lustre_msg_add_version(request->rq_reqmsg, version);
805         request->rq_send_state = LUSTRE_IMP_FULL;
806         request->rq_type = PTL_RPC_MSG_REQUEST;
807
808         request->rq_req_cbid.cbid_fn  = request_out_callback;
809         request->rq_req_cbid.cbid_arg = request;
810
811         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
812         request->rq_reply_cbid.cbid_arg = request;
813
814         request->rq_reply_deadline = 0;
815         request->rq_bulk_deadline = 0;
816         request->rq_req_deadline = 0;
817         request->rq_phase = RQ_PHASE_NEW;
818         request->rq_next_phase = RQ_PHASE_UNDEFINED;
819
820         request->rq_request_portal = imp->imp_client->cli_request_portal;
821         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
822
823         ptlrpc_at_set_req_timeout(request);
824
825         lustre_msg_set_opc(request->rq_reqmsg, opcode);
826
827         /* Let's setup deadline for req/reply/bulk unlink for opcode. */
828         if (cfs_fail_val == opcode) {
829                 time64_t *fail_t = NULL, *fail2_t = NULL;
830
831                 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
832                         fail_t = &request->rq_bulk_deadline;
833                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
834                         fail_t = &request->rq_reply_deadline;
835                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
836                         fail_t = &request->rq_req_deadline;
837                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
838                         fail_t = &request->rq_reply_deadline;
839                         fail2_t = &request->rq_bulk_deadline;
840                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
841                         time64_t now = ktime_get_real_seconds();
842                         u64 xid = ((u64)now >> 4) << 24;
843
844                         atomic64_set(&ptlrpc_last_xid, xid);
845                 }
846
847                 if (fail_t) {
848                         *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
849
850                         if (fail2_t)
851                                 *fail2_t = ktime_get_real_seconds() +
852                                            LONG_UNLINK;
853
854                         /*
855                          * The RPC is infected, let the test to change the
856                          * fail_loc
857                          */
858                         msleep(4 * MSEC_PER_SEC);
859                 }
860         }
861         ptlrpc_assign_next_xid(request);
862
863         RETURN(0);
864
865 out_ctx:
866         LASSERT(!request->rq_pool);
867         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
868 out_free:
869         class_import_put(imp);
870
871         return rc;
872 }
873 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
874
875 /**
876  * Pack request buffers for network transfer, performing necessary encryption
877  * steps if necessary.
878  */
879 int ptlrpc_request_pack(struct ptlrpc_request *request,
880                         __u32 version, int opcode)
881 {
882         return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
883 }
884 EXPORT_SYMBOL(ptlrpc_request_pack);
885
886 /**
887  * Helper function to allocate new request on import \a imp
888  * and possibly using existing request from pool \a pool if provided.
889  * Returns allocated request structure with import field filled or
890  * NULL on error.
891  */
892 static inline
893 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
894                                               struct ptlrpc_request_pool *pool)
895 {
896         struct ptlrpc_request *request = NULL;
897
898         request = ptlrpc_request_cache_alloc(GFP_NOFS);
899
900         if (!request && pool)
901                 request = ptlrpc_prep_req_from_pool(pool);
902
903         if (request) {
904                 ptlrpc_cli_req_init(request);
905
906                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
907                 LASSERT(imp != LP_POISON);
908                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
909                          imp->imp_client);
910                 LASSERT(imp->imp_client != LP_POISON);
911
912                 request->rq_import = class_import_get(imp);
913         } else {
914                 CERROR("request allocation out of memory\n");
915         }
916
917         return request;
918 }
919
920 /**
921  * Helper function for creating a request.
922  * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
923  * buffer structures according to capsule template \a format.
924  * Returns allocated request structure pointer or NULL on error.
925  */
926 static struct ptlrpc_request *
927 ptlrpc_request_alloc_internal(struct obd_import *imp,
928                               struct ptlrpc_request_pool *pool,
929                               const struct req_format *format)
930 {
931         struct ptlrpc_request *request;
932
933         request = __ptlrpc_request_alloc(imp, pool);
934         if (!request)
935                 return NULL;
936
937         /*
938          * initiate connection if needed when the import has been
939          * referenced by the new request to avoid races with disconnect
940          */
941         if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
942                 int rc;
943
944                 CDEBUG_LIMIT(imp->imp_idle_debug,
945                              "%s: reconnect after %llds idle\n",
946                              imp->imp_obd->obd_name, ktime_get_real_seconds() -
947                                                      imp->imp_last_reply_time);
948                 spin_lock(&imp->imp_lock);
949                 if (imp->imp_state == LUSTRE_IMP_IDLE) {
950                         imp->imp_generation++;
951                         imp->imp_initiated_at = imp->imp_generation;
952                         imp->imp_state = LUSTRE_IMP_NEW;
953
954                         /* connect_import_locked releases imp_lock */
955                         rc = ptlrpc_connect_import_locked(imp);
956                         if (rc < 0) {
957                                 ptlrpc_request_free(request);
958                                 return NULL;
959                         }
960                         ptlrpc_pinger_add_import(imp);
961                 } else {
962                         spin_unlock(&imp->imp_lock);
963                 }
964         }
965
966         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
967         req_capsule_set(&request->rq_pill, format);
968         return request;
969 }
970
971 /**
972  * Allocate new request structure for import \a imp and initialize its
973  * buffer structure according to capsule template \a format.
974  */
975 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
976                                             const struct req_format *format)
977 {
978         return ptlrpc_request_alloc_internal(imp, NULL, format);
979 }
980 EXPORT_SYMBOL(ptlrpc_request_alloc);
981
982 /**
983  * Allocate new request structure for import \a imp from pool \a pool and
984  * initialize its buffer structure according to capsule template \a format.
985  */
986 struct ptlrpc_request *
987 ptlrpc_request_alloc_pool(struct obd_import *imp,
988                           struct ptlrpc_request_pool *pool,
989                           const struct req_format *format)
990 {
991         return ptlrpc_request_alloc_internal(imp, pool, format);
992 }
993 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
994
995 /**
996  * For requests not from pool, free memory of the request structure.
997  * For requests obtained from a pool earlier, return request back to pool.
998  */
999 void ptlrpc_request_free(struct ptlrpc_request *request)
1000 {
1001         if (request->rq_pool)
1002                 __ptlrpc_free_req_to_pool(request);
1003         else
1004                 ptlrpc_request_cache_free(request);
1005 }
1006 EXPORT_SYMBOL(ptlrpc_request_free);
1007
1008 /**
1009  * Allocate new request for operatione \a opcode and immediatelly pack it for
1010  * network transfer.
1011  * Only used for simple requests like OBD_PING where the only important
1012  * part of the request is operation itself.
1013  * Returns allocated request or NULL on error.
1014  */
1015 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1016                                                  const struct req_format *format,
1017                                                  __u32 version, int opcode)
1018 {
1019         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1020         int rc;
1021
1022         if (req) {
1023                 rc = ptlrpc_request_pack(req, version, opcode);
1024                 if (rc) {
1025                         ptlrpc_request_free(req);
1026                         req = NULL;
1027                 }
1028         }
1029         return req;
1030 }
1031 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1032
1033 /**
1034  * Allocate and initialize new request set structure on the current CPT.
1035  * Returns a pointer to the newly allocated set structure or NULL on error.
1036  */
1037 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1038 {
1039         struct ptlrpc_request_set *set;
1040         int cpt;
1041
1042         ENTRY;
1043         cpt = cfs_cpt_current(cfs_cpt_table, 0);
1044         OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1045         if (!set)
1046                 RETURN(NULL);
1047         atomic_set(&set->set_refcount, 1);
1048         INIT_LIST_HEAD(&set->set_requests);
1049         init_waitqueue_head(&set->set_waitq);
1050         atomic_set(&set->set_new_count, 0);
1051         atomic_set(&set->set_remaining, 0);
1052         spin_lock_init(&set->set_new_req_lock);
1053         INIT_LIST_HEAD(&set->set_new_requests);
1054         set->set_max_inflight = UINT_MAX;
1055         set->set_producer     = NULL;
1056         set->set_producer_arg = NULL;
1057         set->set_rc           = 0;
1058
1059         RETURN(set);
1060 }
1061 EXPORT_SYMBOL(ptlrpc_prep_set);
1062
1063 /**
1064  * Allocate and initialize new request set structure with flow control
1065  * extension. This extension allows to control the number of requests in-flight
1066  * for the whole set. A callback function to generate requests must be provided
1067  * and the request set will keep the number of requests sent over the wire to
1068  * @max_inflight.
1069  * Returns a pointer to the newly allocated set structure or NULL on error.
1070  */
1071 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1072                                              void *arg)
1073
1074 {
1075         struct ptlrpc_request_set *set;
1076
1077         set = ptlrpc_prep_set();
1078         if (!set)
1079                 RETURN(NULL);
1080
1081         set->set_max_inflight  = max;
1082         set->set_producer      = func;
1083         set->set_producer_arg  = arg;
1084
1085         RETURN(set);
1086 }
1087
1088 /**
1089  * Wind down and free request set structure previously allocated with
1090  * ptlrpc_prep_set.
1091  * Ensures that all requests on the set have completed and removes
1092  * all requests from the request list in a set.
1093  * If any unsent request happen to be on the list, pretends that they got
1094  * an error in flight and calls their completion handler.
1095  */
1096 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1097 {
1098         struct list_head *tmp;
1099         struct list_head *next;
1100         int expected_phase;
1101         int n = 0;
1102
1103         ENTRY;
1104
1105         /* Requests on the set should either all be completed, or all be new */
1106         expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1107                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1108         list_for_each(tmp, &set->set_requests) {
1109                 struct ptlrpc_request *req =
1110                         list_entry(tmp, struct ptlrpc_request,
1111                                    rq_set_chain);
1112
1113                 LASSERT(req->rq_phase == expected_phase);
1114                 n++;
1115         }
1116
1117         LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1118                  atomic_read(&set->set_remaining) == n, "%d / %d\n",
1119                  atomic_read(&set->set_remaining), n);
1120
1121         list_for_each_safe(tmp, next, &set->set_requests) {
1122                 struct ptlrpc_request *req =
1123                         list_entry(tmp, struct ptlrpc_request,
1124                                    rq_set_chain);
1125                 list_del_init(&req->rq_set_chain);
1126
1127                 LASSERT(req->rq_phase == expected_phase);
1128
1129                 if (req->rq_phase == RQ_PHASE_NEW) {
1130                         ptlrpc_req_interpret(NULL, req, -EBADR);
1131                         atomic_dec(&set->set_remaining);
1132                 }
1133
1134                 spin_lock(&req->rq_lock);
1135                 req->rq_set = NULL;
1136                 req->rq_invalid_rqset = 0;
1137                 spin_unlock(&req->rq_lock);
1138
1139                 ptlrpc_req_finished(req);
1140         }
1141
1142         LASSERT(atomic_read(&set->set_remaining) == 0);
1143
1144         ptlrpc_reqset_put(set);
1145         EXIT;
1146 }
1147 EXPORT_SYMBOL(ptlrpc_set_destroy);
1148
1149 /**
1150  * Add a new request to the general purpose request set.
1151  * Assumes request reference from the caller.
1152  */
1153 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1154                         struct ptlrpc_request *req)
1155 {
1156         LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1157         LASSERT(list_empty(&req->rq_set_chain));
1158
1159         if (req->rq_allow_intr)
1160                 set->set_allow_intr = 1;
1161
1162         /* The set takes over the caller's request reference */
1163         list_add_tail(&req->rq_set_chain, &set->set_requests);
1164         req->rq_set = set;
1165         atomic_inc(&set->set_remaining);
1166         req->rq_queued_time = ktime_get_seconds();
1167
1168         if (req->rq_reqmsg)
1169                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1170
1171         if (set->set_producer)
1172                 /*
1173                  * If the request set has a producer callback, the RPC must be
1174                  * sent straight away
1175                  */
1176                 ptlrpc_send_new_req(req);
1177 }
1178 EXPORT_SYMBOL(ptlrpc_set_add_req);
1179
1180 /**
1181  * Add a request to a request with dedicated server thread
1182  * and wake the thread to make any necessary processing.
1183  * Currently only used for ptlrpcd.
1184  */
1185 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1186                             struct ptlrpc_request *req)
1187 {
1188         struct ptlrpc_request_set *set = pc->pc_set;
1189         int count, i;
1190
1191         LASSERT(req->rq_set == NULL);
1192         LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1193
1194         spin_lock(&set->set_new_req_lock);
1195         /*
1196          * The set takes over the caller's request reference.
1197          */
1198         req->rq_set = set;
1199         req->rq_queued_time = ktime_get_seconds();
1200         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1201         count = atomic_inc_return(&set->set_new_count);
1202         spin_unlock(&set->set_new_req_lock);
1203
1204         /* Only need to call wakeup once for the first entry. */
1205         if (count == 1) {
1206                 wake_up(&set->set_waitq);
1207
1208                 /*
1209                  * XXX: It maybe unnecessary to wakeup all the partners. But to
1210                  *      guarantee the async RPC can be processed ASAP, we have
1211                  *      no other better choice. It maybe fixed in future.
1212                  */
1213                 for (i = 0; i < pc->pc_npartners; i++)
1214                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1215         }
1216 }
1217
1218 /**
1219  * Based on the current state of the import, determine if the request
1220  * can be sent, is an error, or should be delayed.
1221  *
1222  * Returns true if this request should be delayed. If false, and
1223  * *status is set, then the request can not be sent and *status is the
1224  * error code.  If false and status is 0, then request can be sent.
1225  *
1226  * The imp->imp_lock must be held.
1227  */
1228 static int ptlrpc_import_delay_req(struct obd_import *imp,
1229                                    struct ptlrpc_request *req, int *status)
1230 {
1231         int delay = 0;
1232
1233         ENTRY;
1234         LASSERT(status);
1235         *status = 0;
1236
1237         if (req->rq_ctx_init || req->rq_ctx_fini) {
1238                 /* always allow ctx init/fini rpc go through */
1239         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1240                 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1241                 *status = -EIO;
1242         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1243                 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1244
1245                 /*
1246                  * pings or MDS-equivalent STATFS may safely
1247                  * race with umount
1248                  */
1249                 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1250                           D_HA : D_ERROR, req, "IMP_CLOSED");
1251                 *status = -EIO;
1252         } else if (ptlrpc_send_limit_expired(req)) {
1253                 /* probably doesn't need to be a D_ERROR afterinitial testing */
1254                 DEBUG_REQ(D_HA, req, "send limit expired");
1255                 *status = -ETIMEDOUT;
1256         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1257                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
1258                 ;/* allow CONNECT even if import is invalid */
1259                 if (atomic_read(&imp->imp_inval_count) != 0) {
1260                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1261                         *status = -EIO;
1262                 }
1263         } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1264                 if (!imp->imp_deactive)
1265                         DEBUG_REQ(D_NET, req, "IMP_INVALID");
1266                 *status = -ESHUTDOWN; /* b=12940 */
1267         } else if (req->rq_import_generation != imp->imp_generation) {
1268                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1269                 *status = -EIO;
1270         } else if (req->rq_send_state != imp->imp_state) {
1271                 /* invalidate in progress - any requests should be drop */
1272                 if (atomic_read(&imp->imp_inval_count) != 0) {
1273                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1274                         *status = -EIO;
1275                 } else if (req->rq_no_delay &&
1276                            imp->imp_generation != imp->imp_initiated_at) {
1277                         /* ignore nodelay for requests initiating connections */
1278                         *status = -EWOULDBLOCK;
1279                 } else if (req->rq_allow_replay &&
1280                            (imp->imp_state == LUSTRE_IMP_REPLAY ||
1281                             imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1282                             imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1283                             imp->imp_state == LUSTRE_IMP_RECOVER)) {
1284                         DEBUG_REQ(D_HA, req, "allow during recovery");
1285                 } else {
1286                         delay = 1;
1287                 }
1288         }
1289
1290         RETURN(delay);
1291 }
1292
1293 /**
1294  * Decide if the error message should be printed to the console or not.
1295  * Makes its decision based on request type, status, and failure frequency.
1296  *
1297  * \param[in] req  request that failed and may need a console message
1298  *
1299  * \retval false if no message should be printed
1300  * \retval true  if console message should be printed
1301  */
1302 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1303 {
1304         LASSERT(req->rq_reqmsg != NULL);
1305
1306         /* Suppress particular reconnect errors which are to be expected. */
1307         if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1308                 /* Suppress timed out reconnect requests */
1309                 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1310                     req->rq_timedout)
1311                         return false;
1312
1313                 /*
1314                  * Suppress most unavailable/again reconnect requests, but
1315                  * print occasionally so it is clear client is trying to
1316                  * connect to a server where no target is running.
1317                  */
1318                 if ((err == -ENODEV || err == -EAGAIN) &&
1319                     req->rq_import->imp_conn_cnt % 30 != 20)
1320                         return false;
1321         }
1322
1323         if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1324                 /* -EAGAIN is normal when using POSIX flocks */
1325                 return false;
1326
1327         if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1328             (req->rq_xid & 0xf) != 10)
1329                 /* Suppress most ping requests, they may fail occasionally */
1330                 return false;
1331
1332         return true;
1333 }
1334
1335 /**
1336  * Check request processing status.
1337  * Returns the status.
1338  */
1339 static int ptlrpc_check_status(struct ptlrpc_request *req)
1340 {
1341         int rc;
1342
1343         ENTRY;
1344         rc = lustre_msg_get_status(req->rq_repmsg);
1345         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1346                 struct obd_import *imp = req->rq_import;
1347                 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1348                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1349
1350                 if (ptlrpc_console_allow(req, opc, rc))
1351                         LCONSOLE_ERROR_MSG(0x11,
1352                                            "%s: operation %s to node %s failed: rc = %d\n",
1353                                            imp->imp_obd->obd_name,
1354                                            ll_opcode2str(opc),
1355                                            libcfs_nid2str(nid), rc);
1356                 RETURN(rc < 0 ? rc : -EINVAL);
1357         }
1358
1359         if (rc)
1360                 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1361
1362         RETURN(rc);
1363 }
1364
1365 /**
1366  * save pre-versions of objects into request for replay.
1367  * Versions are obtained from server reply.
1368  * used for VBR.
1369  */
1370 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1371 {
1372         struct lustre_msg *repmsg = req->rq_repmsg;
1373         struct lustre_msg *reqmsg = req->rq_reqmsg;
1374         __u64 *versions = lustre_msg_get_versions(repmsg);
1375
1376         ENTRY;
1377         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1378                 return;
1379
1380         LASSERT(versions);
1381         lustre_msg_set_versions(reqmsg, versions);
1382         CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1383                versions[0], versions[1]);
1384
1385         EXIT;
1386 }
1387
1388 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1389 {
1390         struct ptlrpc_request *req;
1391
1392         assert_spin_locked(&imp->imp_lock);
1393         if (list_empty(&imp->imp_unreplied_list))
1394                 return 0;
1395
1396         req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1397                          rq_unreplied_list);
1398         LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1399
1400         if (imp->imp_known_replied_xid < req->rq_xid - 1)
1401                 imp->imp_known_replied_xid = req->rq_xid - 1;
1402
1403         return req->rq_xid - 1;
1404 }
1405
1406 /**
1407  * Callback function called when client receives RPC reply for \a req.
1408  * Returns 0 on success or error code.
1409  * The return alue would be assigned to req->rq_status by the caller
1410  * as request processing status.
1411  * This function also decides if the request needs to be saved for later replay.
1412  */
1413 static int after_reply(struct ptlrpc_request *req)
1414 {
1415         struct obd_import *imp = req->rq_import;
1416         struct obd_device *obd = req->rq_import->imp_obd;
1417         ktime_t work_start;
1418         u64 committed;
1419         s64 timediff;
1420         int rc;
1421
1422         ENTRY;
1423         LASSERT(obd != NULL);
1424         /* repbuf must be unlinked */
1425         LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1426
1427         if (req->rq_reply_truncated) {
1428                 if (ptlrpc_no_resend(req)) {
1429                         DEBUG_REQ(D_ERROR, req,
1430                                   "reply buffer overflow, expected=%d, actual size=%d",
1431                                   req->rq_nob_received, req->rq_repbuf_len);
1432                         RETURN(-EOVERFLOW);
1433                 }
1434
1435                 sptlrpc_cli_free_repbuf(req);
1436                 /*
1437                  * Pass the required reply buffer size (include
1438                  * space for early reply).
1439                  * NB: no need to roundup because alloc_repbuf
1440                  * will roundup it
1441                  */
1442                 req->rq_replen = req->rq_nob_received;
1443                 req->rq_nob_received = 0;
1444                 spin_lock(&req->rq_lock);
1445                 req->rq_resend       = 1;
1446                 spin_unlock(&req->rq_lock);
1447                 RETURN(0);
1448         }
1449
1450         work_start = ktime_get_real();
1451         timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1452
1453         /*
1454          * NB Until this point, the whole of the incoming message,
1455          * including buflens, status etc is in the sender's byte order.
1456          */
1457         rc = sptlrpc_cli_unwrap_reply(req);
1458         if (rc) {
1459                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1460                 RETURN(rc);
1461         }
1462
1463         /*
1464          * Security layer unwrap might ask resend this request.
1465          */
1466         if (req->rq_resend)
1467                 RETURN(0);
1468
1469         rc = unpack_reply(req);
1470         if (rc)
1471                 RETURN(rc);
1472
1473         /* retry indefinitely on EINPROGRESS */
1474         if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1475             ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1476                 time64_t now = ktime_get_real_seconds();
1477
1478                 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1479                           D_RPCTRACE, req, "resending request on EINPROGRESS");
1480                 spin_lock(&req->rq_lock);
1481                 req->rq_resend = 1;
1482                 spin_unlock(&req->rq_lock);
1483                 req->rq_nr_resend++;
1484
1485                 /* Readjust the timeout for current conditions */
1486                 ptlrpc_at_set_req_timeout(req);
1487                 /*
1488                  * delay resend to give a chance to the server to get ready.
1489                  * The delay is increased by 1s on every resend and is capped to
1490                  * the current request timeout (i.e. obd_timeout if AT is off,
1491                  * or AT service time x 125% + 5s, see at_est2timeout)
1492                  */
1493                 if (req->rq_nr_resend > req->rq_timeout)
1494                         req->rq_sent = now + req->rq_timeout;
1495                 else
1496                         req->rq_sent = now + req->rq_nr_resend;
1497
1498                 /* Resend for EINPROGRESS will use a new XID */
1499                 spin_lock(&imp->imp_lock);
1500                 list_del_init(&req->rq_unreplied_list);
1501                 spin_unlock(&imp->imp_lock);
1502
1503                 RETURN(0);
1504         }
1505
1506         if (obd->obd_svc_stats) {
1507                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1508                                     timediff);
1509                 ptlrpc_lprocfs_rpc_sent(req, timediff);
1510         }
1511
1512         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1513             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1514                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1515                           lustre_msg_get_type(req->rq_repmsg));
1516                 RETURN(-EPROTO);
1517         }
1518
1519         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1520                 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1521         ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1522         ptlrpc_at_adj_net_latency(req,
1523                                   lustre_msg_get_service_time(req->rq_repmsg));
1524
1525         rc = ptlrpc_check_status(req);
1526
1527         if (rc) {
1528                 /*
1529                  * Either we've been evicted, or the server has failed for
1530                  * some reason. Try to reconnect, and if that fails, punt to
1531                  * the upcall.
1532                  */
1533                 if (ptlrpc_recoverable_error(rc)) {
1534                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
1535                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1536                                 RETURN(rc);
1537                         }
1538                         ptlrpc_request_handle_notconn(req);
1539                         RETURN(rc);
1540                 }
1541         } else {
1542                 /*
1543                  * Let's look if server sent slv. Do it only for RPC with
1544                  * rc == 0.
1545                  */
1546                 ldlm_cli_update_pool(req);
1547         }
1548
1549         /*
1550          * Store transno in reqmsg for replay.
1551          */
1552         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1553                 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1554                 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1555         }
1556
1557         if (imp->imp_replayable) {
1558                 spin_lock(&imp->imp_lock);
1559                 /*
1560                  * No point in adding already-committed requests to the replay
1561                  * list, we will just remove them immediately. b=9829
1562                  */
1563                 if (req->rq_transno != 0 &&
1564                     (req->rq_transno >
1565                      lustre_msg_get_last_committed(req->rq_repmsg) ||
1566                      req->rq_replay)) {
1567                         /** version recovery */
1568                         ptlrpc_save_versions(req);
1569                         ptlrpc_retain_replayable_request(req, imp);
1570                 } else if (req->rq_commit_cb &&
1571                            list_empty(&req->rq_replay_list)) {
1572                         /*
1573                          * NB: don't call rq_commit_cb if it's already on
1574                          * rq_replay_list, ptlrpc_free_committed() will call
1575                          * it later, see LU-3618 for details
1576                          */
1577                         spin_unlock(&imp->imp_lock);
1578                         req->rq_commit_cb(req);
1579                         spin_lock(&imp->imp_lock);
1580                 }
1581
1582                 /*
1583                  * Replay-enabled imports return commit-status information.
1584                  */
1585                 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1586                 if (likely(committed > imp->imp_peer_committed_transno))
1587                         imp->imp_peer_committed_transno = committed;
1588
1589                 ptlrpc_free_committed(imp);
1590
1591                 if (!list_empty(&imp->imp_replay_list)) {
1592                         struct ptlrpc_request *last;
1593
1594                         last = list_entry(imp->imp_replay_list.prev,
1595                                           struct ptlrpc_request,
1596                                           rq_replay_list);
1597                         /*
1598                          * Requests with rq_replay stay on the list even if no
1599                          * commit is expected.
1600                          */
1601                         if (last->rq_transno > imp->imp_peer_committed_transno)
1602                                 ptlrpc_pinger_commit_expected(imp);
1603                 }
1604
1605                 spin_unlock(&imp->imp_lock);
1606         }
1607
1608         RETURN(rc);
1609 }
1610
1611 /**
1612  * Helper function to send request \a req over the network for the first time
1613  * Also adjusts request phase.
1614  * Returns 0 on success or error code.
1615  */
1616 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1617 {
1618         struct obd_import *imp = req->rq_import;
1619         __u64 min_xid = 0;
1620         int rc;
1621
1622         ENTRY;
1623         LASSERT(req->rq_phase == RQ_PHASE_NEW);
1624
1625         /* do not try to go further if there is not enough memory in enc_pool */
1626         if (req->rq_sent && req->rq_bulk)
1627                 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1628                     pool_is_at_full_capacity())
1629                         RETURN(-ENOMEM);
1630
1631         if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1632             (!req->rq_generation_set ||
1633              req->rq_import_generation == imp->imp_generation))
1634                 RETURN(0);
1635
1636         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1637
1638         spin_lock(&imp->imp_lock);
1639
1640         LASSERT(req->rq_xid != 0);
1641         LASSERT(!list_empty(&req->rq_unreplied_list));
1642
1643         if (!req->rq_generation_set)
1644                 req->rq_import_generation = imp->imp_generation;
1645
1646         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1647                 spin_lock(&req->rq_lock);
1648                 req->rq_waiting = 1;
1649                 spin_unlock(&req->rq_lock);
1650
1651                 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1652                           ptlrpc_import_state_name(req->rq_send_state),
1653                           ptlrpc_import_state_name(imp->imp_state));
1654                 LASSERT(list_empty(&req->rq_list));
1655                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1656                 atomic_inc(&req->rq_import->imp_inflight);
1657                 spin_unlock(&imp->imp_lock);
1658                 RETURN(0);
1659         }
1660
1661         if (rc != 0) {
1662                 spin_unlock(&imp->imp_lock);
1663                 req->rq_status = rc;
1664                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1665                 RETURN(rc);
1666         }
1667
1668         LASSERT(list_empty(&req->rq_list));
1669         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1670         atomic_inc(&req->rq_import->imp_inflight);
1671
1672         /*
1673          * find the known replied XID from the unreplied list, CONNECT
1674          * and DISCONNECT requests are skipped to make the sanity check
1675          * on server side happy. see process_req_last_xid().
1676          *
1677          * For CONNECT: Because replay requests have lower XID, it'll
1678          * break the sanity check if CONNECT bump the exp_last_xid on
1679          * server.
1680          *
1681          * For DISCONNECT: Since client will abort inflight RPC before
1682          * sending DISCONNECT, DISCONNECT may carry an XID which higher
1683          * than the inflight RPC.
1684          */
1685         if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1686                 min_xid = ptlrpc_known_replied_xid(imp);
1687         spin_unlock(&imp->imp_lock);
1688
1689         lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1690
1691         lustre_msg_set_status(req->rq_reqmsg, current_pid());
1692
1693         rc = sptlrpc_req_refresh_ctx(req, -1);
1694         if (rc) {
1695                 if (req->rq_err) {
1696                         req->rq_status = rc;
1697                         RETURN(1);
1698                 } else {
1699                         spin_lock(&req->rq_lock);
1700                         req->rq_wait_ctx = 1;
1701                         spin_unlock(&req->rq_lock);
1702                         RETURN(0);
1703                 }
1704         }
1705
1706         CDEBUG(D_RPCTRACE,
1707                "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1708                req, current_comm(),
1709                imp->imp_obd->obd_uuid.uuid,
1710                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1711                obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1712                lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1713
1714         rc = ptl_send_rpc(req, 0);
1715         if (rc == -ENOMEM) {
1716                 spin_lock(&imp->imp_lock);
1717                 if (!list_empty(&req->rq_list)) {
1718                         list_del_init(&req->rq_list);
1719                         atomic_dec(&req->rq_import->imp_inflight);
1720                 }
1721                 spin_unlock(&imp->imp_lock);
1722                 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1723                 RETURN(rc);
1724         }
1725         if (rc) {
1726                 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1727                           rc);
1728                 spin_lock(&req->rq_lock);
1729                 req->rq_net_err = 1;
1730                 spin_unlock(&req->rq_lock);
1731                 RETURN(rc);
1732         }
1733         RETURN(0);
1734 }
1735
1736 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1737 {
1738         int remaining, rc;
1739
1740         ENTRY;
1741         LASSERT(set->set_producer != NULL);
1742
1743         remaining = atomic_read(&set->set_remaining);
1744
1745         /*
1746          * populate the ->set_requests list with requests until we
1747          * reach the maximum number of RPCs in flight for this set
1748          */
1749         while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1750                 rc = set->set_producer(set, set->set_producer_arg);
1751                 if (rc == -ENOENT) {
1752                         /* no more RPC to produce */
1753                         set->set_producer     = NULL;
1754                         set->set_producer_arg = NULL;
1755                         RETURN(0);
1756                 }
1757         }
1758
1759         RETURN((atomic_read(&set->set_remaining) - remaining));
1760 }
1761
1762 /**
1763  * this sends any unsent RPCs in \a set and returns 1 if all are sent
1764  * and no more replies are expected.
1765  * (it is possible to get less replies than requests sent e.g. due to timed out
1766  * requests or requests that we had trouble to send out)
1767  *
1768  * NOTE: This function contains a potential schedule point (cond_resched()).
1769  */
1770 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1771 {
1772         struct list_head *tmp, *next;
1773         struct list_head  comp_reqs;
1774         int force_timer_recalc = 0;
1775
1776         ENTRY;
1777         if (atomic_read(&set->set_remaining) == 0)
1778                 RETURN(1);
1779
1780         INIT_LIST_HEAD(&comp_reqs);
1781         list_for_each_safe(tmp, next, &set->set_requests) {
1782                 struct ptlrpc_request *req =
1783                         list_entry(tmp, struct ptlrpc_request,
1784                                    rq_set_chain);
1785                 struct obd_import *imp = req->rq_import;
1786                 int unregistered = 0;
1787                 int async = 1;
1788                 int rc = 0;
1789
1790                 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1791                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1792                         continue;
1793                 }
1794
1795                 /*
1796                  * This schedule point is mainly for the ptlrpcd caller of this
1797                  * function.  Most ptlrpc sets are not long-lived and unbounded
1798                  * in length, but at the least the set used by the ptlrpcd is.
1799                  * Since the processing time is unbounded, we need to insert an
1800                  * explicit schedule point to make the thread well-behaved.
1801                  */
1802                 cond_resched();
1803
1804                 /*
1805                  * If the caller requires to allow to be interpreted by force
1806                  * and it has really been interpreted, then move the request
1807                  * to RQ_PHASE_INTERPRET phase in spite of what the current
1808                  * phase is.
1809                  */
1810                 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1811                         req->rq_status = -EINTR;
1812                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1813
1814                         /*
1815                          * Since it is interpreted and we have to wait for
1816                          * the reply to be unlinked, then use sync mode.
1817                          */
1818                         async = 0;
1819
1820                         GOTO(interpret, req->rq_status);
1821                 }
1822
1823                 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1824                         force_timer_recalc = 1;
1825
1826                 /* delayed send - skip */
1827                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1828                         continue;
1829
1830                 /* delayed resend - skip */
1831                 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1832                     req->rq_sent > ktime_get_real_seconds())
1833                         continue;
1834
1835                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1836                       req->rq_phase == RQ_PHASE_BULK ||
1837                       req->rq_phase == RQ_PHASE_INTERPRET ||
1838                       req->rq_phase == RQ_PHASE_UNREG_RPC ||
1839                       req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1840                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1841                         LBUG();
1842                 }
1843
1844                 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1845                     req->rq_phase == RQ_PHASE_UNREG_BULK) {
1846                         LASSERT(req->rq_next_phase != req->rq_phase);
1847                         LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1848
1849                         if (req->rq_req_deadline &&
1850                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1851                                 req->rq_req_deadline = 0;
1852                         if (req->rq_reply_deadline &&
1853                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1854                                 req->rq_reply_deadline = 0;
1855                         if (req->rq_bulk_deadline &&
1856                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1857                                 req->rq_bulk_deadline = 0;
1858
1859                         /*
1860                          * Skip processing until reply is unlinked. We
1861                          * can't return to pool before that and we can't
1862                          * call interpret before that. We need to make
1863                          * sure that all rdma transfers finished and will
1864                          * not corrupt any data.
1865                          */
1866                         if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1867                             ptlrpc_client_recv_or_unlink(req))
1868                                 continue;
1869                         if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1870                             ptlrpc_client_bulk_active(req))
1871                                 continue;
1872
1873                         /*
1874                          * Turn fail_loc off to prevent it from looping
1875                          * forever.
1876                          */
1877                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1878                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1879                                                      OBD_FAIL_ONCE);
1880                         }
1881                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1882                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1883                                                      OBD_FAIL_ONCE);
1884                         }
1885
1886                         /*
1887                          * Move to next phase if reply was successfully
1888                          * unlinked.
1889                          */
1890                         ptlrpc_rqphase_move(req, req->rq_next_phase);
1891                 }
1892
1893                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1894                         GOTO(interpret, req->rq_status);
1895
1896                 /*
1897                  * Note that this also will start async reply unlink.
1898                  */
1899                 if (req->rq_net_err && !req->rq_timedout) {
1900                         ptlrpc_expire_one_request(req, 1);
1901
1902                         /*
1903                          * Check if we still need to wait for unlink.
1904                          */
1905                         if (ptlrpc_client_recv_or_unlink(req) ||
1906                             ptlrpc_client_bulk_active(req))
1907                                 continue;
1908                         /* If there is no need to resend, fail it now. */
1909                         if (req->rq_no_resend) {
1910                                 if (req->rq_status == 0)
1911                                         req->rq_status = -EIO;
1912                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1913                                 GOTO(interpret, req->rq_status);
1914                         } else {
1915                                 continue;
1916                         }
1917                 }
1918
1919                 if (req->rq_err) {
1920                         spin_lock(&req->rq_lock);
1921                         req->rq_replied = 0;
1922                         spin_unlock(&req->rq_lock);
1923                         if (req->rq_status == 0)
1924                                 req->rq_status = -EIO;
1925                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1926                         GOTO(interpret, req->rq_status);
1927                 }
1928
1929                 /*
1930                  * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1931                  * so it sets rq_intr regardless of individual rpc
1932                  * timeouts. The synchronous IO waiting path sets
1933                  * rq_intr irrespective of whether ptlrpcd
1934                  * has seen a timeout.  Our policy is to only interpret
1935                  * interrupted rpcs after they have timed out, so we
1936                  * need to enforce that here.
1937                  */
1938
1939                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1940                                      req->rq_wait_ctx)) {
1941                         req->rq_status = -EINTR;
1942                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1943                         GOTO(interpret, req->rq_status);
1944                 }
1945
1946                 if (req->rq_phase == RQ_PHASE_RPC) {
1947                         if (req->rq_timedout || req->rq_resend ||
1948                             req->rq_waiting || req->rq_wait_ctx) {
1949                                 int status;
1950
1951                                 if (!ptlrpc_unregister_reply(req, 1)) {
1952                                         ptlrpc_unregister_bulk(req, 1);
1953                                         continue;
1954                                 }
1955
1956                                 spin_lock(&imp->imp_lock);
1957                                 if (ptlrpc_import_delay_req(imp, req,
1958                                                             &status)) {
1959                                         /*
1960                                          * put on delay list - only if we wait
1961                                          * recovery finished - before send
1962                                          */
1963                                         list_move_tail(&req->rq_list,
1964                                                        &imp->imp_delayed_list);
1965                                         spin_unlock(&imp->imp_lock);
1966                                         continue;
1967                                 }
1968
1969                                 if (status != 0)  {
1970                                         req->rq_status = status;
1971                                         ptlrpc_rqphase_move(req,
1972                                                             RQ_PHASE_INTERPRET);
1973                                         spin_unlock(&imp->imp_lock);
1974                                         GOTO(interpret, req->rq_status);
1975                                 }
1976                                 /* ignore on just initiated connections */
1977                                 if (ptlrpc_no_resend(req) &&
1978                                     !req->rq_wait_ctx &&
1979                                     imp->imp_generation !=
1980                                     imp->imp_initiated_at) {
1981                                         req->rq_status = -ENOTCONN;
1982                                         ptlrpc_rqphase_move(req,
1983                                                             RQ_PHASE_INTERPRET);
1984                                         spin_unlock(&imp->imp_lock);
1985                                         GOTO(interpret, req->rq_status);
1986                                 }
1987
1988                                 list_move_tail(&req->rq_list,
1989                                                &imp->imp_sending_list);
1990
1991                                 spin_unlock(&imp->imp_lock);
1992
1993                                 spin_lock(&req->rq_lock);
1994                                 req->rq_waiting = 0;
1995                                 spin_unlock(&req->rq_lock);
1996
1997                                 if (req->rq_timedout || req->rq_resend) {
1998                                         /*
1999                                          * This is re-sending anyways,
2000                                          * let's mark req as resend.
2001                                          */
2002                                         spin_lock(&req->rq_lock);
2003                                         req->rq_resend = 1;
2004                                         spin_unlock(&req->rq_lock);
2005                                 }
2006                                 /*
2007                                  * rq_wait_ctx is only touched by ptlrpcd,
2008                                  * so no lock is needed here.
2009                                  */
2010                                 status = sptlrpc_req_refresh_ctx(req, -1);
2011                                 if (status) {
2012                                         if (req->rq_err) {
2013                                                 req->rq_status = status;
2014                                                 spin_lock(&req->rq_lock);
2015                                                 req->rq_wait_ctx = 0;
2016                                                 spin_unlock(&req->rq_lock);
2017                                                 force_timer_recalc = 1;
2018                                         } else {
2019                                                 spin_lock(&req->rq_lock);
2020                                                 req->rq_wait_ctx = 1;
2021                                                 spin_unlock(&req->rq_lock);
2022                                         }
2023
2024                                         continue;
2025                                 } else {
2026                                         spin_lock(&req->rq_lock);
2027                                         req->rq_wait_ctx = 0;
2028                                         spin_unlock(&req->rq_lock);
2029                                 }
2030
2031                                 /*
2032                                  * In any case, the previous bulk should be
2033                                  * cleaned up to prepare for the new sending
2034                                  */
2035                                 if (req->rq_bulk &&
2036                                     !ptlrpc_unregister_bulk(req, 1))
2037                                         continue;
2038
2039                                 rc = ptl_send_rpc(req, 0);
2040                                 if (rc == -ENOMEM) {
2041                                         spin_lock(&imp->imp_lock);
2042                                         if (!list_empty(&req->rq_list))
2043                                                 list_del_init(&req->rq_list);
2044                                         spin_unlock(&imp->imp_lock);
2045                                         ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2046                                         continue;
2047                                 }
2048                                 if (rc) {
2049                                         DEBUG_REQ(D_HA, req,
2050                                                   "send failed: rc = %d", rc);
2051                                         force_timer_recalc = 1;
2052                                         spin_lock(&req->rq_lock);
2053                                         req->rq_net_err = 1;
2054                                         spin_unlock(&req->rq_lock);
2055                                         continue;
2056                                 }
2057                                 /* need to reset the timeout */
2058                                 force_timer_recalc = 1;
2059                         }
2060
2061                         spin_lock(&req->rq_lock);
2062
2063                         if (ptlrpc_client_early(req)) {
2064                                 ptlrpc_at_recv_early_reply(req);
2065                                 spin_unlock(&req->rq_lock);
2066                                 continue;
2067                         }
2068
2069                         /* Still waiting for a reply? */
2070                         if (ptlrpc_client_recv(req)) {
2071                                 spin_unlock(&req->rq_lock);
2072                                 continue;
2073                         }
2074
2075                         /* Did we actually receive a reply? */
2076                         if (!ptlrpc_client_replied(req)) {
2077                                 spin_unlock(&req->rq_lock);
2078                                 continue;
2079                         }
2080
2081                         spin_unlock(&req->rq_lock);
2082
2083                         /*
2084                          * unlink from net because we are going to
2085                          * swab in-place of reply buffer
2086                          */
2087                         unregistered = ptlrpc_unregister_reply(req, 1);
2088                         if (!unregistered)
2089                                 continue;
2090
2091                         req->rq_status = after_reply(req);
2092                         if (req->rq_resend)
2093                                 continue;
2094
2095                         /*
2096                          * If there is no bulk associated with this request,
2097                          * then we're done and should let the interpreter
2098                          * process the reply. Similarly if the RPC returned
2099                          * an error, and therefore the bulk will never arrive.
2100                          */
2101                         if (!req->rq_bulk || req->rq_status < 0) {
2102                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2103                                 GOTO(interpret, req->rq_status);
2104                         }
2105
2106                         ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2107                 }
2108
2109                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2110                 if (ptlrpc_client_bulk_active(req))
2111                         continue;
2112
2113                 if (req->rq_bulk->bd_failure) {
2114                         /*
2115                          * The RPC reply arrived OK, but the bulk screwed
2116                          * up!  Dead weird since the server told us the RPC
2117                          * was good after getting the REPLY for her GET or
2118                          * the ACK for her PUT.
2119                          */
2120                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2121                         req->rq_status = -EIO;
2122                 }
2123
2124                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2125
2126 interpret:
2127                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2128
2129                 /*
2130                  * This moves to "unregistering" phase we need to wait for
2131                  * reply unlink.
2132                  */
2133                 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2134                         /* start async bulk unlink too */
2135                         ptlrpc_unregister_bulk(req, 1);
2136                         continue;
2137                 }
2138
2139                 if (!ptlrpc_unregister_bulk(req, async))
2140                         continue;
2141
2142                 /*
2143                  * When calling interpret receiving already should be
2144                  * finished.
2145                  */
2146                 LASSERT(!req->rq_receiving_reply);
2147
2148                 ptlrpc_req_interpret(env, req, req->rq_status);
2149
2150                 if (ptlrpcd_check_work(req)) {
2151                         atomic_dec(&set->set_remaining);
2152                         continue;
2153                 }
2154                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2155
2156                 if (req->rq_reqmsg)
2157                         CDEBUG(D_RPCTRACE,
2158                                "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2159                                req, current_comm(),
2160                                imp->imp_obd->obd_uuid.uuid,
2161                                lustre_msg_get_status(req->rq_reqmsg),
2162                                req->rq_xid,
2163                                obd_import_nid2str(imp),
2164                                lustre_msg_get_opc(req->rq_reqmsg),
2165                                lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2166
2167                 spin_lock(&imp->imp_lock);
2168                 /*
2169                  * Request already may be not on sending or delaying list. This
2170                  * may happen in the case of marking it erroneous for the case
2171                  * ptlrpc_import_delay_req(req, status) find it impossible to
2172                  * allow sending this rpc and returns *status != 0.
2173                  */
2174                 if (!list_empty(&req->rq_list)) {
2175                         list_del_init(&req->rq_list);
2176                         atomic_dec(&imp->imp_inflight);
2177                 }
2178                 list_del_init(&req->rq_unreplied_list);
2179                 spin_unlock(&imp->imp_lock);
2180
2181                 atomic_dec(&set->set_remaining);
2182                 wake_up_all(&imp->imp_recovery_waitq);
2183
2184                 if (set->set_producer) {
2185                         /* produce a new request if possible */
2186                         if (ptlrpc_set_producer(set) > 0)
2187                                 force_timer_recalc = 1;
2188
2189                         /*
2190                          * free the request that has just been completed
2191                          * in order not to pollute set->set_requests
2192                          */
2193                         list_del_init(&req->rq_set_chain);
2194                         spin_lock(&req->rq_lock);
2195                         req->rq_set = NULL;
2196                         req->rq_invalid_rqset = 0;
2197                         spin_unlock(&req->rq_lock);
2198
2199                         /* record rq_status to compute the final status later */
2200                         if (req->rq_status != 0)
2201                                 set->set_rc = req->rq_status;
2202                         ptlrpc_req_finished(req);
2203                 } else {
2204                         list_move_tail(&req->rq_set_chain, &comp_reqs);
2205                 }
2206         }
2207
2208         /*
2209          * move completed request at the head of list so it's easier for
2210          * caller to find them
2211          */
2212         list_splice(&comp_reqs, &set->set_requests);
2213
2214         /* If we hit an error, we want to recover promptly. */
2215         RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2216 }
2217 EXPORT_SYMBOL(ptlrpc_check_set);
2218
2219 /**
2220  * Time out request \a req. is \a async_unlink is set, that means do not wait
2221  * until LNet actually confirms network buffer unlinking.
2222  * Return 1 if we should give up further retrying attempts or 0 otherwise.
2223  */
2224 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2225 {
2226         struct obd_import *imp = req->rq_import;
2227         unsigned int debug_mask = D_RPCTRACE;
2228         int rc = 0;
2229
2230         ENTRY;
2231         spin_lock(&req->rq_lock);
2232         req->rq_timedout = 1;
2233         spin_unlock(&req->rq_lock);
2234
2235         if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2236                                  lustre_msg_get_status(req->rq_reqmsg)))
2237                 debug_mask = D_WARNING;
2238         DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2239                   req->rq_net_err ? "failed due to network error" :
2240                      ((req->rq_real_sent == 0 ||
2241                        req->rq_real_sent < req->rq_sent ||
2242                        req->rq_real_sent >= req->rq_deadline) ?
2243                       "timed out for sent delay" : "timed out for slow reply"),
2244                   (s64)req->rq_sent, (s64)req->rq_real_sent);
2245
2246         if (imp && obd_debug_peer_on_timeout)
2247                 LNetDebugPeer(imp->imp_connection->c_peer);
2248
2249         ptlrpc_unregister_reply(req, async_unlink);
2250         ptlrpc_unregister_bulk(req, async_unlink);
2251
2252         if (obd_dump_on_timeout)
2253                 libcfs_debug_dumplog();
2254
2255         if (!imp) {
2256                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2257                 RETURN(1);
2258         }
2259
2260         atomic_inc(&imp->imp_timeouts);
2261
2262         /* The DLM server doesn't want recovery run on its imports. */
2263         if (imp->imp_dlm_fake)
2264                 RETURN(1);
2265
2266         /*
2267          * If this request is for recovery or other primordial tasks,
2268          * then error it out here.
2269          */
2270         if (req->rq_ctx_init || req->rq_ctx_fini ||
2271             req->rq_send_state != LUSTRE_IMP_FULL ||
2272             imp->imp_obd->obd_no_recov) {
2273                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2274                           ptlrpc_import_state_name(req->rq_send_state),
2275                           ptlrpc_import_state_name(imp->imp_state));
2276                 spin_lock(&req->rq_lock);
2277                 req->rq_status = -ETIMEDOUT;
2278                 req->rq_err = 1;
2279                 spin_unlock(&req->rq_lock);
2280                 RETURN(1);
2281         }
2282
2283         /*
2284          * if a request can't be resent we can't wait for an answer after
2285          * the timeout
2286          */
2287         if (ptlrpc_no_resend(req)) {
2288                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2289                 rc = 1;
2290         }
2291
2292         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2293
2294         RETURN(rc);
2295 }
2296
2297 /**
2298  * Time out all uncompleted requests in request set pointed by \a data
2299  * Callback used when waiting on sets with l_wait_event.
2300  * Always returns 1.
2301  */
2302 int ptlrpc_expired_set(void *data)
2303 {
2304         struct ptlrpc_request_set *set = data;
2305         struct list_head *tmp;
2306         time64_t now = ktime_get_real_seconds();
2307
2308         ENTRY;
2309         LASSERT(set != NULL);
2310
2311         /*
2312          * A timeout expired. See which reqs it applies to...
2313          */
2314         list_for_each(tmp, &set->set_requests) {
2315                 struct ptlrpc_request *req =
2316                         list_entry(tmp, struct ptlrpc_request,
2317                                    rq_set_chain);
2318
2319                 /* don't expire request waiting for context */
2320                 if (req->rq_wait_ctx)
2321                         continue;
2322
2323                 /* Request in-flight? */
2324                 if (!((req->rq_phase == RQ_PHASE_RPC &&
2325                        !req->rq_waiting && !req->rq_resend) ||
2326                       (req->rq_phase == RQ_PHASE_BULK)))
2327                         continue;
2328
2329                 if (req->rq_timedout ||     /* already dealt with */
2330                     req->rq_deadline > now) /* not expired */
2331                         continue;
2332
2333                 /*
2334                  * Deal with this guy. Do it asynchronously to not block
2335                  * ptlrpcd thread.
2336                  */
2337                 ptlrpc_expire_one_request(req, 1);
2338         }
2339
2340         /*
2341          * When waiting for a whole set, we always break out of the
2342          * sleep so we can recalculate the timeout, or enable interrupts
2343          * if everyone's timed out.
2344          */
2345         RETURN(1);
2346 }
2347
2348 /**
2349  * Sets rq_intr flag in \a req under spinlock.
2350  */
2351 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2352 {
2353         spin_lock(&req->rq_lock);
2354         req->rq_intr = 1;
2355         spin_unlock(&req->rq_lock);
2356 }
2357 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2358
2359 /**
2360  * Interrupts (sets interrupted flag) all uncompleted requests in
2361  * a set \a data. Callback for l_wait_event for interruptible waits.
2362  */
2363 static void ptlrpc_interrupted_set(void *data)
2364 {
2365         struct ptlrpc_request_set *set = data;
2366         struct list_head *tmp;
2367
2368         LASSERT(set != NULL);
2369         CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2370
2371         list_for_each(tmp, &set->set_requests) {
2372                 struct ptlrpc_request *req =
2373                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2374
2375                 if (req->rq_intr)
2376                         continue;
2377
2378                 if (req->rq_phase != RQ_PHASE_RPC &&
2379                     req->rq_phase != RQ_PHASE_UNREG_RPC &&
2380                     !req->rq_allow_intr)
2381                         continue;
2382
2383                 ptlrpc_mark_interrupted(req);
2384         }
2385 }
2386
2387 /**
2388  * Get the smallest timeout in the set; this does NOT set a timeout.
2389  */
2390 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2391 {
2392         struct list_head *tmp;
2393         time64_t now = ktime_get_real_seconds();
2394         int timeout = 0;
2395         struct ptlrpc_request *req;
2396         time64_t deadline;
2397
2398         ENTRY;
2399         list_for_each(tmp, &set->set_requests) {
2400                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2401
2402                 /* Request in-flight? */
2403                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2404                       (req->rq_phase == RQ_PHASE_BULK) ||
2405                       (req->rq_phase == RQ_PHASE_NEW)))
2406                         continue;
2407
2408                 /* Already timed out. */
2409                 if (req->rq_timedout)
2410                         continue;
2411
2412                 /* Waiting for ctx. */
2413                 if (req->rq_wait_ctx)
2414                         continue;
2415
2416                 if (req->rq_phase == RQ_PHASE_NEW)
2417                         deadline = req->rq_sent;
2418                 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2419                         deadline = req->rq_sent;
2420                 else
2421                         deadline = req->rq_sent + req->rq_timeout;
2422
2423                 if (deadline <= now)    /* actually expired already */
2424                         timeout = 1;    /* ASAP */
2425                 else if (timeout == 0 || timeout > deadline - now)
2426                         timeout = deadline - now;
2427         }
2428         RETURN(timeout);
2429 }
2430
2431 /**
2432  * Send all unset request from the set and then wait untill all
2433  * requests in the set complete (either get a reply, timeout, get an
2434  * error or otherwise be interrupted).
2435  * Returns 0 on success or error code otherwise.
2436  */
2437 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2438 {
2439         struct list_head *tmp;
2440         struct ptlrpc_request *req;
2441         struct l_wait_info lwi;
2442         time64_t timeout;
2443         int rc;
2444
2445         ENTRY;
2446         if (set->set_producer)
2447                 (void)ptlrpc_set_producer(set);
2448         else
2449                 list_for_each(tmp, &set->set_requests) {
2450                         req = list_entry(tmp, struct ptlrpc_request,
2451                                          rq_set_chain);
2452                         if (req->rq_phase == RQ_PHASE_NEW)
2453                                 (void)ptlrpc_send_new_req(req);
2454                 }
2455
2456         if (list_empty(&set->set_requests))
2457                 RETURN(0);
2458
2459         do {
2460                 timeout = ptlrpc_set_next_timeout(set);
2461
2462                 /*
2463                  * wait until all complete, interrupted, or an in-flight
2464                  * req times out
2465                  */
2466                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2467                        set, timeout);
2468
2469                 if ((timeout == 0 && !signal_pending(current)) ||
2470                     set->set_allow_intr)
2471                         /*
2472                          * No requests are in-flight (ether timed out
2473                          * or delayed), so we can allow interrupts.
2474                          * We still want to block for a limited time,
2475                          * so we allow interrupts during the timeout.
2476                          */
2477                         lwi = LWI_TIMEOUT_INTR_ALL(
2478                                         cfs_time_seconds(timeout ? timeout : 1),
2479                                         ptlrpc_expired_set,
2480                                         ptlrpc_interrupted_set, set);
2481                 else
2482                         /*
2483                          * At least one request is in flight, so no
2484                          * interrupts are allowed. Wait until all
2485                          * complete, or an in-flight req times out.
2486                          */
2487                         lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2488                                           ptlrpc_expired_set, set);
2489
2490                 rc = l_wait_event(set->set_waitq,
2491                                   ptlrpc_check_set(NULL, set), &lwi);
2492
2493                 /*
2494                  * LU-769 - if we ignored the signal because it was already
2495                  * pending when we started, we need to handle it now or we risk
2496                  * it being ignored forever
2497                  */
2498                 if (rc == -ETIMEDOUT &&
2499                     (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2500                     signal_pending(current)) {
2501                         sigset_t blocked_sigs =
2502                                            cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2503
2504                         /*
2505                          * In fact we only interrupt for the "fatal" signals
2506                          * like SIGINT or SIGKILL. We still ignore less
2507                          * important signals since ptlrpc set is not easily
2508                          * reentrant from userspace again
2509                          */
2510                         if (signal_pending(current))
2511                                 ptlrpc_interrupted_set(set);
2512                         cfs_restore_sigs(blocked_sigs);
2513                 }
2514
2515                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2516
2517                 /*
2518                  * -EINTR => all requests have been flagged rq_intr so next
2519                  * check completes.
2520                  * -ETIMEDOUT => someone timed out.  When all reqs have
2521                  * timed out, signals are enabled allowing completion with
2522                  * EINTR.
2523                  * I don't really care if we go once more round the loop in
2524                  * the error cases -eeb.
2525                  */
2526                 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2527                         list_for_each(tmp, &set->set_requests) {
2528                                 req = list_entry(tmp, struct ptlrpc_request,
2529                                                  rq_set_chain);
2530                                 spin_lock(&req->rq_lock);
2531                                 req->rq_invalid_rqset = 1;
2532                                 spin_unlock(&req->rq_lock);
2533                         }
2534                 }
2535         } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2536
2537         LASSERT(atomic_read(&set->set_remaining) == 0);
2538
2539         rc = set->set_rc; /* rq_status of already freed requests if any */
2540         list_for_each(tmp, &set->set_requests) {
2541                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2542
2543                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2544                 if (req->rq_status != 0)
2545                         rc = req->rq_status;
2546         }
2547
2548         RETURN(rc);
2549 }
2550 EXPORT_SYMBOL(ptlrpc_set_wait);
2551
2552 /**
2553  * Helper fuction for request freeing.
2554  * Called when request count reached zero and request needs to be freed.
2555  * Removes request from all sorts of sending/replay lists it might be on,
2556  * frees network buffers if any are present.
2557  * If \a locked is set, that means caller is already holding import imp_lock
2558  * and so we no longer need to reobtain it (for certain lists manipulations)
2559  */
2560 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2561 {
2562         ENTRY;
2563
2564         if (!request)
2565                 RETURN_EXIT;
2566
2567         LASSERT(!request->rq_srv_req);
2568         LASSERT(request->rq_export == NULL);
2569         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2570         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2571         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2572         LASSERTF(!request->rq_replay, "req %p\n", request);
2573
2574         req_capsule_fini(&request->rq_pill);
2575
2576         /*
2577          * We must take it off the imp_replay_list first.  Otherwise, we'll set
2578          * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2579          */
2580         if (request->rq_import) {
2581                 if (!locked)
2582                         spin_lock(&request->rq_import->imp_lock);
2583                 list_del_init(&request->rq_replay_list);
2584                 list_del_init(&request->rq_unreplied_list);
2585                 if (!locked)
2586                         spin_unlock(&request->rq_import->imp_lock);
2587         }
2588         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2589
2590         if (atomic_read(&request->rq_refcount) != 0) {
2591                 DEBUG_REQ(D_ERROR, request,
2592                           "freeing request with nonzero refcount");
2593                 LBUG();
2594         }
2595
2596         if (request->rq_repbuf)
2597                 sptlrpc_cli_free_repbuf(request);
2598
2599         if (request->rq_import) {
2600                 class_import_put(request->rq_import);
2601                 request->rq_import = NULL;
2602         }
2603         if (request->rq_bulk)
2604                 ptlrpc_free_bulk(request->rq_bulk);
2605
2606         if (request->rq_reqbuf || request->rq_clrbuf)
2607                 sptlrpc_cli_free_reqbuf(request);
2608
2609         if (request->rq_cli_ctx)
2610                 sptlrpc_req_put_ctx(request, !locked);
2611
2612         if (request->rq_pool)
2613                 __ptlrpc_free_req_to_pool(request);
2614         else
2615                 ptlrpc_request_cache_free(request);
2616         EXIT;
2617 }
2618
2619 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2620 /**
2621  * Drop one request reference. Must be called with import imp_lock held.
2622  * When reference count drops to zero, request is freed.
2623  */
2624 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2625 {
2626         assert_spin_locked(&request->rq_import->imp_lock);
2627         (void)__ptlrpc_req_finished(request, 1);
2628 }
2629
2630 /**
2631  * Helper function
2632  * Drops one reference count for request \a request.
2633  * \a locked set indicates that caller holds import imp_lock.
2634  * Frees the request whe reference count reaches zero.
2635  *
2636  * \retval 1    the request is freed
2637  * \retval 0    some others still hold references on the request
2638  */
2639 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2640 {
2641         int count;
2642
2643         ENTRY;
2644         if (!request)
2645                 RETURN(1);
2646
2647         LASSERT(request != LP_POISON);
2648         LASSERT(request->rq_reqmsg != LP_POISON);
2649
2650         DEBUG_REQ(D_INFO, request, "refcount now %u",
2651                   atomic_read(&request->rq_refcount) - 1);
2652
2653         spin_lock(&request->rq_lock);
2654         count = atomic_dec_return(&request->rq_refcount);
2655         LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2656
2657         /*
2658          * For open RPC, the client does not know the EA size (LOV, ACL, and
2659          * so on) before replied, then the client has to reserve very large
2660          * reply buffer. Such buffer will not be released until the RPC freed.
2661          * Since The open RPC is replayable, we need to keep it in the replay
2662          * list until close. If there are a lot of files opened concurrently,
2663          * then the client may be OOM.
2664          *
2665          * If fact, it is unnecessary to keep reply buffer for open replay,
2666          * related EAs have already been saved via mdc_save_lovea() before
2667          * coming here. So it is safe to free the reply buffer some earlier
2668          * before releasing the RPC to avoid client OOM. LU-9514
2669          */
2670         if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2671                 spin_lock(&request->rq_early_free_lock);
2672                 sptlrpc_cli_free_repbuf(request);
2673                 request->rq_repbuf = NULL;
2674                 request->rq_repbuf_len = 0;
2675                 request->rq_repdata = NULL;
2676                 request->rq_reqdata_len = 0;
2677                 spin_unlock(&request->rq_early_free_lock);
2678         }
2679         spin_unlock(&request->rq_lock);
2680
2681         if (!count)
2682                 __ptlrpc_free_req(request, locked);
2683
2684         RETURN(!count);
2685 }
2686
2687 /**
2688  * Drops one reference count for a request.
2689  */
2690 void ptlrpc_req_finished(struct ptlrpc_request *request)
2691 {
2692         __ptlrpc_req_finished(request, 0);
2693 }
2694 EXPORT_SYMBOL(ptlrpc_req_finished);
2695
2696 /**
2697  * Returns xid of a \a request
2698  */
2699 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2700 {
2701         return request->rq_xid;
2702 }
2703 EXPORT_SYMBOL(ptlrpc_req_xid);
2704
2705 /**
2706  * Disengage the client's reply buffer from the network
2707  * NB does _NOT_ unregister any client-side bulk.
2708  * IDEMPOTENT, but _not_ safe against concurrent callers.
2709  * The request owner (i.e. the thread doing the I/O) must call...
2710  * Returns 0 on success or 1 if unregistering cannot be made.
2711  */
2712 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2713 {
2714         int rc;
2715         struct l_wait_info lwi;
2716
2717         /*
2718          * Might sleep.
2719          */
2720         LASSERT(!in_interrupt());
2721
2722         /* Let's setup deadline for reply unlink. */
2723         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2724             async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2725                 request->rq_reply_deadline = ktime_get_real_seconds() +
2726                                              LONG_UNLINK;
2727
2728         /*
2729          * Nothing left to do.
2730          */
2731         if (!ptlrpc_client_recv_or_unlink(request))
2732                 RETURN(1);
2733
2734         LNetMDUnlink(request->rq_reply_md_h);
2735
2736         /*
2737          * Let's check it once again.
2738          */
2739         if (!ptlrpc_client_recv_or_unlink(request))
2740                 RETURN(1);
2741
2742         /* Move to "Unregistering" phase as reply was not unlinked yet. */
2743         ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2744
2745         /*
2746          * Do not wait for unlink to finish.
2747          */
2748         if (async)
2749                 RETURN(0);
2750
2751         /*
2752          * We have to l_wait_event() whatever the result, to give liblustre
2753          * a chance to run reply_in_callback(), and to make sure we've
2754          * unlinked before returning a req to the pool.
2755          */
2756         for (;;) {
2757                 /* The wq argument is ignored by user-space wait_event macros */
2758                 wait_queue_head_t *wq = (request->rq_set) ?
2759                                         &request->rq_set->set_waitq :
2760                                         &request->rq_reply_waitq;
2761                 /*
2762                  * Network access will complete in finite time but the HUGE
2763                  * timeout lets us CWARN for visibility of sluggish NALs
2764                  */
2765                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2766                                            cfs_time_seconds(1), NULL, NULL);
2767                 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2768                                   &lwi);
2769                 if (rc == 0) {
2770                         ptlrpc_rqphase_move(request, request->rq_next_phase);
2771                         RETURN(1);
2772                 }
2773
2774                 LASSERT(rc == -ETIMEDOUT);
2775                 DEBUG_REQ(D_WARNING, request,
2776                           "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2777                           request->rq_receiving_reply,
2778                           request->rq_req_unlinked,
2779                           request->rq_reply_unlinked);
2780         }
2781         RETURN(0);
2782 }
2783
2784 static void ptlrpc_free_request(struct ptlrpc_request *req)
2785 {
2786         spin_lock(&req->rq_lock);
2787         req->rq_replay = 0;
2788         spin_unlock(&req->rq_lock);
2789
2790         if (req->rq_commit_cb)
2791                 req->rq_commit_cb(req);
2792         list_del_init(&req->rq_replay_list);
2793
2794         __ptlrpc_req_finished(req, 1);
2795 }
2796
2797 /**
2798  * the request is committed and dropped from the replay list of its import
2799  */
2800 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2801 {
2802         struct obd_import *imp = req->rq_import;
2803
2804         spin_lock(&imp->imp_lock);
2805         if (list_empty(&req->rq_replay_list)) {
2806                 spin_unlock(&imp->imp_lock);
2807                 return;
2808         }
2809
2810         if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2811                 if (imp->imp_replay_cursor == &req->rq_replay_list)
2812                         imp->imp_replay_cursor = req->rq_replay_list.next;
2813                 ptlrpc_free_request(req);
2814         }
2815
2816         spin_unlock(&imp->imp_lock);
2817 }
2818 EXPORT_SYMBOL(ptlrpc_request_committed);
2819
2820 /**
2821  * Iterates through replay_list on import and prunes
2822  * all requests have transno smaller than last_committed for the
2823  * import and don't have rq_replay set.
2824  * Since requests are sorted in transno order, stops when meetign first
2825  * transno bigger than last_committed.
2826  * caller must hold imp->imp_lock
2827  */
2828 void ptlrpc_free_committed(struct obd_import *imp)
2829 {
2830         struct ptlrpc_request *req, *saved;
2831         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2832         bool skip_committed_list = true;
2833
2834         ENTRY;
2835         LASSERT(imp != NULL);
2836         assert_spin_locked(&imp->imp_lock);
2837
2838         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2839             imp->imp_generation == imp->imp_last_generation_checked) {
2840                 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2841                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2842                 RETURN_EXIT;
2843         }
2844         CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2845                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2846                imp->imp_generation);
2847
2848         if (imp->imp_generation != imp->imp_last_generation_checked ||
2849             imp->imp_last_transno_checked == 0)
2850                 skip_committed_list = false;
2851
2852         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2853         imp->imp_last_generation_checked = imp->imp_generation;
2854
2855         list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2856                                  rq_replay_list) {
2857                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
2858                 LASSERT(req != last_req);
2859                 last_req = req;
2860
2861                 if (req->rq_transno == 0) {
2862                         DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2863                         LBUG();
2864                 }
2865                 if (req->rq_import_generation < imp->imp_generation) {
2866                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2867                         GOTO(free_req, 0);
2868                 }
2869
2870                 /* not yet committed */
2871                 if (req->rq_transno > imp->imp_peer_committed_transno) {
2872                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2873                         break;
2874                 }
2875
2876                 if (req->rq_replay) {
2877                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2878                         list_move_tail(&req->rq_replay_list,
2879                                        &imp->imp_committed_list);
2880                         continue;
2881                 }
2882
2883                 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2884                           imp->imp_peer_committed_transno);
2885 free_req:
2886                 ptlrpc_free_request(req);
2887         }
2888
2889         if (skip_committed_list)
2890                 GOTO(out, 0);
2891
2892         list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2893                                  rq_replay_list) {
2894                 LASSERT(req->rq_transno != 0);
2895                 if (req->rq_import_generation < imp->imp_generation ||
2896                     !req->rq_replay) {
2897                         DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2898                                   req->rq_import_generation <
2899                                   imp->imp_generation ? "stale" : "closed");
2900
2901                         if (imp->imp_replay_cursor == &req->rq_replay_list)
2902                                 imp->imp_replay_cursor =
2903                                         req->rq_replay_list.next;
2904
2905                         ptlrpc_free_request(req);
2906                 }
2907         }
2908 out:
2909         EXIT;
2910 }
2911
2912 void ptlrpc_cleanup_client(struct obd_import *imp)
2913 {
2914         ENTRY;
2915         EXIT;
2916 }
2917
2918 /**
2919  * Schedule previously sent request for resend.
2920  * For bulk requests we assign new xid (to avoid problems with
2921  * lost replies and therefore several transfers landing into same buffer
2922  * from different sending attempts).
2923  */
2924 void ptlrpc_resend_req(struct ptlrpc_request *req)
2925 {
2926         DEBUG_REQ(D_HA, req, "going to resend");
2927         spin_lock(&req->rq_lock);
2928
2929         /*
2930          * Request got reply but linked to the import list still.
2931          * Let ptlrpc_check_set() process it.
2932          */
2933         if (ptlrpc_client_replied(req)) {
2934                 spin_unlock(&req->rq_lock);
2935                 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2936                 return;
2937         }
2938
2939         req->rq_status = -EAGAIN;
2940
2941         req->rq_resend = 1;
2942         req->rq_net_err = 0;
2943         req->rq_timedout = 0;
2944
2945         ptlrpc_client_wake_req(req);
2946         spin_unlock(&req->rq_lock);
2947 }
2948
2949 /* XXX: this function and rq_status are currently unused */
2950 void ptlrpc_restart_req(struct ptlrpc_request *req)
2951 {
2952         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2953         req->rq_status = -ERESTARTSYS;
2954
2955         spin_lock(&req->rq_lock);
2956         req->rq_restart = 1;
2957         req->rq_timedout = 0;
2958         ptlrpc_client_wake_req(req);
2959         spin_unlock(&req->rq_lock);
2960 }
2961
2962 /**
2963  * Grab additional reference on a request \a req
2964  */
2965 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2966 {
2967         ENTRY;
2968         atomic_inc(&req->rq_refcount);
2969         RETURN(req);
2970 }
2971 EXPORT_SYMBOL(ptlrpc_request_addref);
2972
2973 /**
2974  * Add a request to import replay_list.
2975  * Must be called under imp_lock
2976  */
2977 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2978                                       struct obd_import *imp)
2979 {
2980         struct list_head *tmp;
2981
2982         assert_spin_locked(&imp->imp_lock);
2983
2984         if (req->rq_transno == 0) {
2985                 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2986                 LBUG();
2987         }
2988
2989         /*
2990          * clear this for new requests that were resent as well
2991          * as resent replayed requests.
2992          */
2993         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2994
2995         /* don't re-add requests that have been replayed */
2996         if (!list_empty(&req->rq_replay_list))
2997                 return;
2998
2999         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3000
3001         spin_lock(&req->rq_lock);
3002         req->rq_resend = 0;
3003         spin_unlock(&req->rq_lock);
3004
3005         LASSERT(imp->imp_replayable);
3006         /* Balanced in ptlrpc_free_committed, usually. */
3007         ptlrpc_request_addref(req);
3008         list_for_each_prev(tmp, &imp->imp_replay_list) {
3009                 struct ptlrpc_request *iter = list_entry(tmp,
3010                                                          struct ptlrpc_request,
3011                                                          rq_replay_list);
3012
3013                 /*
3014                  * We may have duplicate transnos if we create and then
3015                  * open a file, or for closes retained if to match creating
3016                  * opens, so use req->rq_xid as a secondary key.
3017                  * (See bugs 684, 685, and 428.)
3018                  * XXX no longer needed, but all opens need transnos!
3019                  */
3020                 if (iter->rq_transno > req->rq_transno)
3021                         continue;
3022
3023                 if (iter->rq_transno == req->rq_transno) {
3024                         LASSERT(iter->rq_xid != req->rq_xid);
3025                         if (iter->rq_xid > req->rq_xid)
3026                                 continue;
3027                 }
3028
3029                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3030                 return;
3031         }
3032
3033         list_add(&req->rq_replay_list, &imp->imp_replay_list);
3034 }
3035
3036 /**
3037  * Send request and wait until it completes.
3038  * Returns request processing status.
3039  */
3040 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3041 {
3042         struct ptlrpc_request_set *set;
3043         int rc;
3044
3045         ENTRY;
3046         LASSERT(req->rq_set == NULL);
3047         LASSERT(!req->rq_receiving_reply);
3048
3049         set = ptlrpc_prep_set();
3050         if (!set) {
3051                 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3052                 RETURN(-ENOMEM);
3053         }
3054
3055         /* for distributed debugging */
3056         lustre_msg_set_status(req->rq_reqmsg, current_pid());
3057
3058         /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3059         ptlrpc_request_addref(req);
3060         ptlrpc_set_add_req(set, req);
3061         rc = ptlrpc_set_wait(NULL, set);
3062         ptlrpc_set_destroy(set);
3063
3064         RETURN(rc);
3065 }
3066 EXPORT_SYMBOL(ptlrpc_queue_wait);
3067
3068 /**
3069  * Callback used for replayed requests reply processing.
3070  * In case of successful reply calls registered request replay callback.
3071  * In case of error restart replay process.
3072  */
3073 static int ptlrpc_replay_interpret(const struct lu_env *env,
3074                                    struct ptlrpc_request *req,
3075                                    void *args, int rc)
3076 {
3077         struct ptlrpc_replay_async_args *aa = args;
3078         struct obd_import *imp = req->rq_import;
3079
3080         ENTRY;
3081         atomic_dec(&imp->imp_replay_inflight);
3082
3083         /*
3084          * Note: if it is bulk replay (MDS-MDS replay), then even if
3085          * server got the request, but bulk transfer timeout, let's
3086          * replay the bulk req again
3087          */
3088         if (!ptlrpc_client_replied(req) ||
3089             (req->rq_bulk &&
3090              lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3091                 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3092                 GOTO(out, rc = -ETIMEDOUT);
3093         }
3094
3095         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3096             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3097             lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3098                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3099
3100         /** VBR: check version failure */
3101         if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3102                 /** replay was failed due to version mismatch */
3103                 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3104                 spin_lock(&imp->imp_lock);
3105                 imp->imp_vbr_failed = 1;
3106                 spin_unlock(&imp->imp_lock);
3107                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3108         } else {
3109                 /** The transno had better not change over replay. */
3110                 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3111                          lustre_msg_get_transno(req->rq_repmsg) ||
3112                          lustre_msg_get_transno(req->rq_repmsg) == 0,
3113                          "%#llx/%#llx\n",
3114                          lustre_msg_get_transno(req->rq_reqmsg),
3115                          lustre_msg_get_transno(req->rq_repmsg));
3116         }
3117
3118         spin_lock(&imp->imp_lock);
3119         imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3120         spin_unlock(&imp->imp_lock);
3121         LASSERT(imp->imp_last_replay_transno);
3122
3123         /* transaction number shouldn't be bigger than the latest replayed */
3124         if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3125                 DEBUG_REQ(D_ERROR, req,
3126                           "Reported transno=%llu is bigger than replayed=%llu",
3127                           req->rq_transno,
3128                           lustre_msg_get_transno(req->rq_reqmsg));
3129                 GOTO(out, rc = -EINVAL);
3130         }
3131
3132         DEBUG_REQ(D_HA, req, "got reply");
3133
3134         /* let the callback do fixups, possibly including in the request */
3135         if (req->rq_replay_cb)
3136                 req->rq_replay_cb(req);
3137
3138         if (ptlrpc_client_replied(req) &&
3139             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3140                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3141                           lustre_msg_get_status(req->rq_repmsg),
3142                           aa->praa_old_status);
3143
3144                 /*
3145                  * Note: If the replay fails for MDT-MDT recovery, let's
3146                  * abort all of the following requests in the replay
3147                  * and sending list, because MDT-MDT update requests
3148                  * are dependent on each other, see LU-7039
3149                  */
3150                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3151                         struct ptlrpc_request *free_req;
3152                         struct ptlrpc_request *tmp;
3153
3154                         spin_lock(&imp->imp_lock);
3155                         list_for_each_entry_safe(free_req, tmp,
3156                                                  &imp->imp_replay_list,
3157                                                  rq_replay_list) {
3158                                 ptlrpc_free_request(free_req);
3159                         }
3160
3161                         list_for_each_entry_safe(free_req, tmp,
3162                                                  &imp->imp_committed_list,
3163                                                  rq_replay_list) {
3164                                 ptlrpc_free_request(free_req);
3165                         }
3166
3167                         list_for_each_entry_safe(free_req, tmp,
3168                                                  &imp->imp_delayed_list,
3169                                                  rq_list) {
3170                                 spin_lock(&free_req->rq_lock);
3171                                 free_req->rq_err = 1;
3172                                 free_req->rq_status = -EIO;
3173                                 ptlrpc_client_wake_req(free_req);
3174                                 spin_unlock(&free_req->rq_lock);
3175                         }
3176
3177                         list_for_each_entry_safe(free_req, tmp,
3178                                                  &imp->imp_sending_list,
3179                                                  rq_list) {
3180                                 spin_lock(&free_req->rq_lock);
3181                                 free_req->rq_err = 1;
3182                                 free_req->rq_status = -EIO;
3183                                 ptlrpc_client_wake_req(free_req);
3184                                 spin_unlock(&free_req->rq_lock);
3185                         }
3186                         spin_unlock(&imp->imp_lock);
3187                 }
3188         } else {
3189                 /* Put it back for re-replay. */
3190                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3191         }
3192
3193         /*
3194          * Errors while replay can set transno to 0, but
3195          * imp_last_replay_transno shouldn't be set to 0 anyway
3196          */
3197         if (req->rq_transno == 0)
3198                 CERROR("Transno is 0 during replay!\n");
3199
3200         /* continue with recovery */
3201         rc = ptlrpc_import_recovery_state_machine(imp);
3202  out:
3203         req->rq_send_state = aa->praa_old_state;
3204
3205         if (rc != 0)
3206                 /* this replay failed, so restart recovery */
3207                 ptlrpc_connect_import(imp);
3208
3209         RETURN(rc);
3210 }
3211
3212 /**
3213  * Prepares and queues request for replay.
3214  * Adds it to ptlrpcd queue for actual sending.
3215  * Returns 0 on success.
3216  */
3217 int ptlrpc_replay_req(struct ptlrpc_request *req)
3218 {
3219         struct ptlrpc_repl