Whamcloud - gitweb
LU-12368 ptlrpc: make DEBUG_REQ messages consistent
[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_generation = req->rq_import_generation;
221         desc->bd_import = class_import_get(imp);
222         desc->bd_req = req;
223
224         desc->bd_cbid.cbid_fn  = client_bulk_callback;
225         desc->bd_cbid.cbid_arg = desc;
226
227         /* This makes req own desc, and free it when she frees herself */
228         req->rq_bulk = desc;
229
230         return desc;
231 }
232 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
233
234 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
235                              struct page *page, int pageoffset, int len,
236                              int pin)
237 {
238         lnet_kiov_t *kiov;
239
240         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
241         LASSERT(page != NULL);
242         LASSERT(pageoffset >= 0);
243         LASSERT(len > 0);
244         LASSERT(pageoffset + len <= PAGE_SIZE);
245         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
246
247         kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
248
249         desc->bd_nob += len;
250
251         if (pin)
252                 get_page(page);
253
254         kiov->kiov_page = page;
255         kiov->kiov_offset = pageoffset;
256         kiov->kiov_len = len;
257
258         desc->bd_iov_count++;
259 }
260 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
261
262 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
263                           void *frag, int len)
264 {
265         struct kvec *iovec;
266
267         ENTRY;
268
269         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
270         LASSERT(frag != NULL);
271         LASSERT(len > 0);
272         LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
273
274         iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
275
276         desc->bd_nob += len;
277
278         iovec->iov_base = frag;
279         iovec->iov_len = len;
280
281         desc->bd_iov_count++;
282
283         RETURN(desc->bd_nob);
284 }
285 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
286
287 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
288 {
289         ENTRY;
290
291         LASSERT(desc != NULL);
292         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
293         LASSERT(desc->bd_md_count == 0);         /* network hands off */
294         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
295         LASSERT(desc->bd_frag_ops != NULL);
296
297         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
298                 sptlrpc_enc_pool_put_pages(desc);
299
300         if (desc->bd_export)
301                 class_export_put(desc->bd_export);
302         else
303                 class_import_put(desc->bd_import);
304
305         if (desc->bd_frag_ops->release_frags != NULL)
306                 desc->bd_frag_ops->release_frags(desc);
307
308         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
309                 OBD_FREE_LARGE(GET_KIOV(desc),
310                                desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
311         else
312                 OBD_FREE_LARGE(GET_KVEC(desc),
313                                desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
314         OBD_FREE_PTR(desc);
315         EXIT;
316 }
317 EXPORT_SYMBOL(ptlrpc_free_bulk);
318
319 /**
320  * Set server timelimit for this req, i.e. how long are we willing to wait
321  * for reply before timing out this request.
322  */
323 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
324 {
325         __u32 serv_est;
326         int idx;
327         struct imp_at *at;
328
329         LASSERT(req->rq_import);
330
331         if (AT_OFF) {
332                 /* non-AT settings */
333                 /**
334                  * \a imp_server_timeout means this is reverse import and
335                  * we send (currently only) ASTs to the client and cannot afford
336                  * to wait too long for the reply, otherwise the other client
337                  * (because of which we are sending this request) would
338                  * timeout waiting for us
339                  */
340                 req->rq_timeout = req->rq_import->imp_server_timeout ?
341                                   obd_timeout / 2 : obd_timeout;
342         } else {
343                 at = &req->rq_import->imp_at;
344                 idx = import_at_get_index(req->rq_import,
345                                           req->rq_request_portal);
346                 serv_est = at_get(&at->iat_service_estimate[idx]);
347                 req->rq_timeout = at_est2timeout(serv_est);
348         }
349         /*
350          * We could get even fancier here, using history to predict increased
351          * loading...
352          */
353
354         /*
355          * Let the server know what this RPC timeout is by putting it in the
356          * reqmsg
357          */
358         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
359 }
360 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
361
362 /* Adjust max service estimate based on server value */
363 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
364                                   unsigned int serv_est)
365 {
366         int idx;
367         unsigned int oldse;
368         struct imp_at *at;
369
370         LASSERT(req->rq_import);
371         at = &req->rq_import->imp_at;
372
373         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
374         /*
375          * max service estimates are tracked on the server side,
376          * so just keep minimal history here
377          */
378         oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
379         if (oldse != 0)
380                 CDEBUG(D_ADAPTTO,
381                        "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
382                        req->rq_import->imp_obd->obd_name,
383                        req->rq_request_portal,
384                        oldse, at_get(&at->iat_service_estimate[idx]));
385 }
386
387 /* Expected network latency per remote node (secs) */
388 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
389 {
390         return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
391 }
392
393 /* Adjust expected network latency */
394 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
395                                unsigned int service_time)
396 {
397         unsigned int nl, oldnl;
398         struct imp_at *at;
399         time64_t now = ktime_get_real_seconds();
400
401         LASSERT(req->rq_import);
402
403         if (service_time > now - req->rq_sent + 3) {
404                 /*
405                  * b=16408, however, this can also happen if early reply
406                  * is lost and client RPC is expired and resent, early reply
407                  * or reply of original RPC can still be fit in reply buffer
408                  * of resent RPC, now client is measuring time from the
409                  * resent time, but server sent back service time of original
410                  * RPC.
411                  */
412                 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
413                        D_ADAPTTO : D_WARNING,
414                        "Reported service time %u > total measured time %lld\n",
415                        service_time, now - req->rq_sent);
416                 return;
417         }
418
419         /* Network latency is total time less server processing time */
420         nl = max_t(int, now - req->rq_sent -
421                         service_time, 0) + 1; /* st rounding */
422         at = &req->rq_import->imp_at;
423
424         oldnl = at_measured(&at->iat_net_latency, nl);
425         if (oldnl != 0)
426                 CDEBUG(D_ADAPTTO,
427                        "The network latency for %s (nid %s) has changed from %d to %d\n",
428                        req->rq_import->imp_obd->obd_name,
429                        obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
430                        oldnl, at_get(&at->iat_net_latency));
431 }
432
433 static int unpack_reply(struct ptlrpc_request *req)
434 {
435         int rc;
436
437         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
438                 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
439                 if (rc) {
440                         DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
441                                   rc);
442                         return -EPROTO;
443                 }
444         }
445
446         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
447         if (rc) {
448                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
449                           rc);
450                 return -EPROTO;
451         }
452         return 0;
453 }
454
455 /**
456  * Handle an early reply message, called with the rq_lock held.
457  * If anything goes wrong just ignore it - same as if it never happened
458  */
459 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
460 __must_hold(&req->rq_lock)
461 {
462         struct ptlrpc_request *early_req;
463         time64_t olddl;
464         int rc;
465
466         ENTRY;
467         req->rq_early = 0;
468         spin_unlock(&req->rq_lock);
469
470         rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
471         if (rc) {
472                 spin_lock(&req->rq_lock);
473                 RETURN(rc);
474         }
475
476         rc = unpack_reply(early_req);
477         if (rc != 0) {
478                 sptlrpc_cli_finish_early_reply(early_req);
479                 spin_lock(&req->rq_lock);
480                 RETURN(rc);
481         }
482
483         /*
484          * Use new timeout value just to adjust the local value for this
485          * request, don't include it into at_history. It is unclear yet why
486          * service time increased and should it be counted or skipped, e.g.
487          * that can be recovery case or some error or server, the real reply
488          * will add all new data if it is worth to add.
489          */
490         req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
491         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
492
493         /* Network latency can be adjusted, it is pure network delays */
494         ptlrpc_at_adj_net_latency(req,
495                                   lustre_msg_get_service_time(early_req->rq_repmsg));
496
497         sptlrpc_cli_finish_early_reply(early_req);
498
499         spin_lock(&req->rq_lock);
500         olddl = req->rq_deadline;
501         /*
502          * server assumes it now has rq_timeout from when the request
503          * arrived, so the client should give it at least that long.
504          * since we don't know the arrival time we'll use the original
505          * sent time
506          */
507         req->rq_deadline = req->rq_sent + req->rq_timeout +
508                            ptlrpc_at_get_net_latency(req);
509
510         /* The below message is checked in replay-single.sh test_65{a,b} */
511         /* The below message is checked in sanity-{gss,krb5} test_8 */
512         DEBUG_REQ(D_ADAPTTO, req,
513                   "Early reply #%d, new deadline in %llds (%llds)",
514                   req->rq_early_count,
515                   req->rq_deadline - ktime_get_real_seconds(),
516                   req->rq_deadline - olddl);
517
518         RETURN(rc);
519 }
520
521 static struct kmem_cache *request_cache;
522
523 int ptlrpc_request_cache_init(void)
524 {
525         request_cache = kmem_cache_create("ptlrpc_cache",
526                                           sizeof(struct ptlrpc_request),
527                                           0, SLAB_HWCACHE_ALIGN, NULL);
528         return request_cache ? 0 : -ENOMEM;
529 }
530
531 void ptlrpc_request_cache_fini(void)
532 {
533         kmem_cache_destroy(request_cache);
534 }
535
536 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
537 {
538         struct ptlrpc_request *req;
539
540         OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
541         return req;
542 }
543
544 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
545 {
546         OBD_SLAB_FREE_PTR(req, request_cache);
547 }
548
549 /**
550  * Wind down request pool \a pool.
551  * Frees all requests from the pool too
552  */
553 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
554 {
555         struct list_head *l, *tmp;
556         struct ptlrpc_request *req;
557
558         LASSERT(pool != NULL);
559
560         spin_lock(&pool->prp_lock);
561         list_for_each_safe(l, tmp, &pool->prp_req_list) {
562                 req = list_entry(l, struct ptlrpc_request, rq_list);
563                 list_del(&req->rq_list);
564                 LASSERT(req->rq_reqbuf);
565                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
566                 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
567                 ptlrpc_request_cache_free(req);
568         }
569         spin_unlock(&pool->prp_lock);
570         OBD_FREE(pool, sizeof(*pool));
571 }
572 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
573
574 /**
575  * Allocates, initializes and adds \a num_rq requests to the pool \a pool
576  */
577 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
578 {
579         int i;
580         int size = 1;
581
582         while (size < pool->prp_rq_size)
583                 size <<= 1;
584
585         LASSERTF(list_empty(&pool->prp_req_list) ||
586                  size == pool->prp_rq_size,
587                  "Trying to change pool size with nonempty pool from %d to %d bytes\n",
588                  pool->prp_rq_size, size);
589
590         pool->prp_rq_size = size;
591         for (i = 0; i < num_rq; i++) {
592                 struct ptlrpc_request *req;
593                 struct lustre_msg *msg;
594
595                 req = ptlrpc_request_cache_alloc(GFP_NOFS);
596                 if (!req)
597                         return i;
598                 OBD_ALLOC_LARGE(msg, size);
599                 if (!msg) {
600                         ptlrpc_request_cache_free(req);
601                         return i;
602                 }
603                 req->rq_reqbuf = msg;
604                 req->rq_reqbuf_len = size;
605                 req->rq_pool = pool;
606                 spin_lock(&pool->prp_lock);
607                 list_add_tail(&req->rq_list, &pool->prp_req_list);
608                 spin_unlock(&pool->prp_lock);
609         }
610         return num_rq;
611 }
612 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
613
614 /**
615  * Create and initialize new request pool with given attributes:
616  * \a num_rq - initial number of requests to create for the pool
617  * \a msgsize - maximum message size possible for requests in thid pool
618  * \a populate_pool - function to be called when more requests need to be added
619  *                    to the pool
620  * Returns pointer to newly created pool or NULL on error.
621  */
622 struct ptlrpc_request_pool *
623 ptlrpc_init_rq_pool(int num_rq, int msgsize,
624                     int (*populate_pool)(struct ptlrpc_request_pool *, int))
625 {
626         struct ptlrpc_request_pool *pool;
627
628         OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
629         if (!pool)
630                 return NULL;
631
632         /*
633          * Request next power of two for the allocation, because internally
634          * kernel would do exactly this
635          */
636         spin_lock_init(&pool->prp_lock);
637         INIT_LIST_HEAD(&pool->prp_req_list);
638         pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
639         pool->prp_populate = populate_pool;
640
641         populate_pool(pool, num_rq);
642
643         return pool;
644 }
645 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
646
647 /**
648  * Fetches one request from pool \a pool
649  */
650 static struct ptlrpc_request *
651 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
652 {
653         struct ptlrpc_request *request;
654         struct lustre_msg *reqbuf;
655
656         if (!pool)
657                 return NULL;
658
659         spin_lock(&pool->prp_lock);
660
661         /*
662          * See if we have anything in a pool, and bail out if nothing,
663          * in writeout path, where this matters, this is safe to do, because
664          * nothing is lost in this case, and when some in-flight requests
665          * complete, this code will be called again.
666          */
667         if (unlikely(list_empty(&pool->prp_req_list))) {
668                 spin_unlock(&pool->prp_lock);
669                 return NULL;
670         }
671
672         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
673                              rq_list);
674         list_del_init(&request->rq_list);
675         spin_unlock(&pool->prp_lock);
676
677         LASSERT(request->rq_reqbuf);
678         LASSERT(request->rq_pool);
679
680         reqbuf = request->rq_reqbuf;
681         memset(request, 0, sizeof(*request));
682         request->rq_reqbuf = reqbuf;
683         request->rq_reqbuf_len = pool->prp_rq_size;
684         request->rq_pool = pool;
685
686         return request;
687 }
688
689 /**
690  * Returns freed \a request to pool.
691  */
692 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
693 {
694         struct ptlrpc_request_pool *pool = request->rq_pool;
695
696         spin_lock(&pool->prp_lock);
697         LASSERT(list_empty(&request->rq_list));
698         LASSERT(!request->rq_receiving_reply);
699         list_add_tail(&request->rq_list, &pool->prp_req_list);
700         spin_unlock(&pool->prp_lock);
701 }
702
703 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
704 {
705         struct obd_import *imp = req->rq_import;
706         struct list_head *tmp;
707         struct ptlrpc_request *iter;
708
709         assert_spin_locked(&imp->imp_lock);
710         LASSERT(list_empty(&req->rq_unreplied_list));
711
712         /* unreplied list is sorted by xid in ascending order */
713         list_for_each_prev(tmp, &imp->imp_unreplied_list) {
714                 iter = list_entry(tmp, struct ptlrpc_request,
715                                   rq_unreplied_list);
716
717                 LASSERT(req->rq_xid != iter->rq_xid);
718                 if (req->rq_xid < iter->rq_xid)
719                         continue;
720                 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
721                 return;
722         }
723         list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
724 }
725
726 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
727 {
728         req->rq_xid = ptlrpc_next_xid();
729         ptlrpc_add_unreplied(req);
730 }
731
732 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
733 {
734         spin_lock(&req->rq_import->imp_lock);
735         ptlrpc_assign_next_xid_nolock(req);
736         spin_unlock(&req->rq_import->imp_lock);
737 }
738
739 static atomic64_t ptlrpc_last_xid;
740
741 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
742                              __u32 version, int opcode, char **bufs,
743                              struct ptlrpc_cli_ctx *ctx)
744 {
745         int count;
746         struct obd_import *imp;
747         __u32 *lengths;
748         int rc;
749
750         ENTRY;
751
752         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
753         imp = request->rq_import;
754         lengths = request->rq_pill.rc_area[RCL_CLIENT];
755
756         if (ctx) {
757                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
758         } else {
759                 rc = sptlrpc_req_get_ctx(request);
760                 if (rc)
761                         GOTO(out_free, rc);
762         }
763         sptlrpc_req_set_flavor(request, opcode);
764
765         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
766                                  lengths, bufs);
767         if (rc)
768                 GOTO(out_ctx, rc);
769
770         lustre_msg_add_version(request->rq_reqmsg, version);
771         request->rq_send_state = LUSTRE_IMP_FULL;
772         request->rq_type = PTL_RPC_MSG_REQUEST;
773
774         request->rq_req_cbid.cbid_fn  = request_out_callback;
775         request->rq_req_cbid.cbid_arg = request;
776
777         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
778         request->rq_reply_cbid.cbid_arg = request;
779
780         request->rq_reply_deadline = 0;
781         request->rq_bulk_deadline = 0;
782         request->rq_req_deadline = 0;
783         request->rq_phase = RQ_PHASE_NEW;
784         request->rq_next_phase = RQ_PHASE_UNDEFINED;
785
786         request->rq_request_portal = imp->imp_client->cli_request_portal;
787         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
788
789         ptlrpc_at_set_req_timeout(request);
790
791         lustre_msg_set_opc(request->rq_reqmsg, opcode);
792
793         /* Let's setup deadline for req/reply/bulk unlink for opcode. */
794         if (cfs_fail_val == opcode) {
795                 time64_t *fail_t = NULL, *fail2_t = NULL;
796
797                 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
798                         fail_t = &request->rq_bulk_deadline;
799                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
800                         fail_t = &request->rq_reply_deadline;
801                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
802                         fail_t = &request->rq_req_deadline;
803                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
804                         fail_t = &request->rq_reply_deadline;
805                         fail2_t = &request->rq_bulk_deadline;
806                 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
807                         time64_t now = ktime_get_real_seconds();
808                         u64 xid = ((u64)now >> 4) << 24;
809
810                         atomic64_set(&ptlrpc_last_xid, xid);
811                 }
812
813                 if (fail_t) {
814                         *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
815
816                         if (fail2_t)
817                                 *fail2_t = ktime_get_real_seconds() +
818                                            LONG_UNLINK;
819
820                         /*
821                          * The RPC is infected, let the test to change the
822                          * fail_loc
823                          */
824                         msleep(4 * MSEC_PER_SEC);
825                 }
826         }
827         ptlrpc_assign_next_xid(request);
828
829         RETURN(0);
830
831 out_ctx:
832         LASSERT(!request->rq_pool);
833         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
834 out_free:
835         class_import_put(imp);
836
837         return rc;
838 }
839 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
840
841 /**
842  * Pack request buffers for network transfer, performing necessary encryption
843  * steps if necessary.
844  */
845 int ptlrpc_request_pack(struct ptlrpc_request *request,
846                         __u32 version, int opcode)
847 {
848         return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
849 }
850 EXPORT_SYMBOL(ptlrpc_request_pack);
851
852 /**
853  * Helper function to allocate new request on import \a imp
854  * and possibly using existing request from pool \a pool if provided.
855  * Returns allocated request structure with import field filled or
856  * NULL on error.
857  */
858 static inline
859 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
860                                               struct ptlrpc_request_pool *pool)
861 {
862         struct ptlrpc_request *request = NULL;
863
864         request = ptlrpc_request_cache_alloc(GFP_NOFS);
865
866         if (!request && pool)
867                 request = ptlrpc_prep_req_from_pool(pool);
868
869         if (request) {
870                 ptlrpc_cli_req_init(request);
871
872                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
873                 LASSERT(imp != LP_POISON);
874                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
875                          imp->imp_client);
876                 LASSERT(imp->imp_client != LP_POISON);
877
878                 request->rq_import = class_import_get(imp);
879         } else {
880                 CERROR("request allocation out of memory\n");
881         }
882
883         return request;
884 }
885
886 /**
887  * Helper function for creating a request.
888  * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
889  * buffer structures according to capsule template \a format.
890  * Returns allocated request structure pointer or NULL on error.
891  */
892 static struct ptlrpc_request *
893 ptlrpc_request_alloc_internal(struct obd_import *imp,
894                               struct ptlrpc_request_pool *pool,
895                               const struct req_format *format)
896 {
897         struct ptlrpc_request *request;
898         int connect = 0;
899
900         request = __ptlrpc_request_alloc(imp, pool);
901         if (!request)
902                 return NULL;
903
904         /*
905          * initiate connection if needed when the import has been
906          * referenced by the new request to avoid races with disconnect
907          */
908         if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
909                 int rc;
910
911                 CDEBUG_LIMIT(imp->imp_idle_debug,
912                              "%s: reconnect after %llds idle\n",
913                              imp->imp_obd->obd_name, ktime_get_real_seconds() -
914                                                      imp->imp_last_reply_time);
915                 spin_lock(&imp->imp_lock);
916                 if (imp->imp_state == LUSTRE_IMP_IDLE) {
917                         imp->imp_generation++;
918                         imp->imp_initiated_at = imp->imp_generation;
919                         imp->imp_state =  LUSTRE_IMP_NEW;
920                         connect = 1;
921                 }
922                 spin_unlock(&imp->imp_lock);
923                 if (connect) {
924                         rc = ptlrpc_connect_import(imp);
925                         if (rc < 0) {
926                                 ptlrpc_request_free(request);
927                                 return NULL;
928                         }
929                         ptlrpc_pinger_add_import(imp);
930                 }
931         }
932
933         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
934         req_capsule_set(&request->rq_pill, format);
935         return request;
936 }
937
938 /**
939  * Allocate new request structure for import \a imp and initialize its
940  * buffer structure according to capsule template \a format.
941  */
942 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
943                                             const struct req_format *format)
944 {
945         return ptlrpc_request_alloc_internal(imp, NULL, format);
946 }
947 EXPORT_SYMBOL(ptlrpc_request_alloc);
948
949 /**
950  * Allocate new request structure for import \a imp from pool \a pool and
951  * initialize its buffer structure according to capsule template \a format.
952  */
953 struct ptlrpc_request *
954 ptlrpc_request_alloc_pool(struct obd_import *imp,
955                           struct ptlrpc_request_pool *pool,
956                           const struct req_format *format)
957 {
958         return ptlrpc_request_alloc_internal(imp, pool, format);
959 }
960 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
961
962 /**
963  * For requests not from pool, free memory of the request structure.
964  * For requests obtained from a pool earlier, return request back to pool.
965  */
966 void ptlrpc_request_free(struct ptlrpc_request *request)
967 {
968         if (request->rq_pool)
969                 __ptlrpc_free_req_to_pool(request);
970         else
971                 ptlrpc_request_cache_free(request);
972 }
973 EXPORT_SYMBOL(ptlrpc_request_free);
974
975 /**
976  * Allocate new request for operatione \a opcode and immediatelly pack it for
977  * network transfer.
978  * Only used for simple requests like OBD_PING where the only important
979  * part of the request is operation itself.
980  * Returns allocated request or NULL on error.
981  */
982 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
983                                                  const struct req_format *format,
984                                                  __u32 version, int opcode)
985 {
986         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
987         int rc;
988
989         if (req) {
990                 rc = ptlrpc_request_pack(req, version, opcode);
991                 if (rc) {
992                         ptlrpc_request_free(req);
993                         req = NULL;
994                 }
995         }
996         return req;
997 }
998 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
999
1000 /**
1001  * Allocate and initialize new request set structure on the current CPT.
1002  * Returns a pointer to the newly allocated set structure or NULL on error.
1003  */
1004 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1005 {
1006         struct ptlrpc_request_set *set;
1007         int cpt;
1008
1009         ENTRY;
1010         cpt = cfs_cpt_current(cfs_cpt_table, 0);
1011         OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1012         if (!set)
1013                 RETURN(NULL);
1014         atomic_set(&set->set_refcount, 1);
1015         INIT_LIST_HEAD(&set->set_requests);
1016         init_waitqueue_head(&set->set_waitq);
1017         atomic_set(&set->set_new_count, 0);
1018         atomic_set(&set->set_remaining, 0);
1019         spin_lock_init(&set->set_new_req_lock);
1020         INIT_LIST_HEAD(&set->set_new_requests);
1021         set->set_max_inflight = UINT_MAX;
1022         set->set_producer     = NULL;
1023         set->set_producer_arg = NULL;
1024         set->set_rc           = 0;
1025
1026         RETURN(set);
1027 }
1028 EXPORT_SYMBOL(ptlrpc_prep_set);
1029
1030 /**
1031  * Allocate and initialize new request set structure with flow control
1032  * extension. This extension allows to control the number of requests in-flight
1033  * for the whole set. A callback function to generate requests must be provided
1034  * and the request set will keep the number of requests sent over the wire to
1035  * @max_inflight.
1036  * Returns a pointer to the newly allocated set structure or NULL on error.
1037  */
1038 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1039                                              void *arg)
1040
1041 {
1042         struct ptlrpc_request_set *set;
1043
1044         set = ptlrpc_prep_set();
1045         if (!set)
1046                 RETURN(NULL);
1047
1048         set->set_max_inflight  = max;
1049         set->set_producer      = func;
1050         set->set_producer_arg  = arg;
1051
1052         RETURN(set);
1053 }
1054
1055 /**
1056  * Wind down and free request set structure previously allocated with
1057  * ptlrpc_prep_set.
1058  * Ensures that all requests on the set have completed and removes
1059  * all requests from the request list in a set.
1060  * If any unsent request happen to be on the list, pretends that they got
1061  * an error in flight and calls their completion handler.
1062  */
1063 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1064 {
1065         struct list_head *tmp;
1066         struct list_head *next;
1067         int expected_phase;
1068         int n = 0;
1069
1070         ENTRY;
1071
1072         /* Requests on the set should either all be completed, or all be new */
1073         expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1074                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1075         list_for_each(tmp, &set->set_requests) {
1076                 struct ptlrpc_request *req =
1077                         list_entry(tmp, struct ptlrpc_request,
1078                                    rq_set_chain);
1079
1080                 LASSERT(req->rq_phase == expected_phase);
1081                 n++;
1082         }
1083
1084         LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1085                  atomic_read(&set->set_remaining) == n, "%d / %d\n",
1086                  atomic_read(&set->set_remaining), n);
1087
1088         list_for_each_safe(tmp, next, &set->set_requests) {
1089                 struct ptlrpc_request *req =
1090                         list_entry(tmp, struct ptlrpc_request,
1091                                    rq_set_chain);
1092                 list_del_init(&req->rq_set_chain);
1093
1094                 LASSERT(req->rq_phase == expected_phase);
1095
1096                 if (req->rq_phase == RQ_PHASE_NEW) {
1097                         ptlrpc_req_interpret(NULL, req, -EBADR);
1098                         atomic_dec(&set->set_remaining);
1099                 }
1100
1101                 spin_lock(&req->rq_lock);
1102                 req->rq_set = NULL;
1103                 req->rq_invalid_rqset = 0;
1104                 spin_unlock(&req->rq_lock);
1105
1106                 ptlrpc_req_finished(req);
1107         }
1108
1109         LASSERT(atomic_read(&set->set_remaining) == 0);
1110
1111         ptlrpc_reqset_put(set);
1112         EXIT;
1113 }
1114 EXPORT_SYMBOL(ptlrpc_set_destroy);
1115
1116 /**
1117  * Add a new request to the general purpose request set.
1118  * Assumes request reference from the caller.
1119  */
1120 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1121                         struct ptlrpc_request *req)
1122 {
1123         LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1124         LASSERT(list_empty(&req->rq_set_chain));
1125
1126         if (req->rq_allow_intr)
1127                 set->set_allow_intr = 1;
1128
1129         /* The set takes over the caller's request reference */
1130         list_add_tail(&req->rq_set_chain, &set->set_requests);
1131         req->rq_set = set;
1132         atomic_inc(&set->set_remaining);
1133         req->rq_queued_time = ktime_get_seconds();
1134
1135         if (req->rq_reqmsg)
1136                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1137
1138         if (set->set_producer)
1139                 /*
1140                  * If the request set has a producer callback, the RPC must be
1141                  * sent straight away
1142                  */
1143                 ptlrpc_send_new_req(req);
1144 }
1145 EXPORT_SYMBOL(ptlrpc_set_add_req);
1146
1147 /**
1148  * Add a request to a request with dedicated server thread
1149  * and wake the thread to make any necessary processing.
1150  * Currently only used for ptlrpcd.
1151  */
1152 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1153                             struct ptlrpc_request *req)
1154 {
1155         struct ptlrpc_request_set *set = pc->pc_set;
1156         int count, i;
1157
1158         LASSERT(req->rq_set == NULL);
1159         LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1160
1161         spin_lock(&set->set_new_req_lock);
1162         /*
1163          * The set takes over the caller's request reference.
1164          */
1165         req->rq_set = set;
1166         req->rq_queued_time = ktime_get_seconds();
1167         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1168         count = atomic_inc_return(&set->set_new_count);
1169         spin_unlock(&set->set_new_req_lock);
1170
1171         /* Only need to call wakeup once for the first entry. */
1172         if (count == 1) {
1173                 wake_up(&set->set_waitq);
1174
1175                 /*
1176                  * XXX: It maybe unnecessary to wakeup all the partners. But to
1177                  *      guarantee the async RPC can be processed ASAP, we have
1178                  *      no other better choice. It maybe fixed in future.
1179                  */
1180                 for (i = 0; i < pc->pc_npartners; i++)
1181                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1182         }
1183 }
1184
1185 /**
1186  * Based on the current state of the import, determine if the request
1187  * can be sent, is an error, or should be delayed.
1188  *
1189  * Returns true if this request should be delayed. If false, and
1190  * *status is set, then the request can not be sent and *status is the
1191  * error code.  If false and status is 0, then request can be sent.
1192  *
1193  * The imp->imp_lock must be held.
1194  */
1195 static int ptlrpc_import_delay_req(struct obd_import *imp,
1196                                    struct ptlrpc_request *req, int *status)
1197 {
1198         int delay = 0;
1199
1200         ENTRY;
1201         LASSERT(status);
1202         *status = 0;
1203
1204         if (req->rq_ctx_init || req->rq_ctx_fini) {
1205                 /* always allow ctx init/fini rpc go through */
1206         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1207                 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1208                 *status = -EIO;
1209         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1210                 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1211
1212                 /*
1213                  * pings or MDS-equivalent STATFS may safely
1214                  * race with umount
1215                  */
1216                 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1217                           D_HA : D_ERROR, req, "IMP_CLOSED");
1218                 *status = -EIO;
1219         } else if (ptlrpc_send_limit_expired(req)) {
1220                 /* probably doesn't need to be a D_ERROR afterinitial testing */
1221                 DEBUG_REQ(D_HA, req, "send limit expired");
1222                 *status = -ETIMEDOUT;
1223         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1224                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
1225                 ;/* allow CONNECT even if import is invalid */
1226                 if (atomic_read(&imp->imp_inval_count) != 0) {
1227                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1228                         *status = -EIO;
1229                 }
1230         } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1231                 if (!imp->imp_deactive)
1232                         DEBUG_REQ(D_NET, req, "IMP_INVALID");
1233                 *status = -ESHUTDOWN; /* b=12940 */
1234         } else if (req->rq_import_generation != imp->imp_generation) {
1235                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1236                 *status = -EIO;
1237         } else if (req->rq_send_state != imp->imp_state) {
1238                 /* invalidate in progress - any requests should be drop */
1239                 if (atomic_read(&imp->imp_inval_count) != 0) {
1240                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1241                         *status = -EIO;
1242                 } else if (req->rq_no_delay &&
1243                            imp->imp_generation != imp->imp_initiated_at) {
1244                         /* ignore nodelay for requests initiating connections */
1245                         *status = -EWOULDBLOCK;
1246                 } else if (req->rq_allow_replay &&
1247                            (imp->imp_state == LUSTRE_IMP_REPLAY ||
1248                             imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1249                             imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1250                             imp->imp_state == LUSTRE_IMP_RECOVER)) {
1251                         DEBUG_REQ(D_HA, req, "allow during recovery");
1252                 } else {
1253                         delay = 1;
1254                 }
1255         }
1256
1257         RETURN(delay);
1258 }
1259
1260 /**
1261  * Decide if the error message should be printed to the console or not.
1262  * Makes its decision based on request type, status, and failure frequency.
1263  *
1264  * \param[in] req  request that failed and may need a console message
1265  *
1266  * \retval false if no message should be printed
1267  * \retval true  if console message should be printed
1268  */
1269 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1270 {
1271         LASSERT(req->rq_reqmsg != NULL);
1272
1273         /* Suppress particular reconnect errors which are to be expected. */
1274         if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1275                 /* Suppress timed out reconnect requests */
1276                 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1277                     req->rq_timedout)
1278                         return false;
1279
1280                 /*
1281                  * Suppress most unavailable/again reconnect requests, but
1282                  * print occasionally so it is clear client is trying to
1283                  * connect to a server where no target is running.
1284                  */
1285                 if ((err == -ENODEV || err == -EAGAIN) &&
1286                     req->rq_import->imp_conn_cnt % 30 != 20)
1287                         return false;
1288         }
1289
1290         if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1291                 /* -EAGAIN is normal when using POSIX flocks */
1292                 return false;
1293
1294         if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1295             (req->rq_xid & 0xf) != 10)
1296                 /* Suppress most ping requests, they may fail occasionally */
1297                 return false;
1298
1299         return true;
1300 }
1301
1302 /**
1303  * Check request processing status.
1304  * Returns the status.
1305  */
1306 static int ptlrpc_check_status(struct ptlrpc_request *req)
1307 {
1308         int rc;
1309
1310         ENTRY;
1311         rc = lustre_msg_get_status(req->rq_repmsg);
1312         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1313                 struct obd_import *imp = req->rq_import;
1314                 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1315                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1316
1317                 if (ptlrpc_console_allow(req, opc, rc))
1318                         LCONSOLE_ERROR_MSG(0x11,
1319                                            "%s: operation %s to node %s failed: rc = %d\n",
1320                                            imp->imp_obd->obd_name,
1321                                            ll_opcode2str(opc),
1322                                            libcfs_nid2str(nid), rc);
1323                 RETURN(rc < 0 ? rc : -EINVAL);
1324         }
1325
1326         if (rc)
1327                 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1328
1329         RETURN(rc);
1330 }
1331
1332 /**
1333  * save pre-versions of objects into request for replay.
1334  * Versions are obtained from server reply.
1335  * used for VBR.
1336  */
1337 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1338 {
1339         struct lustre_msg *repmsg = req->rq_repmsg;
1340         struct lustre_msg *reqmsg = req->rq_reqmsg;
1341         __u64 *versions = lustre_msg_get_versions(repmsg);
1342
1343         ENTRY;
1344         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1345                 return;
1346
1347         LASSERT(versions);
1348         lustre_msg_set_versions(reqmsg, versions);
1349         CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1350                versions[0], versions[1]);
1351
1352         EXIT;
1353 }
1354
1355 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1356 {
1357         struct ptlrpc_request *req;
1358
1359         assert_spin_locked(&imp->imp_lock);
1360         if (list_empty(&imp->imp_unreplied_list))
1361                 return 0;
1362
1363         req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1364                          rq_unreplied_list);
1365         LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1366
1367         if (imp->imp_known_replied_xid < req->rq_xid - 1)
1368                 imp->imp_known_replied_xid = req->rq_xid - 1;
1369
1370         return req->rq_xid - 1;
1371 }
1372
1373 /**
1374  * Callback function called when client receives RPC reply for \a req.
1375  * Returns 0 on success or error code.
1376  * The return alue would be assigned to req->rq_status by the caller
1377  * as request processing status.
1378  * This function also decides if the request needs to be saved for later replay.
1379  */
1380 static int after_reply(struct ptlrpc_request *req)
1381 {
1382         struct obd_import *imp = req->rq_import;
1383         struct obd_device *obd = req->rq_import->imp_obd;
1384         ktime_t work_start;
1385         u64 committed;
1386         s64 timediff;
1387         int rc;
1388
1389         ENTRY;
1390         LASSERT(obd != NULL);
1391         /* repbuf must be unlinked */
1392         LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1393
1394         if (req->rq_reply_truncated) {
1395                 if (ptlrpc_no_resend(req)) {
1396                         DEBUG_REQ(D_ERROR, req,
1397                                   "reply buffer overflow, expected=%d, actual size=%d",
1398                                   req->rq_nob_received, req->rq_repbuf_len);
1399                         RETURN(-EOVERFLOW);
1400                 }
1401
1402                 sptlrpc_cli_free_repbuf(req);
1403                 /*
1404                  * Pass the required reply buffer size (include
1405                  * space for early reply).
1406                  * NB: no need to roundup because alloc_repbuf
1407                  * will roundup it
1408                  */
1409                 req->rq_replen = req->rq_nob_received;
1410                 req->rq_nob_received = 0;
1411                 spin_lock(&req->rq_lock);
1412                 req->rq_resend       = 1;
1413                 spin_unlock(&req->rq_lock);
1414                 RETURN(0);
1415         }
1416
1417         work_start = ktime_get_real();
1418         timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1419
1420         /*
1421          * NB Until this point, the whole of the incoming message,
1422          * including buflens, status etc is in the sender's byte order.
1423          */
1424         rc = sptlrpc_cli_unwrap_reply(req);
1425         if (rc) {
1426                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1427                 RETURN(rc);
1428         }
1429
1430         /*
1431          * Security layer unwrap might ask resend this request.
1432          */
1433         if (req->rq_resend)
1434                 RETURN(0);
1435
1436         rc = unpack_reply(req);
1437         if (rc)
1438                 RETURN(rc);
1439
1440         /* retry indefinitely on EINPROGRESS */
1441         if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1442             ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1443                 time64_t now = ktime_get_real_seconds();
1444
1445                 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1446                           D_RPCTRACE, req, "resending request on EINPROGRESS");
1447                 spin_lock(&req->rq_lock);
1448                 req->rq_resend = 1;
1449                 spin_unlock(&req->rq_lock);
1450                 req->rq_nr_resend++;
1451
1452                 /* Readjust the timeout for current conditions */
1453                 ptlrpc_at_set_req_timeout(req);
1454                 /*
1455                  * delay resend to give a chance to the server to get ready.
1456                  * The delay is increased by 1s on every resend and is capped to
1457                  * the current request timeout (i.e. obd_timeout if AT is off,
1458                  * or AT service time x 125% + 5s, see at_est2timeout)
1459                  */
1460                 if (req->rq_nr_resend > req->rq_timeout)
1461                         req->rq_sent = now + req->rq_timeout;
1462                 else
1463                         req->rq_sent = now + req->rq_nr_resend;
1464
1465                 /* Resend for EINPROGRESS will use a new XID */
1466                 spin_lock(&imp->imp_lock);
1467                 list_del_init(&req->rq_unreplied_list);
1468                 spin_unlock(&imp->imp_lock);
1469
1470                 RETURN(0);
1471         }
1472
1473         if (obd->obd_svc_stats) {
1474                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1475                                     timediff);
1476                 ptlrpc_lprocfs_rpc_sent(req, timediff);
1477         }
1478
1479         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1480             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1481                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1482                           lustre_msg_get_type(req->rq_repmsg));
1483                 RETURN(-EPROTO);
1484         }
1485
1486         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1487                 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1488         ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1489         ptlrpc_at_adj_net_latency(req,
1490                                   lustre_msg_get_service_time(req->rq_repmsg));
1491
1492         rc = ptlrpc_check_status(req);
1493
1494         if (rc) {
1495                 /*
1496                  * Either we've been evicted, or the server has failed for
1497                  * some reason. Try to reconnect, and if that fails, punt to
1498                  * the upcall.
1499                  */
1500                 if (ptlrpc_recoverable_error(rc)) {
1501                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
1502                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1503                                 RETURN(rc);
1504                         }
1505                         ptlrpc_request_handle_notconn(req);
1506                         RETURN(rc);
1507                 }
1508         } else {
1509                 /*
1510                  * Let's look if server sent slv. Do it only for RPC with
1511                  * rc == 0.
1512                  */
1513                 ldlm_cli_update_pool(req);
1514         }
1515
1516         /*
1517          * Store transno in reqmsg for replay.
1518          */
1519         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1520                 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1521                 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1522         }
1523
1524         if (imp->imp_replayable) {
1525                 spin_lock(&imp->imp_lock);
1526                 /*
1527                  * No point in adding already-committed requests to the replay
1528                  * list, we will just remove them immediately. b=9829
1529                  */
1530                 if (req->rq_transno != 0 &&
1531                     (req->rq_transno >
1532                      lustre_msg_get_last_committed(req->rq_repmsg) ||
1533                      req->rq_replay)) {
1534                         /** version recovery */
1535                         ptlrpc_save_versions(req);
1536                         ptlrpc_retain_replayable_request(req, imp);
1537                 } else if (req->rq_commit_cb &&
1538                            list_empty(&req->rq_replay_list)) {
1539                         /*
1540                          * NB: don't call rq_commit_cb if it's already on
1541                          * rq_replay_list, ptlrpc_free_committed() will call
1542                          * it later, see LU-3618 for details
1543                          */
1544                         spin_unlock(&imp->imp_lock);
1545                         req->rq_commit_cb(req);
1546                         spin_lock(&imp->imp_lock);
1547                 }
1548
1549                 /*
1550                  * Replay-enabled imports return commit-status information.
1551                  */
1552                 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1553                 if (likely(committed > imp->imp_peer_committed_transno))
1554                         imp->imp_peer_committed_transno = committed;
1555
1556                 ptlrpc_free_committed(imp);
1557
1558                 if (!list_empty(&imp->imp_replay_list)) {
1559                         struct ptlrpc_request *last;
1560
1561                         last = list_entry(imp->imp_replay_list.prev,
1562                                           struct ptlrpc_request,
1563                                           rq_replay_list);
1564                         /*
1565                          * Requests with rq_replay stay on the list even if no
1566                          * commit is expected.
1567                          */
1568                         if (last->rq_transno > imp->imp_peer_committed_transno)
1569                                 ptlrpc_pinger_commit_expected(imp);
1570                 }
1571
1572                 spin_unlock(&imp->imp_lock);
1573         }
1574
1575         RETURN(rc);
1576 }
1577
1578 /**
1579  * Helper function to send request \a req over the network for the first time
1580  * Also adjusts request phase.
1581  * Returns 0 on success or error code.
1582  */
1583 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1584 {
1585         struct obd_import *imp = req->rq_import;
1586         __u64 min_xid = 0;
1587         int rc;
1588
1589         ENTRY;
1590         LASSERT(req->rq_phase == RQ_PHASE_NEW);
1591
1592         /* do not try to go further if there is not enough memory in enc_pool */
1593         if (req->rq_sent && req->rq_bulk)
1594                 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1595                     pool_is_at_full_capacity())
1596                         RETURN(-ENOMEM);
1597
1598         if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1599             (!req->rq_generation_set ||
1600              req->rq_import_generation == imp->imp_generation))
1601                 RETURN(0);
1602
1603         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1604
1605         spin_lock(&imp->imp_lock);
1606
1607         LASSERT(req->rq_xid != 0);
1608         LASSERT(!list_empty(&req->rq_unreplied_list));
1609
1610         if (!req->rq_generation_set)
1611                 req->rq_import_generation = imp->imp_generation;
1612
1613         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1614                 spin_lock(&req->rq_lock);
1615                 req->rq_waiting = 1;
1616                 spin_unlock(&req->rq_lock);
1617
1618                 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1619                           ptlrpc_import_state_name(req->rq_send_state),
1620                           ptlrpc_import_state_name(imp->imp_state));
1621                 LASSERT(list_empty(&req->rq_list));
1622                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1623                 atomic_inc(&req->rq_import->imp_inflight);
1624                 spin_unlock(&imp->imp_lock);
1625                 RETURN(0);
1626         }
1627
1628         if (rc != 0) {
1629                 spin_unlock(&imp->imp_lock);
1630                 req->rq_status = rc;
1631                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1632                 RETURN(rc);
1633         }
1634
1635         LASSERT(list_empty(&req->rq_list));
1636         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1637         atomic_inc(&req->rq_import->imp_inflight);
1638
1639         /*
1640          * find the known replied XID from the unreplied list, CONNECT
1641          * and DISCONNECT requests are skipped to make the sanity check
1642          * on server side happy. see process_req_last_xid().
1643          *
1644          * For CONNECT: Because replay requests have lower XID, it'll
1645          * break the sanity check if CONNECT bump the exp_last_xid on
1646          * server.
1647          *
1648          * For DISCONNECT: Since client will abort inflight RPC before
1649          * sending DISCONNECT, DISCONNECT may carry an XID which higher
1650          * than the inflight RPC.
1651          */
1652         if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1653                 min_xid = ptlrpc_known_replied_xid(imp);
1654         spin_unlock(&imp->imp_lock);
1655
1656         lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1657
1658         lustre_msg_set_status(req->rq_reqmsg, current_pid());
1659
1660         rc = sptlrpc_req_refresh_ctx(req, -1);
1661         if (rc) {
1662                 if (req->rq_err) {
1663                         req->rq_status = rc;
1664                         RETURN(1);
1665                 } else {
1666                         spin_lock(&req->rq_lock);
1667                         req->rq_wait_ctx = 1;
1668                         spin_unlock(&req->rq_lock);
1669                         RETURN(0);
1670                 }
1671         }
1672
1673         CDEBUG(D_RPCTRACE,
1674                "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1675                req, current_comm(),
1676                imp->imp_obd->obd_uuid.uuid,
1677                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1678                obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1679                lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1680
1681         rc = ptl_send_rpc(req, 0);
1682         if (rc == -ENOMEM) {
1683                 spin_lock(&imp->imp_lock);
1684                 if (!list_empty(&req->rq_list)) {
1685                         list_del_init(&req->rq_list);
1686                         atomic_dec(&req->rq_import->imp_inflight);
1687                 }
1688                 spin_unlock(&imp->imp_lock);
1689                 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1690                 RETURN(rc);
1691         }
1692         if (rc) {
1693                 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1694                           rc);
1695                 spin_lock(&req->rq_lock);
1696                 req->rq_net_err = 1;
1697                 spin_unlock(&req->rq_lock);
1698                 RETURN(rc);
1699         }
1700         RETURN(0);
1701 }
1702
1703 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1704 {
1705         int remaining, rc;
1706
1707         ENTRY;
1708         LASSERT(set->set_producer != NULL);
1709
1710         remaining = atomic_read(&set->set_remaining);
1711
1712         /*
1713          * populate the ->set_requests list with requests until we
1714          * reach the maximum number of RPCs in flight for this set
1715          */
1716         while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1717                 rc = set->set_producer(set, set->set_producer_arg);
1718                 if (rc == -ENOENT) {
1719                         /* no more RPC to produce */
1720                         set->set_producer     = NULL;
1721                         set->set_producer_arg = NULL;
1722                         RETURN(0);
1723                 }
1724         }
1725
1726         RETURN((atomic_read(&set->set_remaining) - remaining));
1727 }
1728
1729 /**
1730  * this sends any unsent RPCs in \a set and returns 1 if all are sent
1731  * and no more replies are expected.
1732  * (it is possible to get less replies than requests sent e.g. due to timed out
1733  * requests or requests that we had trouble to send out)
1734  *
1735  * NOTE: This function contains a potential schedule point (cond_resched()).
1736  */
1737 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1738 {
1739         struct list_head *tmp, *next;
1740         struct list_head  comp_reqs;
1741         int force_timer_recalc = 0;
1742
1743         ENTRY;
1744         if (atomic_read(&set->set_remaining) == 0)
1745                 RETURN(1);
1746
1747         INIT_LIST_HEAD(&comp_reqs);
1748         list_for_each_safe(tmp, next, &set->set_requests) {
1749                 struct ptlrpc_request *req =
1750                         list_entry(tmp, struct ptlrpc_request,
1751                                    rq_set_chain);
1752                 struct obd_import *imp = req->rq_import;
1753                 int unregistered = 0;
1754                 int async = 1;
1755                 int rc = 0;
1756
1757                 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1758                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1759                         continue;
1760                 }
1761
1762                 /*
1763                  * This schedule point is mainly for the ptlrpcd caller of this
1764                  * function.  Most ptlrpc sets are not long-lived and unbounded
1765                  * in length, but at the least the set used by the ptlrpcd is.
1766                  * Since the processing time is unbounded, we need to insert an
1767                  * explicit schedule point to make the thread well-behaved.
1768                  */
1769                 cond_resched();
1770
1771                 /*
1772                  * If the caller requires to allow to be interpreted by force
1773                  * and it has really been interpreted, then move the request
1774                  * to RQ_PHASE_INTERPRET phase in spite of what the current
1775                  * phase is.
1776                  */
1777                 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1778                         req->rq_status = -EINTR;
1779                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1780
1781                         /*
1782                          * Since it is interpreted and we have to wait for
1783                          * the reply to be unlinked, then use sync mode.
1784                          */
1785                         async = 0;
1786
1787                         GOTO(interpret, req->rq_status);
1788                 }
1789
1790                 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1791                         force_timer_recalc = 1;
1792
1793                 /* delayed send - skip */
1794                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1795                         continue;
1796
1797                 /* delayed resend - skip */
1798                 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1799                     req->rq_sent > ktime_get_real_seconds())
1800                         continue;
1801
1802                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1803                       req->rq_phase == RQ_PHASE_BULK ||
1804                       req->rq_phase == RQ_PHASE_INTERPRET ||
1805                       req->rq_phase == RQ_PHASE_UNREG_RPC ||
1806                       req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1807                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1808                         LBUG();
1809                 }
1810
1811                 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1812                     req->rq_phase == RQ_PHASE_UNREG_BULK) {
1813                         LASSERT(req->rq_next_phase != req->rq_phase);
1814                         LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1815
1816                         if (req->rq_req_deadline &&
1817                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1818                                 req->rq_req_deadline = 0;
1819                         if (req->rq_reply_deadline &&
1820                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1821                                 req->rq_reply_deadline = 0;
1822                         if (req->rq_bulk_deadline &&
1823                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1824                                 req->rq_bulk_deadline = 0;
1825
1826                         /*
1827                          * Skip processing until reply is unlinked. We
1828                          * can't return to pool before that and we can't
1829                          * call interpret before that. We need to make
1830                          * sure that all rdma transfers finished and will
1831                          * not corrupt any data.
1832                          */
1833                         if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1834                             ptlrpc_client_recv_or_unlink(req))
1835                                 continue;
1836                         if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1837                             ptlrpc_client_bulk_active(req))
1838                                 continue;
1839
1840                         /*
1841                          * Turn fail_loc off to prevent it from looping
1842                          * forever.
1843                          */
1844                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1845                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1846                                                      OBD_FAIL_ONCE);
1847                         }
1848                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1849                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1850                                                      OBD_FAIL_ONCE);
1851                         }
1852
1853                         /*
1854                          * Move to next phase if reply was successfully
1855                          * unlinked.
1856                          */
1857                         ptlrpc_rqphase_move(req, req->rq_next_phase);
1858                 }
1859
1860                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1861                         GOTO(interpret, req->rq_status);
1862
1863                 /*
1864                  * Note that this also will start async reply unlink.
1865                  */
1866                 if (req->rq_net_err && !req->rq_timedout) {
1867                         ptlrpc_expire_one_request(req, 1);
1868
1869                         /*
1870                          * Check if we still need to wait for unlink.
1871                          */
1872                         if (ptlrpc_client_recv_or_unlink(req) ||
1873                             ptlrpc_client_bulk_active(req))
1874                                 continue;
1875                         /* If there is no need to resend, fail it now. */
1876                         if (req->rq_no_resend) {
1877                                 if (req->rq_status == 0)
1878                                         req->rq_status = -EIO;
1879                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1880                                 GOTO(interpret, req->rq_status);
1881                         } else {
1882                                 continue;
1883                         }
1884                 }
1885
1886                 if (req->rq_err) {
1887                         spin_lock(&req->rq_lock);
1888                         req->rq_replied = 0;
1889                         spin_unlock(&req->rq_lock);
1890                         if (req->rq_status == 0)
1891                                 req->rq_status = -EIO;
1892                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1893                         GOTO(interpret, req->rq_status);
1894                 }
1895
1896                 /*
1897                  * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1898                  * so it sets rq_intr regardless of individual rpc
1899                  * timeouts. The synchronous IO waiting path sets
1900                  * rq_intr irrespective of whether ptlrpcd
1901                  * has seen a timeout.  Our policy is to only interpret
1902                  * interrupted rpcs after they have timed out, so we
1903                  * need to enforce that here.
1904                  */
1905
1906                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1907                                      req->rq_wait_ctx)) {
1908                         req->rq_status = -EINTR;
1909                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1910                         GOTO(interpret, req->rq_status);
1911                 }
1912
1913                 if (req->rq_phase == RQ_PHASE_RPC) {
1914                         if (req->rq_timedout || req->rq_resend ||
1915                             req->rq_waiting || req->rq_wait_ctx) {
1916                                 int status;
1917
1918                                 if (!ptlrpc_unregister_reply(req, 1)) {
1919                                         ptlrpc_unregister_bulk(req, 1);
1920                                         continue;
1921                                 }
1922
1923                                 spin_lock(&imp->imp_lock);
1924                                 if (ptlrpc_import_delay_req(imp, req,
1925                                                             &status)) {
1926                                         /*
1927                                          * put on delay list - only if we wait
1928                                          * recovery finished - before send
1929                                          */
1930                                         list_move_tail(&req->rq_list,
1931                                                        &imp->imp_delayed_list);
1932                                         spin_unlock(&imp->imp_lock);
1933                                         continue;
1934                                 }
1935
1936                                 if (status != 0)  {
1937                                         req->rq_status = status;
1938                                         ptlrpc_rqphase_move(req,
1939                                                             RQ_PHASE_INTERPRET);
1940                                         spin_unlock(&imp->imp_lock);
1941                                         GOTO(interpret, req->rq_status);
1942                                 }
1943                                 /* ignore on just initiated connections */
1944                                 if (ptlrpc_no_resend(req) &&
1945                                     !req->rq_wait_ctx &&
1946                                     imp->imp_generation !=
1947                                     imp->imp_initiated_at) {
1948                                         req->rq_status = -ENOTCONN;
1949                                         ptlrpc_rqphase_move(req,
1950                                                             RQ_PHASE_INTERPRET);
1951                                         spin_unlock(&imp->imp_lock);
1952                                         GOTO(interpret, req->rq_status);
1953                                 }
1954
1955                                 list_move_tail(&req->rq_list,
1956                                                &imp->imp_sending_list);
1957
1958                                 spin_unlock(&imp->imp_lock);
1959
1960                                 spin_lock(&req->rq_lock);
1961                                 req->rq_waiting = 0;
1962                                 spin_unlock(&req->rq_lock);
1963
1964                                 if (req->rq_timedout || req->rq_resend) {
1965                                         /*
1966                                          * This is re-sending anyways,
1967                                          * let's mark req as resend.
1968                                          */
1969                                         spin_lock(&req->rq_lock);
1970                                         req->rq_resend = 1;
1971                                         spin_unlock(&req->rq_lock);
1972                                 }
1973                                 /*
1974                                  * rq_wait_ctx is only touched by ptlrpcd,
1975                                  * so no lock is needed here.
1976                                  */
1977                                 status = sptlrpc_req_refresh_ctx(req, -1);
1978                                 if (status) {
1979                                         if (req->rq_err) {
1980                                                 req->rq_status = status;
1981                                                 spin_lock(&req->rq_lock);
1982                                                 req->rq_wait_ctx = 0;
1983                                                 spin_unlock(&req->rq_lock);
1984                                                 force_timer_recalc = 1;
1985                                         } else {
1986                                                 spin_lock(&req->rq_lock);
1987                                                 req->rq_wait_ctx = 1;
1988                                                 spin_unlock(&req->rq_lock);
1989                                         }
1990
1991                                         continue;
1992                                 } else {
1993                                         spin_lock(&req->rq_lock);
1994                                         req->rq_wait_ctx = 0;
1995                                         spin_unlock(&req->rq_lock);
1996                                 }
1997
1998                                 /*
1999                                  * In any case, the previous bulk should be
2000                                  * cleaned up to prepare for the new sending
2001                                  */
2002                                 if (req->rq_bulk &&
2003                                     !ptlrpc_unregister_bulk(req, 1))
2004                                         continue;
2005
2006                                 rc = ptl_send_rpc(req, 0);
2007                                 if (rc == -ENOMEM) {
2008                                         spin_lock(&imp->imp_lock);
2009                                         if (!list_empty(&req->rq_list))
2010                                                 list_del_init(&req->rq_list);
2011                                         spin_unlock(&imp->imp_lock);
2012                                         ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2013                                         continue;
2014                                 }
2015                                 if (rc) {
2016                                         DEBUG_REQ(D_HA, req,
2017                                                   "send failed: rc = %d", rc);
2018                                         force_timer_recalc = 1;
2019                                         spin_lock(&req->rq_lock);
2020                                         req->rq_net_err = 1;
2021                                         spin_unlock(&req->rq_lock);
2022                                         continue;
2023                                 }
2024                                 /* need to reset the timeout */
2025                                 force_timer_recalc = 1;
2026                         }
2027
2028                         spin_lock(&req->rq_lock);
2029
2030                         if (ptlrpc_client_early(req)) {
2031                                 ptlrpc_at_recv_early_reply(req);
2032                                 spin_unlock(&req->rq_lock);
2033                                 continue;
2034                         }
2035
2036                         /* Still waiting for a reply? */
2037                         if (ptlrpc_client_recv(req)) {
2038                                 spin_unlock(&req->rq_lock);
2039                                 continue;
2040                         }
2041
2042                         /* Did we actually receive a reply? */
2043                         if (!ptlrpc_client_replied(req)) {
2044                                 spin_unlock(&req->rq_lock);
2045                                 continue;
2046                         }
2047
2048                         spin_unlock(&req->rq_lock);
2049
2050                         /*
2051                          * unlink from net because we are going to
2052                          * swab in-place of reply buffer
2053                          */
2054                         unregistered = ptlrpc_unregister_reply(req, 1);
2055                         if (!unregistered)
2056                                 continue;
2057
2058                         req->rq_status = after_reply(req);
2059                         if (req->rq_resend)
2060                                 continue;
2061
2062                         /*
2063                          * If there is no bulk associated with this request,
2064                          * then we're done and should let the interpreter
2065                          * process the reply. Similarly if the RPC returned
2066                          * an error, and therefore the bulk will never arrive.
2067                          */
2068                         if (!req->rq_bulk || req->rq_status < 0) {
2069                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2070                                 GOTO(interpret, req->rq_status);
2071                         }
2072
2073                         ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2074                 }
2075
2076                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2077                 if (ptlrpc_client_bulk_active(req))
2078                         continue;
2079
2080                 if (req->rq_bulk->bd_failure) {
2081                         /*
2082                          * The RPC reply arrived OK, but the bulk screwed
2083                          * up!  Dead weird since the server told us the RPC
2084                          * was good after getting the REPLY for her GET or
2085                          * the ACK for her PUT.
2086                          */
2087                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2088                         req->rq_status = -EIO;
2089                 }
2090
2091                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2092
2093 interpret:
2094                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2095
2096                 /*
2097                  * This moves to "unregistering" phase we need to wait for
2098                  * reply unlink.
2099                  */
2100                 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2101                         /* start async bulk unlink too */
2102                         ptlrpc_unregister_bulk(req, 1);
2103                         continue;
2104                 }
2105
2106                 if (!ptlrpc_unregister_bulk(req, async))
2107                         continue;
2108
2109                 /*
2110                  * When calling interpret receiving already should be
2111                  * finished.
2112                  */
2113                 LASSERT(!req->rq_receiving_reply);
2114
2115                 ptlrpc_req_interpret(env, req, req->rq_status);
2116
2117                 if (ptlrpcd_check_work(req)) {
2118                         atomic_dec(&set->set_remaining);
2119                         continue;
2120                 }
2121                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2122
2123                 if (req->rq_reqmsg)
2124                         CDEBUG(D_RPCTRACE,
2125                                "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2126                                req, current_comm(),
2127                                imp->imp_obd->obd_uuid.uuid,
2128                                lustre_msg_get_status(req->rq_reqmsg),
2129                                req->rq_xid,
2130                                obd_import_nid2str(imp),
2131                                lustre_msg_get_opc(req->rq_reqmsg),
2132                                lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2133
2134                 spin_lock(&imp->imp_lock);
2135                 /*
2136                  * Request already may be not on sending or delaying list. This
2137                  * may happen in the case of marking it erroneous for the case
2138                  * ptlrpc_import_delay_req(req, status) find it impossible to
2139                  * allow sending this rpc and returns *status != 0.
2140                  */
2141                 if (!list_empty(&req->rq_list)) {
2142                         list_del_init(&req->rq_list);
2143                         atomic_dec(&imp->imp_inflight);
2144                 }
2145                 list_del_init(&req->rq_unreplied_list);
2146                 spin_unlock(&imp->imp_lock);
2147
2148                 atomic_dec(&set->set_remaining);
2149                 wake_up_all(&imp->imp_recovery_waitq);
2150
2151                 if (set->set_producer) {
2152                         /* produce a new request if possible */
2153                         if (ptlrpc_set_producer(set) > 0)
2154                                 force_timer_recalc = 1;
2155
2156                         /*
2157                          * free the request that has just been completed
2158                          * in order not to pollute set->set_requests
2159                          */
2160                         list_del_init(&req->rq_set_chain);
2161                         spin_lock(&req->rq_lock);
2162                         req->rq_set = NULL;
2163                         req->rq_invalid_rqset = 0;
2164                         spin_unlock(&req->rq_lock);
2165
2166                         /* record rq_status to compute the final status later */
2167                         if (req->rq_status != 0)
2168                                 set->set_rc = req->rq_status;
2169                         ptlrpc_req_finished(req);
2170                 } else {
2171                         list_move_tail(&req->rq_set_chain, &comp_reqs);
2172                 }
2173         }
2174
2175         /*
2176          * move completed request at the head of list so it's easier for
2177          * caller to find them
2178          */
2179         list_splice(&comp_reqs, &set->set_requests);
2180
2181         /* If we hit an error, we want to recover promptly. */
2182         RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2183 }
2184 EXPORT_SYMBOL(ptlrpc_check_set);
2185
2186 /**
2187  * Time out request \a req. is \a async_unlink is set, that means do not wait
2188  * until LNet actually confirms network buffer unlinking.
2189  * Return 1 if we should give up further retrying attempts or 0 otherwise.
2190  */
2191 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2192 {
2193         struct obd_import *imp = req->rq_import;
2194         unsigned int debug_mask = D_RPCTRACE;
2195         int rc = 0;
2196
2197         ENTRY;
2198         spin_lock(&req->rq_lock);
2199         req->rq_timedout = 1;
2200         spin_unlock(&req->rq_lock);
2201
2202         if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2203                                  lustre_msg_get_status(req->rq_reqmsg)))
2204                 debug_mask = D_WARNING;
2205         DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2206                   req->rq_net_err ? "failed due to network error" :
2207                      ((req->rq_real_sent == 0 ||
2208                        req->rq_real_sent < req->rq_sent ||
2209                        req->rq_real_sent >= req->rq_deadline) ?
2210                       "timed out for sent delay" : "timed out for slow reply"),
2211                   (s64)req->rq_sent, (s64)req->rq_real_sent);
2212
2213         if (imp && obd_debug_peer_on_timeout)
2214                 LNetDebugPeer(imp->imp_connection->c_peer);
2215
2216         ptlrpc_unregister_reply(req, async_unlink);
2217         ptlrpc_unregister_bulk(req, async_unlink);
2218
2219         if (obd_dump_on_timeout)
2220                 libcfs_debug_dumplog();
2221
2222         if (!imp) {
2223                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2224                 RETURN(1);
2225         }
2226
2227         atomic_inc(&imp->imp_timeouts);
2228
2229         /* The DLM server doesn't want recovery run on its imports. */
2230         if (imp->imp_dlm_fake)
2231                 RETURN(1);
2232
2233         /*
2234          * If this request is for recovery or other primordial tasks,
2235          * then error it out here.
2236          */
2237         if (req->rq_ctx_init || req->rq_ctx_fini ||
2238             req->rq_send_state != LUSTRE_IMP_FULL ||
2239             imp->imp_obd->obd_no_recov) {
2240                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2241                           ptlrpc_import_state_name(req->rq_send_state),
2242                           ptlrpc_import_state_name(imp->imp_state));
2243                 spin_lock(&req->rq_lock);
2244                 req->rq_status = -ETIMEDOUT;
2245                 req->rq_err = 1;
2246                 spin_unlock(&req->rq_lock);
2247                 RETURN(1);
2248         }
2249
2250         /*
2251          * if a request can't be resent we can't wait for an answer after
2252          * the timeout
2253          */
2254         if (ptlrpc_no_resend(req)) {
2255                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2256                 rc = 1;
2257         }
2258
2259         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2260
2261         RETURN(rc);
2262 }
2263
2264 /**
2265  * Time out all uncompleted requests in request set pointed by \a data
2266  * Callback used when waiting on sets with l_wait_event.
2267  * Always returns 1.
2268  */
2269 int ptlrpc_expired_set(void *data)
2270 {
2271         struct ptlrpc_request_set *set = data;
2272         struct list_head *tmp;
2273         time64_t now = ktime_get_real_seconds();
2274
2275         ENTRY;
2276         LASSERT(set != NULL);
2277
2278         /*
2279          * A timeout expired. See which reqs it applies to...
2280          */
2281         list_for_each(tmp, &set->set_requests) {
2282                 struct ptlrpc_request *req =
2283                         list_entry(tmp, struct ptlrpc_request,
2284                                    rq_set_chain);
2285
2286                 /* don't expire request waiting for context */
2287                 if (req->rq_wait_ctx)
2288                         continue;
2289
2290                 /* Request in-flight? */
2291                 if (!((req->rq_phase == RQ_PHASE_RPC &&
2292                        !req->rq_waiting && !req->rq_resend) ||
2293                       (req->rq_phase == RQ_PHASE_BULK)))
2294                         continue;
2295
2296                 if (req->rq_timedout ||     /* already dealt with */
2297                     req->rq_deadline > now) /* not expired */
2298                         continue;
2299
2300                 /*
2301                  * Deal with this guy. Do it asynchronously to not block
2302                  * ptlrpcd thread.
2303                  */
2304                 ptlrpc_expire_one_request(req, 1);
2305         }
2306
2307         /*
2308          * When waiting for a whole set, we always break out of the
2309          * sleep so we can recalculate the timeout, or enable interrupts
2310          * if everyone's timed out.
2311          */
2312         RETURN(1);
2313 }
2314
2315 /**
2316  * Sets rq_intr flag in \a req under spinlock.
2317  */
2318 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2319 {
2320         spin_lock(&req->rq_lock);
2321         req->rq_intr = 1;
2322         spin_unlock(&req->rq_lock);
2323 }
2324 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2325
2326 /**
2327  * Interrupts (sets interrupted flag) all uncompleted requests in
2328  * a set \a data. Callback for l_wait_event for interruptible waits.
2329  */
2330 static void ptlrpc_interrupted_set(void *data)
2331 {
2332         struct ptlrpc_request_set *set = data;
2333         struct list_head *tmp;
2334
2335         LASSERT(set != NULL);
2336         CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2337
2338         list_for_each(tmp, &set->set_requests) {
2339                 struct ptlrpc_request *req =
2340                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2341
2342                 if (req->rq_intr)
2343                         continue;
2344
2345                 if (req->rq_phase != RQ_PHASE_RPC &&
2346                     req->rq_phase != RQ_PHASE_UNREG_RPC &&
2347                     !req->rq_allow_intr)
2348                         continue;
2349
2350                 ptlrpc_mark_interrupted(req);
2351         }
2352 }
2353
2354 /**
2355  * Get the smallest timeout in the set; this does NOT set a timeout.
2356  */
2357 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2358 {
2359         struct list_head *tmp;
2360         time64_t now = ktime_get_real_seconds();
2361         int timeout = 0;
2362         struct ptlrpc_request *req;
2363         time64_t deadline;
2364
2365         ENTRY;
2366         list_for_each(tmp, &set->set_requests) {
2367                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2368
2369                 /* Request in-flight? */
2370                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2371                       (req->rq_phase == RQ_PHASE_BULK) ||
2372                       (req->rq_phase == RQ_PHASE_NEW)))
2373                         continue;
2374
2375                 /* Already timed out. */
2376                 if (req->rq_timedout)
2377                         continue;
2378
2379                 /* Waiting for ctx. */
2380                 if (req->rq_wait_ctx)
2381                         continue;
2382
2383                 if (req->rq_phase == RQ_PHASE_NEW)
2384                         deadline = req->rq_sent;
2385                 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2386                         deadline = req->rq_sent;
2387                 else
2388                         deadline = req->rq_sent + req->rq_timeout;
2389
2390                 if (deadline <= now)    /* actually expired already */
2391                         timeout = 1;    /* ASAP */
2392                 else if (timeout == 0 || timeout > deadline - now)
2393                         timeout = deadline - now;
2394         }
2395         RETURN(timeout);
2396 }
2397
2398 /**
2399  * Send all unset request from the set and then wait untill all
2400  * requests in the set complete (either get a reply, timeout, get an
2401  * error or otherwise be interrupted).
2402  * Returns 0 on success or error code otherwise.
2403  */
2404 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2405 {
2406         struct list_head *tmp;
2407         struct ptlrpc_request *req;
2408         struct l_wait_info lwi;
2409         time64_t timeout;
2410         int rc;
2411
2412         ENTRY;
2413         if (set->set_producer)
2414                 (void)ptlrpc_set_producer(set);
2415         else
2416                 list_for_each(tmp, &set->set_requests) {
2417                         req = list_entry(tmp, struct ptlrpc_request,
2418                                          rq_set_chain);
2419                         if (req->rq_phase == RQ_PHASE_NEW)
2420                                 (void)ptlrpc_send_new_req(req);
2421                 }
2422
2423         if (list_empty(&set->set_requests))
2424                 RETURN(0);
2425
2426         do {
2427                 timeout = ptlrpc_set_next_timeout(set);
2428
2429                 /*
2430                  * wait until all complete, interrupted, or an in-flight
2431                  * req times out
2432                  */
2433                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2434                        set, timeout);
2435
2436                 if ((timeout == 0 && !signal_pending(current)) ||
2437                     set->set_allow_intr)
2438                         /*
2439                          * No requests are in-flight (ether timed out
2440                          * or delayed), so we can allow interrupts.
2441                          * We still want to block for a limited time,
2442                          * so we allow interrupts during the timeout.
2443                          */
2444                         lwi = LWI_TIMEOUT_INTR_ALL(
2445                                         cfs_time_seconds(timeout ? timeout : 1),
2446                                         ptlrpc_expired_set,
2447                                         ptlrpc_interrupted_set, set);
2448                 else
2449                         /*
2450                          * At least one request is in flight, so no
2451                          * interrupts are allowed. Wait until all
2452                          * complete, or an in-flight req times out.
2453                          */
2454                         lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2455                                           ptlrpc_expired_set, set);
2456
2457                 rc = l_wait_event(set->set_waitq,
2458                                   ptlrpc_check_set(NULL, set), &lwi);
2459
2460                 /*
2461                  * LU-769 - if we ignored the signal because it was already
2462                  * pending when we started, we need to handle it now or we risk
2463                  * it being ignored forever
2464                  */
2465                 if (rc == -ETIMEDOUT &&
2466                     (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2467                     signal_pending(current)) {
2468                         sigset_t blocked_sigs =
2469                                            cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2470
2471                         /*
2472                          * In fact we only interrupt for the "fatal" signals
2473                          * like SIGINT or SIGKILL. We still ignore less
2474                          * important signals since ptlrpc set is not easily
2475                          * reentrant from userspace again
2476                          */
2477                         if (signal_pending(current))
2478                                 ptlrpc_interrupted_set(set);
2479                         cfs_restore_sigs(blocked_sigs);
2480                 }
2481
2482                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2483
2484                 /*
2485                  * -EINTR => all requests have been flagged rq_intr so next
2486                  * check completes.
2487                  * -ETIMEDOUT => someone timed out.  When all reqs have
2488                  * timed out, signals are enabled allowing completion with
2489                  * EINTR.
2490                  * I don't really care if we go once more round the loop in
2491                  * the error cases -eeb.
2492                  */
2493                 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2494                         list_for_each(tmp, &set->set_requests) {
2495                                 req = list_entry(tmp, struct ptlrpc_request,
2496                                                  rq_set_chain);
2497                                 spin_lock(&req->rq_lock);
2498                                 req->rq_invalid_rqset = 1;
2499                                 spin_unlock(&req->rq_lock);
2500                         }
2501                 }
2502         } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2503
2504         LASSERT(atomic_read(&set->set_remaining) == 0);
2505
2506         rc = set->set_rc; /* rq_status of already freed requests if any */
2507         list_for_each(tmp, &set->set_requests) {
2508                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2509
2510                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2511                 if (req->rq_status != 0)
2512                         rc = req->rq_status;
2513         }
2514
2515         RETURN(rc);
2516 }
2517 EXPORT_SYMBOL(ptlrpc_set_wait);
2518
2519 /**
2520  * Helper fuction for request freeing.
2521  * Called when request count reached zero and request needs to be freed.
2522  * Removes request from all sorts of sending/replay lists it might be on,
2523  * frees network buffers if any are present.
2524  * If \a locked is set, that means caller is already holding import imp_lock
2525  * and so we no longer need to reobtain it (for certain lists manipulations)
2526  */
2527 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2528 {
2529         ENTRY;
2530
2531         if (!request)
2532                 RETURN_EXIT;
2533
2534         LASSERT(!request->rq_srv_req);
2535         LASSERT(request->rq_export == NULL);
2536         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2537         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2538         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2539         LASSERTF(!request->rq_replay, "req %p\n", request);
2540
2541         req_capsule_fini(&request->rq_pill);
2542
2543         /*
2544          * We must take it off the imp_replay_list first.  Otherwise, we'll set
2545          * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2546          */
2547         if (request->rq_import) {
2548                 if (!locked)
2549                         spin_lock(&request->rq_import->imp_lock);
2550                 list_del_init(&request->rq_replay_list);
2551                 list_del_init(&request->rq_unreplied_list);
2552                 if (!locked)
2553                         spin_unlock(&request->rq_import->imp_lock);
2554         }
2555         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2556
2557         if (atomic_read(&request->rq_refcount) != 0) {
2558                 DEBUG_REQ(D_ERROR, request,
2559                           "freeing request with nonzero refcount");
2560                 LBUG();
2561         }
2562
2563         if (request->rq_repbuf)
2564                 sptlrpc_cli_free_repbuf(request);
2565
2566         if (request->rq_import) {
2567                 class_import_put(request->rq_import);
2568                 request->rq_import = NULL;
2569         }
2570         if (request->rq_bulk)
2571                 ptlrpc_free_bulk(request->rq_bulk);
2572
2573         if (request->rq_reqbuf || request->rq_clrbuf)
2574                 sptlrpc_cli_free_reqbuf(request);
2575
2576         if (request->rq_cli_ctx)
2577                 sptlrpc_req_put_ctx(request, !locked);
2578
2579         if (request->rq_pool)
2580                 __ptlrpc_free_req_to_pool(request);
2581         else
2582                 ptlrpc_request_cache_free(request);
2583         EXIT;
2584 }
2585
2586 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2587 /**
2588  * Drop one request reference. Must be called with import imp_lock held.
2589  * When reference count drops to zero, request is freed.
2590  */
2591 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2592 {
2593         assert_spin_locked(&request->rq_import->imp_lock);
2594         (void)__ptlrpc_req_finished(request, 1);
2595 }
2596
2597 /**
2598  * Helper function
2599  * Drops one reference count for request \a request.
2600  * \a locked set indicates that caller holds import imp_lock.
2601  * Frees the request whe reference count reaches zero.
2602  *
2603  * \retval 1    the request is freed
2604  * \retval 0    some others still hold references on the request
2605  */
2606 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2607 {
2608         int count;
2609
2610         ENTRY;
2611         if (!request)
2612                 RETURN(1);
2613
2614         LASSERT(request != LP_POISON);
2615         LASSERT(request->rq_reqmsg != LP_POISON);
2616
2617         DEBUG_REQ(D_INFO, request, "refcount now %u",
2618                   atomic_read(&request->rq_refcount) - 1);
2619
2620         spin_lock(&request->rq_lock);
2621         count = atomic_dec_return(&request->rq_refcount);
2622         LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2623
2624         /*
2625          * For open RPC, the client does not know the EA size (LOV, ACL, and
2626          * so on) before replied, then the client has to reserve very large
2627          * reply buffer. Such buffer will not be released until the RPC freed.
2628          * Since The open RPC is replayable, we need to keep it in the replay
2629          * list until close. If there are a lot of files opened concurrently,
2630          * then the client may be OOM.
2631          *
2632          * If fact, it is unnecessary to keep reply buffer for open replay,
2633          * related EAs have already been saved via mdc_save_lovea() before
2634          * coming here. So it is safe to free the reply buffer some earlier
2635          * before releasing the RPC to avoid client OOM. LU-9514
2636          */
2637         if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2638                 spin_lock(&request->rq_early_free_lock);
2639                 sptlrpc_cli_free_repbuf(request);
2640                 request->rq_repbuf = NULL;
2641                 request->rq_repbuf_len = 0;
2642                 request->rq_repdata = NULL;
2643                 request->rq_reqdata_len = 0;
2644                 spin_unlock(&request->rq_early_free_lock);
2645         }
2646         spin_unlock(&request->rq_lock);
2647
2648         if (!count)
2649                 __ptlrpc_free_req(request, locked);
2650
2651         RETURN(!count);
2652 }
2653
2654 /**
2655  * Drops one reference count for a request.
2656  */
2657 void ptlrpc_req_finished(struct ptlrpc_request *request)
2658 {
2659         __ptlrpc_req_finished(request, 0);
2660 }
2661 EXPORT_SYMBOL(ptlrpc_req_finished);
2662
2663 /**
2664  * Returns xid of a \a request
2665  */
2666 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2667 {
2668         return request->rq_xid;
2669 }
2670 EXPORT_SYMBOL(ptlrpc_req_xid);
2671
2672 /**
2673  * Disengage the client's reply buffer from the network
2674  * NB does _NOT_ unregister any client-side bulk.
2675  * IDEMPOTENT, but _not_ safe against concurrent callers.
2676  * The request owner (i.e. the thread doing the I/O) must call...
2677  * Returns 0 on success or 1 if unregistering cannot be made.
2678  */
2679 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2680 {
2681         int rc;
2682         struct l_wait_info lwi;
2683
2684         /*
2685          * Might sleep.
2686          */
2687         LASSERT(!in_interrupt());
2688
2689         /* Let's setup deadline for reply unlink. */
2690         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2691             async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2692                 request->rq_reply_deadline = ktime_get_real_seconds() +
2693                                              LONG_UNLINK;
2694
2695         /*
2696          * Nothing left to do.
2697          */
2698         if (!ptlrpc_client_recv_or_unlink(request))
2699                 RETURN(1);
2700
2701         LNetMDUnlink(request->rq_reply_md_h);
2702
2703         /*
2704          * Let's check it once again.
2705          */
2706         if (!ptlrpc_client_recv_or_unlink(request))
2707                 RETURN(1);
2708
2709         /* Move to "Unregistering" phase as reply was not unlinked yet. */
2710         ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2711
2712         /*
2713          * Do not wait for unlink to finish.
2714          */
2715         if (async)
2716                 RETURN(0);
2717
2718         /*
2719          * We have to l_wait_event() whatever the result, to give liblustre
2720          * a chance to run reply_in_callback(), and to make sure we've
2721          * unlinked before returning a req to the pool.
2722          */
2723         for (;;) {
2724                 /* The wq argument is ignored by user-space wait_event macros */
2725                 wait_queue_head_t *wq = (request->rq_set) ?
2726                                         &request->rq_set->set_waitq :
2727                                         &request->rq_reply_waitq;
2728                 /*
2729                  * Network access will complete in finite time but the HUGE
2730                  * timeout lets us CWARN for visibility of sluggish NALs
2731                  */
2732                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2733                                            cfs_time_seconds(1), NULL, NULL);
2734                 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2735                                   &lwi);
2736                 if (rc == 0) {
2737                         ptlrpc_rqphase_move(request, request->rq_next_phase);
2738                         RETURN(1);
2739                 }
2740
2741                 LASSERT(rc == -ETIMEDOUT);
2742                 DEBUG_REQ(D_WARNING, request,
2743                           "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2744                           request->rq_receiving_reply,
2745                           request->rq_req_unlinked,
2746                           request->rq_reply_unlinked);
2747         }
2748         RETURN(0);
2749 }
2750
2751 static void ptlrpc_free_request(struct ptlrpc_request *req)
2752 {
2753         spin_lock(&req->rq_lock);
2754         req->rq_replay = 0;
2755         spin_unlock(&req->rq_lock);
2756
2757         if (req->rq_commit_cb)
2758                 req->rq_commit_cb(req);
2759         list_del_init(&req->rq_replay_list);
2760
2761         __ptlrpc_req_finished(req, 1);
2762 }
2763
2764 /**
2765  * the request is committed and dropped from the replay list of its import
2766  */
2767 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2768 {
2769         struct obd_import *imp = req->rq_import;
2770
2771         spin_lock(&imp->imp_lock);
2772         if (list_empty(&req->rq_replay_list)) {
2773                 spin_unlock(&imp->imp_lock);
2774                 return;
2775         }
2776
2777         if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2778                 if (imp->imp_replay_cursor == &req->rq_replay_list)
2779                         imp->imp_replay_cursor = req->rq_replay_list.next;
2780                 ptlrpc_free_request(req);
2781         }
2782
2783         spin_unlock(&imp->imp_lock);
2784 }
2785 EXPORT_SYMBOL(ptlrpc_request_committed);
2786
2787 /**
2788  * Iterates through replay_list on import and prunes
2789  * all requests have transno smaller than last_committed for the
2790  * import and don't have rq_replay set.
2791  * Since requests are sorted in transno order, stops when meetign first
2792  * transno bigger than last_committed.
2793  * caller must hold imp->imp_lock
2794  */
2795 void ptlrpc_free_committed(struct obd_import *imp)
2796 {
2797         struct ptlrpc_request *req, *saved;
2798         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2799         bool skip_committed_list = true;
2800
2801         ENTRY;
2802         LASSERT(imp != NULL);
2803         assert_spin_locked(&imp->imp_lock);
2804
2805         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2806             imp->imp_generation == imp->imp_last_generation_checked) {
2807                 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2808                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2809                 RETURN_EXIT;
2810         }
2811         CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2812                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2813                imp->imp_generation);
2814
2815         if (imp->imp_generation != imp->imp_last_generation_checked ||
2816             imp->imp_last_transno_checked == 0)
2817                 skip_committed_list = false;
2818
2819         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2820         imp->imp_last_generation_checked = imp->imp_generation;
2821
2822         list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2823                                  rq_replay_list) {
2824                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
2825                 LASSERT(req != last_req);
2826                 last_req = req;
2827
2828                 if (req->rq_transno == 0) {
2829                         DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2830                         LBUG();
2831                 }
2832                 if (req->rq_import_generation < imp->imp_generation) {
2833                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2834                         GOTO(free_req, 0);
2835                 }
2836
2837                 /* not yet committed */
2838                 if (req->rq_transno > imp->imp_peer_committed_transno) {
2839                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2840                         break;
2841                 }
2842
2843                 if (req->rq_replay) {
2844                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2845                         list_move_tail(&req->rq_replay_list,
2846                                        &imp->imp_committed_list);
2847                         continue;
2848                 }
2849
2850                 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2851                           imp->imp_peer_committed_transno);
2852 free_req:
2853                 ptlrpc_free_request(req);
2854         }
2855
2856         if (skip_committed_list)
2857                 GOTO(out, 0);
2858
2859         list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2860                                  rq_replay_list) {
2861                 LASSERT(req->rq_transno != 0);
2862                 if (req->rq_import_generation < imp->imp_generation ||
2863                     !req->rq_replay) {
2864                         DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2865                                   req->rq_import_generation <
2866                                   imp->imp_generation ? "stale" : "closed");
2867
2868                         if (imp->imp_replay_cursor == &req->rq_replay_list)
2869                                 imp->imp_replay_cursor =
2870                                         req->rq_replay_list.next;
2871
2872                         ptlrpc_free_request(req);
2873                 }
2874         }
2875 out:
2876         EXIT;
2877 }
2878
2879 void ptlrpc_cleanup_client(struct obd_import *imp)
2880 {
2881         ENTRY;
2882         EXIT;
2883 }
2884
2885 /**
2886  * Schedule previously sent request for resend.
2887  * For bulk requests we assign new xid (to avoid problems with
2888  * lost replies and therefore several transfers landing into same buffer
2889  * from different sending attempts).
2890  */
2891 void ptlrpc_resend_req(struct ptlrpc_request *req)
2892 {
2893         DEBUG_REQ(D_HA, req, "going to resend");
2894         spin_lock(&req->rq_lock);
2895
2896         /*
2897          * Request got reply but linked to the import list still.
2898          * Let ptlrpc_check_set() process it.
2899          */
2900         if (ptlrpc_client_replied(req)) {
2901                 spin_unlock(&req->rq_lock);
2902                 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2903                 return;
2904         }
2905
2906         req->rq_status = -EAGAIN;
2907
2908         req->rq_resend = 1;
2909         req->rq_net_err = 0;
2910         req->rq_timedout = 0;
2911
2912         ptlrpc_client_wake_req(req);
2913         spin_unlock(&req->rq_lock);
2914 }
2915
2916 /* XXX: this function and rq_status are currently unused */
2917 void ptlrpc_restart_req(struct ptlrpc_request *req)
2918 {
2919         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2920         req->rq_status = -ERESTARTSYS;
2921
2922         spin_lock(&req->rq_lock);
2923         req->rq_restart = 1;
2924         req->rq_timedout = 0;
2925         ptlrpc_client_wake_req(req);
2926         spin_unlock(&req->rq_lock);
2927 }
2928
2929 /**
2930  * Grab additional reference on a request \a req
2931  */
2932 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2933 {
2934         ENTRY;
2935         atomic_inc(&req->rq_refcount);
2936         RETURN(req);
2937 }
2938 EXPORT_SYMBOL(ptlrpc_request_addref);
2939
2940 /**
2941  * Add a request to import replay_list.
2942  * Must be called under imp_lock
2943  */
2944 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2945                                       struct obd_import *imp)
2946 {
2947         struct list_head *tmp;
2948
2949         assert_spin_locked(&imp->imp_lock);
2950
2951         if (req->rq_transno == 0) {
2952                 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2953                 LBUG();
2954         }
2955
2956         /*
2957          * clear this for new requests that were resent as well
2958          * as resent replayed requests.
2959          */
2960         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2961
2962         /* don't re-add requests that have been replayed */
2963         if (!list_empty(&req->rq_replay_list))
2964                 return;
2965
2966         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2967
2968         spin_lock(&req->rq_lock);
2969         req->rq_resend = 0;
2970         spin_unlock(&req->rq_lock);
2971
2972         LASSERT(imp->imp_replayable);
2973         /* Balanced in ptlrpc_free_committed, usually. */
2974         ptlrpc_request_addref(req);
2975         list_for_each_prev(tmp, &imp->imp_replay_list) {
2976                 struct ptlrpc_request *iter = list_entry(tmp,
2977                                                          struct ptlrpc_request,
2978                                                          rq_replay_list);
2979
2980                 /*
2981                  * We may have duplicate transnos if we create and then
2982                  * open a file, or for closes retained if to match creating
2983                  * opens, so use req->rq_xid as a secondary key.
2984                  * (See bugs 684, 685, and 428.)
2985                  * XXX no longer needed, but all opens need transnos!
2986                  */
2987                 if (iter->rq_transno > req->rq_transno)
2988                         continue;
2989
2990                 if (iter->rq_transno == req->rq_transno) {
2991                         LASSERT(iter->rq_xid != req->rq_xid);
2992                         if (iter->rq_xid > req->rq_xid)
2993                                 continue;
2994                 }
2995
2996                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2997                 return;
2998         }
2999
3000         list_add(&req->rq_replay_list, &imp->imp_replay_list);
3001 }
3002
3003 /**
3004  * Send request and wait until it completes.
3005  * Returns request processing status.
3006  */
3007 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3008 {
3009         struct ptlrpc_request_set *set;
3010         int rc;
3011
3012         ENTRY;
3013         LASSERT(req->rq_set == NULL);
3014         LASSERT(!req->rq_receiving_reply);
3015
3016         set = ptlrpc_prep_set();
3017         if (!set) {
3018                 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3019                 RETURN(-ENOMEM);
3020         }
3021
3022         /* for distributed debugging */
3023         lustre_msg_set_status(req->rq_reqmsg, current_pid());
3024
3025         /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3026         ptlrpc_request_addref(req);
3027         ptlrpc_set_add_req(set, req);
3028         rc = ptlrpc_set_wait(NULL, set);
3029         ptlrpc_set_destroy(set);
3030
3031         RETURN(rc);
3032 }
3033 EXPORT_SYMBOL(ptlrpc_queue_wait);
3034
3035 /**
3036  * Callback used for replayed requests reply processing.
3037  * In case of successful reply calls registered request replay callback.
3038  * In case of error restart replay process.
3039  */
3040 static int ptlrpc_replay_interpret(const struct lu_env *env,
3041                                    struct ptlrpc_request *req,
3042                                    void *args, int rc)
3043 {
3044         struct ptlrpc_replay_async_args *aa = args;
3045         struct obd_import *imp = req->rq_import;
3046
3047         ENTRY;
3048         atomic_dec(&imp->imp_replay_inflight);
3049
3050         /*
3051          * Note: if it is bulk replay (MDS-MDS replay), then even if
3052          * server got the request, but bulk transfer timeout, let's
3053          * replay the bulk req again
3054          */
3055         if (!ptlrpc_client_replied(req) ||
3056             (req->rq_bulk &&
3057              lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3058                 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3059                 GOTO(out, rc = -ETIMEDOUT);
3060         }
3061
3062         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3063             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3064             lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3065                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3066
3067         /** VBR: check version failure */
3068         if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3069                 /** replay was failed due to version mismatch */
3070                 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3071                 spin_lock(&imp->imp_lock);
3072                 imp->imp_vbr_failed = 1;
3073                 spin_unlock(&imp->imp_lock);
3074                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3075         } else {
3076                 /** The transno had better not change over replay. */
3077                 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3078                          lustre_msg_get_transno(req->rq_repmsg) ||
3079                          lustre_msg_get_transno(req->rq_repmsg) == 0,
3080                          "%#llx/%#llx\n",
3081                          lustre_msg_get_transno(req->rq_reqmsg),
3082                          lustre_msg_get_transno(req->rq_repmsg));
3083         }
3084
3085         spin_lock(&imp->imp_lock);
3086         imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3087         spin_unlock(&imp->imp_lock);
3088         LASSERT(imp->imp_last_replay_transno);
3089
3090         /* transaction number shouldn't be bigger than the latest replayed */
3091         if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3092                 DEBUG_REQ(D_ERROR, req,
3093                           "Reported transno=%llu is bigger than replayed=%llu",
3094                           req->rq_transno,
3095                           lustre_msg_get_transno(req->rq_reqmsg));
3096                 GOTO(out, rc = -EINVAL);
3097         }
3098
3099         DEBUG_REQ(D_HA, req, "got reply");
3100
3101         /* let the callback do fixups, possibly including in the request */
3102         if (req->rq_replay_cb)
3103                 req->rq_replay_cb(req);
3104
3105         if (ptlrpc_client_replied(req) &&
3106             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3107                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3108                           lustre_msg_get_status(req->rq_repmsg),
3109                           aa->praa_old_status);
3110
3111                 /*
3112                  * Note: If the replay fails for MDT-MDT recovery, let's
3113                  * abort all of the following requests in the replay
3114                  * and sending list, because MDT-MDT update requests
3115                  * are dependent on each other, see LU-7039
3116                  */
3117                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3118                         struct ptlrpc_request *free_req;
3119                         struct ptlrpc_request *tmp;
3120
3121                         spin_lock(&imp->imp_lock);
3122                         list_for_each_entry_safe(free_req, tmp,
3123                                                  &imp->imp_replay_list,
3124                                                  rq_replay_list) {
3125                                 ptlrpc_free_request(free_req);
3126                         }
3127
3128                         list_for_each_entry_safe(free_req, tmp,
3129                                                  &imp->imp_committed_list,
3130                                                  rq_replay_list) {
3131                                 ptlrpc_free_request(free_req);
3132                         }
3133
3134                         list_for_each_entry_safe(free_req, tmp,
3135                                                  &imp->imp_delayed_list,
3136                                                  rq_list) {
3137                                 spin_lock(&free_req->rq_lock);
3138                                 free_req->rq_err = 1;
3139                                 free_req->rq_status = -EIO;
3140                                 ptlrpc_client_wake_req(free_req);
3141                                 spin_unlock(&free_req->rq_lock);
3142                         }
3143
3144                         list_for_each_entry_safe(free_req, tmp,
3145                                                  &imp->imp_sending_list,
3146                                                  rq_list) {
3147                                 spin_lock(&free_req->rq_lock);
3148                                 free_req->rq_err = 1;
3149                                 free_req->rq_status = -EIO;
3150                                 ptlrpc_client_wake_req(free_req);
3151                                 spin_unlock(&free_req->rq_lock);
3152                         }
3153                         spin_unlock(&imp->imp_lock);
3154                 }
3155         } else {
3156                 /* Put it back for re-replay. */
3157                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3158         }
3159
3160         /*
3161          * Errors while replay can set transno to 0, but
3162          * imp_last_replay_transno shouldn't be set to 0 anyway
3163          */
3164         if (req->rq_transno == 0)
3165                 CERROR("Transno is 0 during replay!\n");
3166
3167         /* continue with recovery */
3168         rc = ptlrpc_import_recovery_state_machine(imp);
3169  out:
3170         req->rq_send_state = aa->praa_old_state;
3171
3172         if (rc != 0)
3173                 /* this replay failed, so restart recovery */
3174                 ptlrpc_connect_import(imp);
3175
3176         RETURN(rc);
3177 }
3178
3179 /**
3180  * Prepares and queues request for replay.
3181  * Adds it to ptlrpcd queue for actual sending.
3182  * Returns 0 on success.
3183  */
3184 int ptlrpc_replay_req(struct ptlrpc_request *req)
3185 {
3186         struct ptlrpc_replay_async_args *aa;
3187
3188         ENTRY;
3189
3190         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3191
3192         aa = ptlrpc_req_async_args(aa, req);
3193         memset(aa, 0, sizeof(*aa));
3194
3195         /* Prepare request to be resent with ptlrpcd */
3196         aa->praa_old_state = req->rq_send_state;
3197         req->rq_send_state = LUSTRE_IMP_REPLAY;
3198         req->rq_phase = RQ_PHASE_NEW;
3199         req->rq_next_phase = RQ_PHASE_UNDEFINED;
3200         if (req->rq_repmsg)
3201                 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3202         req->rq_status = 0;
3203         req->rq_interpret_reply = ptlrpc_replay_interpret;
3204         /* Readjust the timeout for current conditions */
3205         ptlrpc_at_set_req_timeout(req);
3206
3207         /* Tell server net_latency to calculate how long to wait for reply. */
3208         lustre_msg_set_service_time(req->rq_reqmsg,
3209                                     ptlrpc_at_get_net_latency(req));
3210         DEBUG_REQ(D_HA, req, "REPLAY");
3211
3212         atomic_inc(&req->rq_import->imp_replay_inflight);