Whamcloud - gitweb
LU-9683 ptlrpc: fix argument misorder
[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, 2016, 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 <obd_support.h>
38 #include <obd_class.h>
39 #include <lustre_lib.h>
40 #include <lustre_ha.h>
41 #include <lustre_import.h>
42 #include <lustre_req_layout.h>
43
44 #include "ptlrpc_internal.h"
45
46 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
47         .add_kiov_frag  = ptlrpc_prep_bulk_page_pin,
48         .release_frags  = ptlrpc_release_bulk_page_pin,
49 };
50 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
51
52 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
53         .add_kiov_frag  = ptlrpc_prep_bulk_page_nopin,
54         .release_frags  = ptlrpc_release_bulk_noop,
55 };
56 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
57
58 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
59         .add_iov_frag = ptlrpc_prep_bulk_frag,
60 };
61 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
62
63 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
64 static int ptlrpcd_check_work(struct ptlrpc_request *req);
65 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
66
67 /**
68  * Initialize passed in client structure \a cl.
69  */
70 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
71                         struct ptlrpc_client *cl)
72 {
73         cl->cli_request_portal = req_portal;
74         cl->cli_reply_portal   = rep_portal;
75         cl->cli_name           = name;
76 }
77 EXPORT_SYMBOL(ptlrpc_init_client);
78
79 /**
80  * Return PortalRPC connection for remore uud \a uuid
81  */
82 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
83                                                     lnet_nid_t nid4refnet)
84 {
85         struct ptlrpc_connection *c;
86         lnet_nid_t                self;
87         struct lnet_process_id peer;
88         int                       err;
89
90         /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
91          * before accessing its values. */
92         /* coverity[uninit_use_in_call] */
93         peer.nid = nid4refnet;
94         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
95         if (err != 0) {
96                 CNETERR("cannot find peer %s!\n", uuid->uuid);
97                 return NULL;
98         }
99
100         c = ptlrpc_connection_get(peer, self, uuid);
101         if (c) {
102                 memcpy(c->c_remote_uuid.uuid,
103                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
104         }
105
106         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
107
108         return c;
109 }
110
111 /**
112  * Allocate and initialize new bulk descriptor on the sender.
113  * Returns pointer to the descriptor or NULL on error.
114  */
115 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
116                                          enum ptlrpc_bulk_op_type type,
117                                          unsigned portal,
118                                          const struct ptlrpc_bulk_frag_ops *ops)
119 {
120         struct ptlrpc_bulk_desc *desc;
121         int i;
122
123         /* ensure that only one of KIOV or IOVEC is set but not both */
124         LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
125                  ops->add_kiov_frag != NULL) ||
126                 (ptlrpc_is_bulk_desc_kvec(type) &&
127                  ops->add_iov_frag != NULL));
128
129         OBD_ALLOC_PTR(desc);
130         if (desc == NULL)
131                 return NULL;
132         if (type & PTLRPC_BULK_BUF_KIOV) {
133                 OBD_ALLOC_LARGE(GET_KIOV(desc),
134                                 nfrags * sizeof(*GET_KIOV(desc)));
135                 if (GET_KIOV(desc) == NULL)
136                         goto out;
137         } else {
138                 OBD_ALLOC_LARGE(GET_KVEC(desc),
139                                 nfrags * sizeof(*GET_KVEC(desc)));
140                 if (GET_KVEC(desc) == NULL)
141                         goto out;
142         }
143
144         spin_lock_init(&desc->bd_lock);
145         init_waitqueue_head(&desc->bd_waitq);
146         desc->bd_max_iov = nfrags;
147         desc->bd_iov_count = 0;
148         desc->bd_portal = portal;
149         desc->bd_type = type;
150         desc->bd_md_count = 0;
151         desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
152         LASSERT(max_brw > 0);
153         desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
154         /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
155          * node. Negotiated ocd_brw_size will always be <= this number. */
156         for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
157                 LNetInvalidateMDHandle(&desc->bd_mds[i]);
158
159         return desc;
160 out:
161         OBD_FREE_PTR(desc);
162         return NULL;
163 }
164
165 /**
166  * Prepare bulk descriptor for specified outgoing request \a req that
167  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
168  * the bulk to be sent. Used on client-side.
169  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
170  * error.
171  */
172 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
173                                               unsigned nfrags, unsigned max_brw,
174                                               unsigned int type,
175                                               unsigned portal,
176                                               const struct ptlrpc_bulk_frag_ops
177                                                 *ops)
178 {
179         struct obd_import *imp = req->rq_import;
180         struct ptlrpc_bulk_desc *desc;
181
182         ENTRY;
183         LASSERT(ptlrpc_is_bulk_op_passive(type));
184
185         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
186         if (desc == NULL)
187                 RETURN(NULL);
188
189         desc->bd_import_generation = req->rq_import_generation;
190         desc->bd_import = class_import_get(imp);
191         desc->bd_req = req;
192
193         desc->bd_cbid.cbid_fn  = client_bulk_callback;
194         desc->bd_cbid.cbid_arg = desc;
195
196         /* This makes req own desc, and free it when she frees herself */
197         req->rq_bulk = desc;
198
199         return desc;
200 }
201 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
202
203 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
204                              struct page *page, int pageoffset, int len,
205                              int pin)
206 {
207         lnet_kiov_t *kiov;
208
209         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
210         LASSERT(page != NULL);
211         LASSERT(pageoffset >= 0);
212         LASSERT(len > 0);
213         LASSERT(pageoffset + len <= PAGE_SIZE);
214         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
215
216         kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
217
218         desc->bd_nob += len;
219
220         if (pin)
221                 get_page(page);
222
223         kiov->kiov_page = page;
224         kiov->kiov_offset = pageoffset;
225         kiov->kiov_len = len;
226
227         desc->bd_iov_count++;
228 }
229 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
230
231 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
232                           void *frag, int len)
233 {
234         struct kvec *iovec;
235         ENTRY;
236
237         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
238         LASSERT(frag != NULL);
239         LASSERT(len > 0);
240         LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
241
242         iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
243
244         desc->bd_nob += len;
245
246         iovec->iov_base = frag;
247         iovec->iov_len = len;
248
249         desc->bd_iov_count++;
250
251         RETURN(desc->bd_nob);
252 }
253 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
254
255 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
256 {
257         ENTRY;
258
259         LASSERT(desc != NULL);
260         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
261         LASSERT(desc->bd_md_count == 0);         /* network hands off */
262         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
263         LASSERT(desc->bd_frag_ops != NULL);
264
265         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
266                 sptlrpc_enc_pool_put_pages(desc);
267
268         if (desc->bd_export)
269                 class_export_put(desc->bd_export);
270         else
271                 class_import_put(desc->bd_import);
272
273         if (desc->bd_frag_ops->release_frags != NULL)
274                 desc->bd_frag_ops->release_frags(desc);
275
276         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
277                 OBD_FREE_LARGE(GET_KIOV(desc),
278                         desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
279         else
280                 OBD_FREE_LARGE(GET_KVEC(desc),
281                         desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
282         OBD_FREE_PTR(desc);
283         EXIT;
284 }
285 EXPORT_SYMBOL(ptlrpc_free_bulk);
286
287 /**
288  * Set server timelimit for this req, i.e. how long are we willing to wait
289  * for reply before timing out this request.
290  */
291 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
292 {
293         __u32 serv_est;
294         int idx;
295         struct imp_at *at;
296
297         LASSERT(req->rq_import);
298
299         if (AT_OFF) {
300                 /* non-AT settings */
301                 /**
302                  * \a imp_server_timeout means this is reverse import and
303                  * we send (currently only) ASTs to the client and cannot afford
304                  * to wait too long for the reply, otherwise the other client
305                  * (because of which we are sending this request) would
306                  * timeout waiting for us
307                  */
308                 req->rq_timeout = req->rq_import->imp_server_timeout ?
309                                   obd_timeout / 2 : obd_timeout;
310         } else {
311                 at = &req->rq_import->imp_at;
312                 idx = import_at_get_index(req->rq_import,
313                                           req->rq_request_portal);
314                 serv_est = at_get(&at->iat_service_estimate[idx]);
315                 req->rq_timeout = at_est2timeout(serv_est);
316         }
317         /* We could get even fancier here, using history to predict increased
318            loading... */
319
320         /* Let the server know what this RPC timeout is by putting it in the
321            reqmsg*/
322         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
323 }
324 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
325
326 /* Adjust max service estimate based on server value */
327 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
328                                   unsigned int serv_est)
329 {
330         int idx;
331         unsigned int oldse;
332         struct imp_at *at;
333
334         LASSERT(req->rq_import);
335         at = &req->rq_import->imp_at;
336
337         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
338         /* max service estimates are tracked on the server side,
339            so just keep minimal history here */
340         oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
341         if (oldse != 0)
342                 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
343                        "has changed from %d to %d\n",
344                        req->rq_import->imp_obd->obd_name,req->rq_request_portal,
345                        oldse, at_get(&at->iat_service_estimate[idx]));
346 }
347
348 /* Expected network latency per remote node (secs) */
349 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
350 {
351         return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
352 }
353
354 /* Adjust expected network latency */
355 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
356                                unsigned int service_time)
357 {
358         unsigned int nl, oldnl;
359         struct imp_at *at;
360         time64_t now = ktime_get_real_seconds();
361
362         LASSERT(req->rq_import);
363
364         if (service_time > now - req->rq_sent + 3) {
365                 /* bz16408, however, this can also happen if early reply
366                  * is lost and client RPC is expired and resent, early reply
367                  * or reply of original RPC can still be fit in reply buffer
368                  * of resent RPC, now client is measuring time from the
369                  * resent time, but server sent back service time of original
370                  * RPC.
371                  */
372                 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
373                        D_ADAPTTO : D_WARNING,
374                        "Reported service time %u > total measured time %lld\n",
375                        service_time, now - req->rq_sent);
376                 return;
377         }
378
379         /* Network latency is total time less server processing time */
380         nl = max_t(int, now - req->rq_sent -
381                         service_time, 0) + 1; /* st rounding */
382         at = &req->rq_import->imp_at;
383
384         oldnl = at_measured(&at->iat_net_latency, nl);
385         if (oldnl != 0)
386                 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
387                        "has changed from %d to %d\n",
388                        req->rq_import->imp_obd->obd_name,
389                        obd_uuid2str(
390                                &req->rq_import->imp_connection->c_remote_uuid),
391                        oldnl, at_get(&at->iat_net_latency));
392 }
393
394 static int unpack_reply(struct ptlrpc_request *req)
395 {
396         int rc;
397
398         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
399                 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
400                 if (rc) {
401                         DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
402                         return(-EPROTO);
403                 }
404         }
405
406         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
407         if (rc) {
408                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
409                 return(-EPROTO);
410         }
411         return 0;
412 }
413
414 /**
415  * Handle an early reply message, called with the rq_lock held.
416  * If anything goes wrong just ignore it - same as if it never happened
417  */
418 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
419 __must_hold(&req->rq_lock)
420 {
421         struct ptlrpc_request *early_req;
422         time64_t olddl;
423         int rc;
424
425         ENTRY;
426         req->rq_early = 0;
427         spin_unlock(&req->rq_lock);
428
429         rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
430         if (rc) {
431                 spin_lock(&req->rq_lock);
432                 RETURN(rc);
433         }
434
435         rc = unpack_reply(early_req);
436         if (rc != 0) {
437                 sptlrpc_cli_finish_early_reply(early_req);
438                 spin_lock(&req->rq_lock);
439                 RETURN(rc);
440         }
441
442         /* Use new timeout value just to adjust the local value for this
443          * request, don't include it into at_history. It is unclear yet why
444          * service time increased and should it be counted or skipped, e.g.
445          * that can be recovery case or some error or server, the real reply
446          * will add all new data if it is worth to add. */
447         req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
448         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
449
450         /* Network latency can be adjusted, it is pure network delays */
451         ptlrpc_at_adj_net_latency(req,
452                         lustre_msg_get_service_time(early_req->rq_repmsg));
453
454         sptlrpc_cli_finish_early_reply(early_req);
455
456         spin_lock(&req->rq_lock);
457         olddl = req->rq_deadline;
458         /* server assumes it now has rq_timeout from when the request
459          * arrived, so the client should give it at least that long.
460          * since we don't know the arrival time we'll use the original
461          * sent time */
462         req->rq_deadline = req->rq_sent + req->rq_timeout +
463                            ptlrpc_at_get_net_latency(req);
464
465         DEBUG_REQ(D_ADAPTTO, req,
466                   "Early reply #%d, new deadline in %llds (%llds)",
467                   req->rq_early_count,
468                   req->rq_deadline - ktime_get_real_seconds(),
469                   req->rq_deadline - olddl);
470
471         RETURN(rc);
472 }
473
474 static struct kmem_cache *request_cache;
475
476 int ptlrpc_request_cache_init(void)
477 {
478         request_cache = kmem_cache_create("ptlrpc_cache",
479                                           sizeof(struct ptlrpc_request),
480                                           0, SLAB_HWCACHE_ALIGN, NULL);
481         return request_cache == NULL ? -ENOMEM : 0;
482 }
483
484 void ptlrpc_request_cache_fini(void)
485 {
486         kmem_cache_destroy(request_cache);
487 }
488
489 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
490 {
491         struct ptlrpc_request *req;
492
493         OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
494         return req;
495 }
496
497 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
498 {
499         OBD_SLAB_FREE_PTR(req, request_cache);
500 }
501
502 /**
503  * Wind down request pool \a pool.
504  * Frees all requests from the pool too
505  */
506 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
507 {
508         struct list_head *l, *tmp;
509         struct ptlrpc_request *req;
510
511         LASSERT(pool != NULL);
512
513         spin_lock(&pool->prp_lock);
514         list_for_each_safe(l, tmp, &pool->prp_req_list) {
515                 req = list_entry(l, struct ptlrpc_request, rq_list);
516                 list_del(&req->rq_list);
517                 LASSERT(req->rq_reqbuf);
518                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
519                 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
520                 ptlrpc_request_cache_free(req);
521         }
522         spin_unlock(&pool->prp_lock);
523         OBD_FREE(pool, sizeof(*pool));
524 }
525 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
526
527 /**
528  * Allocates, initializes and adds \a num_rq requests to the pool \a pool
529  */
530 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
531 {
532         int i;
533         int size = 1;
534
535         while (size < pool->prp_rq_size)
536                 size <<= 1;
537
538         LASSERTF(list_empty(&pool->prp_req_list) ||
539                  size == pool->prp_rq_size,
540                  "Trying to change pool size with nonempty pool "
541                  "from %d to %d bytes\n", pool->prp_rq_size, size);
542
543         spin_lock(&pool->prp_lock);
544         pool->prp_rq_size = size;
545         for (i = 0; i < num_rq; i++) {
546                 struct ptlrpc_request *req;
547                 struct lustre_msg *msg;
548
549                 spin_unlock(&pool->prp_lock);
550                 req = ptlrpc_request_cache_alloc(GFP_NOFS);
551                 if (!req)
552                         return i;
553                 OBD_ALLOC_LARGE(msg, size);
554                 if (!msg) {
555                         ptlrpc_request_cache_free(req);
556                         return i;
557                 }
558                 req->rq_reqbuf = msg;
559                 req->rq_reqbuf_len = size;
560                 req->rq_pool = pool;
561                 spin_lock(&pool->prp_lock);
562                 list_add_tail(&req->rq_list, &pool->prp_req_list);
563         }
564         spin_unlock(&pool->prp_lock);
565         return num_rq;
566 }
567 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
568
569 /**
570  * Create and initialize new request pool with given attributes:
571  * \a num_rq - initial number of requests to create for the pool
572  * \a msgsize - maximum message size possible for requests in thid pool
573  * \a populate_pool - function to be called when more requests need to be added
574  *                    to the pool
575  * Returns pointer to newly created pool or NULL on error.
576  */
577 struct ptlrpc_request_pool *
578 ptlrpc_init_rq_pool(int num_rq, int msgsize,
579                     int (*populate_pool)(struct ptlrpc_request_pool *, int))
580 {
581         struct ptlrpc_request_pool *pool;
582
583         OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
584         if (!pool)
585                 return NULL;
586
587         /* Request next power of two for the allocation, because internally
588            kernel would do exactly this */
589
590         spin_lock_init(&pool->prp_lock);
591         INIT_LIST_HEAD(&pool->prp_req_list);
592         pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
593         pool->prp_populate = populate_pool;
594
595         populate_pool(pool, num_rq);
596
597         return pool;
598 }
599 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
600
601 /**
602  * Fetches one request from pool \a pool
603  */
604 static struct ptlrpc_request *
605 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
606 {
607         struct ptlrpc_request *request;
608         struct lustre_msg *reqbuf;
609
610         if (!pool)
611                 return NULL;
612
613         spin_lock(&pool->prp_lock);
614
615         /* See if we have anything in a pool, and bail out if nothing,
616          * in writeout path, where this matters, this is safe to do, because
617          * nothing is lost in this case, and when some in-flight requests
618          * complete, this code will be called again. */
619         if (unlikely(list_empty(&pool->prp_req_list))) {
620                 spin_unlock(&pool->prp_lock);
621                 return NULL;
622         }
623
624         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
625                              rq_list);
626         list_del_init(&request->rq_list);
627         spin_unlock(&pool->prp_lock);
628
629         LASSERT(request->rq_reqbuf);
630         LASSERT(request->rq_pool);
631
632         reqbuf = request->rq_reqbuf;
633         memset(request, 0, sizeof(*request));
634         request->rq_reqbuf = reqbuf;
635         request->rq_reqbuf_len = pool->prp_rq_size;
636         request->rq_pool = pool;
637
638         return request;
639 }
640
641 /**
642  * Returns freed \a request to pool.
643  */
644 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
645 {
646         struct ptlrpc_request_pool *pool = request->rq_pool;
647
648         spin_lock(&pool->prp_lock);
649         LASSERT(list_empty(&request->rq_list));
650         LASSERT(!request->rq_receiving_reply);
651         list_add_tail(&request->rq_list, &pool->prp_req_list);
652         spin_unlock(&pool->prp_lock);
653 }
654
655 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
656 {
657         struct obd_import       *imp = req->rq_import;
658         struct list_head        *tmp;
659         struct ptlrpc_request   *iter;
660
661         assert_spin_locked(&imp->imp_lock);
662         LASSERT(list_empty(&req->rq_unreplied_list));
663
664         /* unreplied list is sorted by xid in ascending order */
665         list_for_each_prev(tmp, &imp->imp_unreplied_list) {
666                 iter = list_entry(tmp, struct ptlrpc_request,
667                                   rq_unreplied_list);
668
669                 LASSERT(req->rq_xid != iter->rq_xid);
670                 if (req->rq_xid < iter->rq_xid)
671                         continue;
672                 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
673                 return;
674         }
675         list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
676 }
677
678 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
679 {
680         req->rq_xid = ptlrpc_next_xid();
681         ptlrpc_add_unreplied(req);
682 }
683
684 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
685 {
686         spin_lock(&req->rq_import->imp_lock);
687         ptlrpc_assign_next_xid_nolock(req);
688         spin_unlock(&req->rq_import->imp_lock);
689 }
690
691 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
692                              __u32 version, int opcode, char **bufs,
693                              struct ptlrpc_cli_ctx *ctx)
694 {
695         int count;
696         struct obd_import *imp;
697         __u32 *lengths;
698         int rc;
699
700         ENTRY;
701
702         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
703         imp = request->rq_import;
704         lengths = request->rq_pill.rc_area[RCL_CLIENT];
705
706         if (ctx != NULL) {
707                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
708         } else {
709                 rc = sptlrpc_req_get_ctx(request);
710                 if (rc)
711                         GOTO(out_free, rc);
712         }
713         sptlrpc_req_set_flavor(request, opcode);
714
715         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
716                                  lengths, bufs);
717         if (rc)
718                 GOTO(out_ctx, rc);
719
720         lustre_msg_add_version(request->rq_reqmsg, version);
721         request->rq_send_state = LUSTRE_IMP_FULL;
722         request->rq_type = PTL_RPC_MSG_REQUEST;
723
724         request->rq_req_cbid.cbid_fn  = request_out_callback;
725         request->rq_req_cbid.cbid_arg = request;
726
727         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
728         request->rq_reply_cbid.cbid_arg = request;
729
730         request->rq_reply_deadline = 0;
731         request->rq_bulk_deadline = 0;
732         request->rq_req_deadline = 0;
733         request->rq_phase = RQ_PHASE_NEW;
734         request->rq_next_phase = RQ_PHASE_UNDEFINED;
735
736         request->rq_request_portal = imp->imp_client->cli_request_portal;
737         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
738
739         ptlrpc_at_set_req_timeout(request);
740
741         lustre_msg_set_opc(request->rq_reqmsg, opcode);
742         ptlrpc_assign_next_xid(request);
743
744         /* Let's setup deadline for req/reply/bulk unlink for opcode. */
745         if (cfs_fail_val == opcode) {
746                 time64_t *fail_t = NULL, *fail2_t = NULL;
747
748                 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
749                         fail_t = &request->rq_bulk_deadline;
750                 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
751                         fail_t = &request->rq_reply_deadline;
752                 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
753                         fail_t = &request->rq_req_deadline;
754                 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
755                         fail_t = &request->rq_reply_deadline;
756                         fail2_t = &request->rq_bulk_deadline;
757                 }
758
759                 if (fail_t) {
760                         *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
761
762                         if (fail2_t)
763                                 *fail2_t = ktime_get_real_seconds() +
764                                            LONG_UNLINK;
765
766                         /*
767                          * The RPC is infected, let the test to change the
768                          * fail_loc
769                          */
770                         msleep(4 * MSEC_PER_SEC);
771                 }
772         }
773
774         RETURN(0);
775
776 out_ctx:
777         LASSERT(!request->rq_pool);
778         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
779 out_free:
780         class_import_put(imp);
781
782         return rc;
783
784 }
785 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
786
787 /**
788  * Pack request buffers for network transfer, performing necessary encryption
789  * steps if necessary.
790  */
791 int ptlrpc_request_pack(struct ptlrpc_request *request,
792                         __u32 version, int opcode)
793 {
794         int rc;
795         rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
796         if (rc)
797                 return rc;
798
799         /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
800          * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
801          * have to send old ptlrpc_body to keep interoprability with these
802          * clients.
803          *
804          * Only three kinds of server->client RPCs so far:
805          *  - LDLM_BL_CALLBACK
806          *  - LDLM_CP_CALLBACK
807          *  - LDLM_GL_CALLBACK
808          *
809          * XXX This should be removed whenever we drop the interoprability with
810          *     the these old clients.
811          */
812         if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
813             opcode == LDLM_GL_CALLBACK)
814                 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
815                                    sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
816
817         return rc;
818 }
819 EXPORT_SYMBOL(ptlrpc_request_pack);
820
821 /**
822  * Helper function to allocate new request on import \a imp
823  * and possibly using existing request from pool \a pool if provided.
824  * Returns allocated request structure with import field filled or
825  * NULL on error.
826  */
827 static inline
828 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
829                                               struct ptlrpc_request_pool *pool)
830 {
831         struct ptlrpc_request *request = NULL;
832
833         request = ptlrpc_request_cache_alloc(GFP_NOFS);
834
835         if (!request && pool)
836                 request = ptlrpc_prep_req_from_pool(pool);
837
838         if (request) {
839                 ptlrpc_cli_req_init(request);
840
841                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
842                 LASSERT(imp != LP_POISON);
843                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
844                         imp->imp_client);
845                 LASSERT(imp->imp_client != LP_POISON);
846
847                 request->rq_import = class_import_get(imp);
848         } else {
849                 CERROR("request allocation out of memory\n");
850         }
851
852         return request;
853 }
854
855 /**
856  * Helper function for creating a request.
857  * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
858  * buffer structures according to capsule template \a format.
859  * Returns allocated request structure pointer or NULL on error.
860  */
861 static struct ptlrpc_request *
862 ptlrpc_request_alloc_internal(struct obd_import *imp,
863                               struct ptlrpc_request_pool * pool,
864                               const struct req_format *format)
865 {
866         struct ptlrpc_request *request;
867
868         request = __ptlrpc_request_alloc(imp, pool);
869         if (request == NULL)
870                 return NULL;
871
872         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
873         req_capsule_set(&request->rq_pill, format);
874         return request;
875 }
876
877 /**
878  * Allocate new request structure for import \a imp and initialize its
879  * buffer structure according to capsule template \a format.
880  */
881 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
882                                             const struct req_format *format)
883 {
884         return ptlrpc_request_alloc_internal(imp, NULL, format);
885 }
886 EXPORT_SYMBOL(ptlrpc_request_alloc);
887
888 /**
889  * Allocate new request structure for import \a imp from pool \a pool and
890  * initialize its buffer structure according to capsule template \a format.
891  */
892 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
893                                             struct ptlrpc_request_pool * pool,
894                                             const struct req_format *format)
895 {
896         return ptlrpc_request_alloc_internal(imp, pool, format);
897 }
898 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
899
900 /**
901  * For requests not from pool, free memory of the request structure.
902  * For requests obtained from a pool earlier, return request back to pool.
903  */
904 void ptlrpc_request_free(struct ptlrpc_request *request)
905 {
906         if (request->rq_pool)
907                 __ptlrpc_free_req_to_pool(request);
908         else
909                 ptlrpc_request_cache_free(request);
910 }
911 EXPORT_SYMBOL(ptlrpc_request_free);
912
913 /**
914  * Allocate new request for operatione \a opcode and immediatelly pack it for
915  * network transfer.
916  * Only used for simple requests like OBD_PING where the only important
917  * part of the request is operation itself.
918  * Returns allocated request or NULL on error.
919  */
920 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
921                                                 const struct req_format *format,
922                                                 __u32 version, int opcode)
923 {
924         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
925         int                    rc;
926
927         if (req) {
928                 rc = ptlrpc_request_pack(req, version, opcode);
929                 if (rc) {
930                         ptlrpc_request_free(req);
931                         req = NULL;
932                 }
933         }
934         return req;
935 }
936 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
937
938 /**
939  * Allocate and initialize new request set structure on the current CPT.
940  * Returns a pointer to the newly allocated set structure or NULL on error.
941  */
942 struct ptlrpc_request_set *ptlrpc_prep_set(void)
943 {
944         struct ptlrpc_request_set       *set;
945         int                             cpt;
946
947         ENTRY;
948         cpt = cfs_cpt_current(cfs_cpt_table, 0);
949         OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
950         if (!set)
951                 RETURN(NULL);
952         atomic_set(&set->set_refcount, 1);
953         INIT_LIST_HEAD(&set->set_requests);
954         init_waitqueue_head(&set->set_waitq);
955         atomic_set(&set->set_new_count, 0);
956         atomic_set(&set->set_remaining, 0);
957         spin_lock_init(&set->set_new_req_lock);
958         INIT_LIST_HEAD(&set->set_new_requests);
959         INIT_LIST_HEAD(&set->set_cblist);
960         set->set_max_inflight = UINT_MAX;
961         set->set_producer     = NULL;
962         set->set_producer_arg = NULL;
963         set->set_rc           = 0;
964
965         RETURN(set);
966 }
967 EXPORT_SYMBOL(ptlrpc_prep_set);
968
969 /**
970  * Allocate and initialize new request set structure with flow control
971  * extension. This extension allows to control the number of requests in-flight
972  * for the whole set. A callback function to generate requests must be provided
973  * and the request set will keep the number of requests sent over the wire to
974  * @max_inflight.
975  * Returns a pointer to the newly allocated set structure or NULL on error.
976  */
977 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
978                                              void *arg)
979
980 {
981         struct ptlrpc_request_set *set;
982
983         set = ptlrpc_prep_set();
984         if (!set)
985                 RETURN(NULL);
986
987         set->set_max_inflight  = max;
988         set->set_producer      = func;
989         set->set_producer_arg  = arg;
990
991         RETURN(set);
992 }
993
994 /**
995  * Wind down and free request set structure previously allocated with
996  * ptlrpc_prep_set.
997  * Ensures that all requests on the set have completed and removes
998  * all requests from the request list in a set.
999  * If any unsent request happen to be on the list, pretends that they got
1000  * an error in flight and calls their completion handler.
1001  */
1002 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1003 {
1004         struct list_head        *tmp;
1005         struct list_head        *next;
1006         int                      expected_phase;
1007         int                      n = 0;
1008         ENTRY;
1009
1010         /* Requests on the set should either all be completed, or all be new */
1011         expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1012                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1013         list_for_each(tmp, &set->set_requests) {
1014                 struct ptlrpc_request *req =
1015                         list_entry(tmp, struct ptlrpc_request,
1016                                    rq_set_chain);
1017
1018                 LASSERT(req->rq_phase == expected_phase);
1019                 n++;
1020         }
1021
1022         LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1023                  atomic_read(&set->set_remaining) == n, "%d / %d\n",
1024                  atomic_read(&set->set_remaining), n);
1025
1026         list_for_each_safe(tmp, next, &set->set_requests) {
1027                 struct ptlrpc_request *req =
1028                         list_entry(tmp, struct ptlrpc_request,
1029                                    rq_set_chain);
1030                 list_del_init(&req->rq_set_chain);
1031
1032                 LASSERT(req->rq_phase == expected_phase);
1033
1034                 if (req->rq_phase == RQ_PHASE_NEW) {
1035                         ptlrpc_req_interpret(NULL, req, -EBADR);
1036                         atomic_dec(&set->set_remaining);
1037                 }
1038
1039                 spin_lock(&req->rq_lock);
1040                 req->rq_set = NULL;
1041                 req->rq_invalid_rqset = 0;
1042                 spin_unlock(&req->rq_lock);
1043
1044                 ptlrpc_req_finished (req);
1045         }
1046
1047         LASSERT(atomic_read(&set->set_remaining) == 0);
1048
1049         ptlrpc_reqset_put(set);
1050         EXIT;
1051 }
1052 EXPORT_SYMBOL(ptlrpc_set_destroy);
1053
1054 /**
1055  * Add a callback function \a fn to the set.
1056  * This function would be called when all requests on this set are completed.
1057  * The function will be passed \a data argument.
1058  */
1059 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1060                       set_interpreter_func fn, void *data)
1061 {
1062         struct ptlrpc_set_cbdata *cbdata;
1063
1064         OBD_ALLOC_PTR(cbdata);
1065         if (cbdata == NULL)
1066                 RETURN(-ENOMEM);
1067
1068         cbdata->psc_interpret = fn;
1069         cbdata->psc_data = data;
1070         list_add_tail(&cbdata->psc_item, &set->set_cblist);
1071
1072         RETURN(0);
1073 }
1074
1075 /**
1076  * Add a new request to the general purpose request set.
1077  * Assumes request reference from the caller.
1078  */
1079 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1080                         struct ptlrpc_request *req)
1081 {
1082         LASSERT(list_empty(&req->rq_set_chain));
1083
1084         if (req->rq_allow_intr)
1085                 set->set_allow_intr = 1;
1086
1087         /* The set takes over the caller's request reference */
1088         list_add_tail(&req->rq_set_chain, &set->set_requests);
1089         req->rq_set = set;
1090         atomic_inc(&set->set_remaining);
1091         req->rq_queued_time = cfs_time_current();
1092
1093         if (req->rq_reqmsg != NULL)
1094                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1095
1096         if (set->set_producer != NULL)
1097                 /* If the request set has a producer callback, the RPC must be
1098                  * sent straight away */
1099                 ptlrpc_send_new_req(req);
1100 }
1101 EXPORT_SYMBOL(ptlrpc_set_add_req);
1102
1103 /**
1104  * Add a request to a request with dedicated server thread
1105  * and wake the thread to make any necessary processing.
1106  * Currently only used for ptlrpcd.
1107  */
1108 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1109                            struct ptlrpc_request *req)
1110 {
1111         struct ptlrpc_request_set *set = pc->pc_set;
1112         int count, i;
1113
1114         LASSERT(req->rq_set == NULL);
1115         LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1116
1117         spin_lock(&set->set_new_req_lock);
1118         /*
1119          * The set takes over the caller's request reference.
1120          */
1121         req->rq_set = set;
1122         req->rq_queued_time = cfs_time_current();
1123         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1124         count = atomic_inc_return(&set->set_new_count);
1125         spin_unlock(&set->set_new_req_lock);
1126
1127         /* Only need to call wakeup once for the first entry. */
1128         if (count == 1) {
1129                 wake_up(&set->set_waitq);
1130
1131                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1132                  *      guarantee the async RPC can be processed ASAP, we have
1133                  *      no other better choice. It maybe fixed in future. */
1134                 for (i = 0; i < pc->pc_npartners; i++)
1135                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1136         }
1137 }
1138
1139 /**
1140  * Based on the current state of the import, determine if the request
1141  * can be sent, is an error, or should be delayed.
1142  *
1143  * Returns true if this request should be delayed. If false, and
1144  * *status is set, then the request can not be sent and *status is the
1145  * error code.  If false and status is 0, then request can be sent.
1146  *
1147  * The imp->imp_lock must be held.
1148  */
1149 static int ptlrpc_import_delay_req(struct obd_import *imp,
1150                                    struct ptlrpc_request *req, int *status)
1151 {
1152         int delay = 0;
1153         ENTRY;
1154
1155         LASSERT (status != NULL);
1156         *status = 0;
1157
1158         if (req->rq_ctx_init || req->rq_ctx_fini) {
1159                 /* always allow ctx init/fini rpc go through */
1160         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1161                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1162                 *status = -EIO;
1163         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1164                 /* pings may safely race with umount */
1165                 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1166                           D_HA : D_ERROR, req, "IMP_CLOSED ");
1167                 *status = -EIO;
1168         } else if (ptlrpc_send_limit_expired(req)) {
1169                 /* probably doesn't need to be a D_ERROR after initial testing*/
1170                 DEBUG_REQ(D_HA, req, "send limit expired ");
1171                 *status = -ETIMEDOUT;
1172         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1173                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
1174                 /* allow CONNECT even if import is invalid */ ;
1175                 if (atomic_read(&imp->imp_inval_count) != 0) {
1176                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1177                         *status = -EIO;
1178                 }
1179         } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1180                 if (!imp->imp_deactive)
1181                         DEBUG_REQ(D_NET, req, "IMP_INVALID");
1182                 *status = -ESHUTDOWN; /* bz 12940 */
1183         } else if (req->rq_import_generation != imp->imp_generation) {
1184                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1185                 *status = -EIO;
1186         } else if (req->rq_send_state != imp->imp_state) {
1187                 /* invalidate in progress - any requests should be drop */
1188                 if (atomic_read(&imp->imp_inval_count) != 0) {
1189                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1190                         *status = -EIO;
1191                 } else if (req->rq_no_delay) {
1192                         *status = -EWOULDBLOCK;
1193                 } else if (req->rq_allow_replay &&
1194                           (imp->imp_state == LUSTRE_IMP_REPLAY ||
1195                            imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1196                            imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1197                            imp->imp_state == LUSTRE_IMP_RECOVER)) {
1198                         DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1199                 } else {
1200                         delay = 1;
1201                 }
1202         }
1203
1204         RETURN(delay);
1205 }
1206
1207 /**
1208  * Decide if the error message should be printed to the console or not.
1209  * Makes its decision based on request type, status, and failure frequency.
1210  *
1211  * \param[in] req  request that failed and may need a console message
1212  *
1213  * \retval false if no message should be printed
1214  * \retval true  if console message should be printed
1215  */
1216 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1217 {
1218         __u32 opc;
1219
1220         LASSERT(req->rq_reqmsg != NULL);
1221         opc = lustre_msg_get_opc(req->rq_reqmsg);
1222
1223         /* Suppress particular reconnect errors which are to be expected. */
1224         if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1225                 int err;
1226
1227                 /* Suppress timed out reconnect requests */
1228                 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1229                     req->rq_timedout)
1230                         return false;
1231
1232                 /* Suppress most unavailable/again reconnect requests, but
1233                  * print occasionally so it is clear client is trying to
1234                  * connect to a server where no target is running. */
1235                 err = lustre_msg_get_status(req->rq_repmsg);
1236                 if ((err == -ENODEV || err == -EAGAIN) &&
1237                     req->rq_import->imp_conn_cnt % 30 != 20)
1238                         return false;
1239         }
1240
1241         return true;
1242 }
1243
1244 /**
1245  * Check request processing status.
1246  * Returns the status.
1247  */
1248 static int ptlrpc_check_status(struct ptlrpc_request *req)
1249 {
1250         int err;
1251         ENTRY;
1252
1253         err = lustre_msg_get_status(req->rq_repmsg);
1254         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1255                 struct obd_import *imp = req->rq_import;
1256                 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1257                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1258
1259                 /* -EAGAIN is normal when using POSIX flocks */
1260                 if (ptlrpc_console_allow(req) &&
1261                     !(opc == LDLM_ENQUEUE && err == -EAGAIN))
1262                         LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1263                                            "failed: rc = %d\n",
1264                                            imp->imp_obd->obd_name,
1265                                            ll_opcode2str(opc),
1266                                            libcfs_nid2str(nid), err);
1267                 RETURN(err < 0 ? err : -EINVAL);
1268         }
1269
1270         if (err < 0) {
1271                 DEBUG_REQ(D_INFO, req, "status is %d", err);
1272         } else if (err > 0) {
1273                 /* XXX: translate this error from net to host */
1274                 DEBUG_REQ(D_INFO, req, "status is %d", err);
1275         }
1276
1277         RETURN(err);
1278 }
1279
1280 /**
1281  * save pre-versions of objects into request for replay.
1282  * Versions are obtained from server reply.
1283  * used for VBR.
1284  */
1285 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1286 {
1287         struct lustre_msg *repmsg = req->rq_repmsg;
1288         struct lustre_msg *reqmsg = req->rq_reqmsg;
1289         __u64 *versions = lustre_msg_get_versions(repmsg);
1290         ENTRY;
1291
1292         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1293                 return;
1294
1295         LASSERT(versions);
1296         lustre_msg_set_versions(reqmsg, versions);
1297         CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1298                versions[0], versions[1]);
1299
1300         EXIT;
1301 }
1302
1303 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1304 {
1305         struct ptlrpc_request *req;
1306
1307         assert_spin_locked(&imp->imp_lock);
1308         if (list_empty(&imp->imp_unreplied_list))
1309                 return 0;
1310
1311         req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1312                          rq_unreplied_list);
1313         LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1314
1315         if (imp->imp_known_replied_xid < req->rq_xid - 1)
1316                 imp->imp_known_replied_xid = req->rq_xid - 1;
1317
1318         return req->rq_xid - 1;
1319 }
1320
1321 /**
1322  * Callback function called when client receives RPC reply for \a req.
1323  * Returns 0 on success or error code.
1324  * The return alue would be assigned to req->rq_status by the caller
1325  * as request processing status.
1326  * This function also decides if the request needs to be saved for later replay.
1327  */
1328 static int after_reply(struct ptlrpc_request *req)
1329 {
1330         struct obd_import *imp = req->rq_import;
1331         struct obd_device *obd = req->rq_import->imp_obd;
1332         ktime_t work_start;
1333         u64 committed;
1334         s64 timediff;
1335         int rc;
1336
1337         ENTRY;
1338         LASSERT(obd != NULL);
1339         /* repbuf must be unlinked */
1340         LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1341
1342         if (req->rq_reply_truncated) {
1343                 if (ptlrpc_no_resend(req)) {
1344                         DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1345                                   " expected: %d, actual size: %d",
1346                                   req->rq_nob_received, req->rq_repbuf_len);
1347                         RETURN(-EOVERFLOW);
1348                 }
1349
1350                 sptlrpc_cli_free_repbuf(req);
1351                 /* Pass the required reply buffer size (include
1352                  * space for early reply).
1353                  * NB: no need to roundup because alloc_repbuf
1354                  * will roundup it */
1355                 req->rq_replen       = req->rq_nob_received;
1356                 req->rq_nob_received = 0;
1357                 spin_lock(&req->rq_lock);
1358                 req->rq_resend       = 1;
1359                 spin_unlock(&req->rq_lock);
1360                 RETURN(0);
1361         }
1362
1363         work_start = ktime_get_real();
1364         timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1365
1366         /*
1367          * NB Until this point, the whole of the incoming message,
1368          * including buflens, status etc is in the sender's byte order.
1369          */
1370         rc = sptlrpc_cli_unwrap_reply(req);
1371         if (rc) {
1372                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1373                 RETURN(rc);
1374         }
1375
1376         /*
1377          * Security layer unwrap might ask resend this request.
1378          */
1379         if (req->rq_resend)
1380                 RETURN(0);
1381
1382         rc = unpack_reply(req);
1383         if (rc)
1384                 RETURN(rc);
1385
1386         /* retry indefinitely on EINPROGRESS */
1387         if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1388             ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1389                 time64_t now = ktime_get_real_seconds();
1390
1391                 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1392                 spin_lock(&req->rq_lock);
1393                 req->rq_resend = 1;
1394                 spin_unlock(&req->rq_lock);
1395                 req->rq_nr_resend++;
1396
1397                 /* Readjust the timeout for current conditions */
1398                 ptlrpc_at_set_req_timeout(req);
1399                 /* delay resend to give a chance to the server to get ready.
1400                  * The delay is increased by 1s on every resend and is capped to
1401                  * the current request timeout (i.e. obd_timeout if AT is off,
1402                  * or AT service time x 125% + 5s, see at_est2timeout) */
1403                 if (req->rq_nr_resend > req->rq_timeout)
1404                         req->rq_sent = now + req->rq_timeout;
1405                 else
1406                         req->rq_sent = now + req->rq_nr_resend;
1407
1408                 /* Resend for EINPROGRESS will use a new XID */
1409                 spin_lock(&imp->imp_lock);
1410                 list_del_init(&req->rq_unreplied_list);
1411                 spin_unlock(&imp->imp_lock);
1412
1413                 RETURN(0);
1414         }
1415
1416         if (obd->obd_svc_stats != NULL) {
1417                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1418                                     timediff);
1419                 ptlrpc_lprocfs_rpc_sent(req, timediff);
1420         }
1421
1422         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1423             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1424                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1425                           lustre_msg_get_type(req->rq_repmsg));
1426                 RETURN(-EPROTO);
1427         }
1428
1429         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1430                 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1431         ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1432         ptlrpc_at_adj_net_latency(req,
1433                                   lustre_msg_get_service_time(req->rq_repmsg));
1434
1435         rc = ptlrpc_check_status(req);
1436         imp->imp_connect_error = rc;
1437
1438         if (rc) {
1439                 /*
1440                  * Either we've been evicted, or the server has failed for
1441                  * some reason. Try to reconnect, and if that fails, punt to
1442                  * the upcall.
1443                  */
1444                 if (ptlrpc_recoverable_error(rc)) {
1445                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
1446                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1447                                 RETURN(rc);
1448                         }
1449                         ptlrpc_request_handle_notconn(req);
1450                         RETURN(rc);
1451                 }
1452         } else {
1453                 /*
1454                  * Let's look if server sent slv. Do it only for RPC with
1455                  * rc == 0.
1456                  */
1457                 ldlm_cli_update_pool(req);
1458         }
1459
1460         /*
1461          * Store transno in reqmsg for replay.
1462          */
1463         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1464                 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1465                 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1466         }
1467
1468         if (imp->imp_replayable) {
1469                 spin_lock(&imp->imp_lock);
1470                 /*
1471                  * No point in adding already-committed requests to the replay
1472                  * list, we will just remove them immediately. b=9829
1473                  */
1474                 if (req->rq_transno != 0 &&
1475                     (req->rq_transno >
1476                      lustre_msg_get_last_committed(req->rq_repmsg) ||
1477                      req->rq_replay)) {
1478                         /** version recovery */
1479                         ptlrpc_save_versions(req);
1480                         ptlrpc_retain_replayable_request(req, imp);
1481                 } else if (req->rq_commit_cb != NULL &&
1482                            list_empty(&req->rq_replay_list)) {
1483                         /* NB: don't call rq_commit_cb if it's already on
1484                          * rq_replay_list, ptlrpc_free_committed() will call
1485                          * it later, see LU-3618 for details */
1486                         spin_unlock(&imp->imp_lock);
1487                         req->rq_commit_cb(req);
1488                         spin_lock(&imp->imp_lock);
1489                 }
1490
1491                 /*
1492                  * Replay-enabled imports return commit-status information.
1493                  */
1494                 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1495                 if (likely(committed > imp->imp_peer_committed_transno))
1496                         imp->imp_peer_committed_transno = committed;
1497
1498                 ptlrpc_free_committed(imp);
1499
1500                 if (!list_empty(&imp->imp_replay_list)) {
1501                         struct ptlrpc_request *last;
1502
1503                         last = list_entry(imp->imp_replay_list.prev,
1504                                           struct ptlrpc_request,
1505                                           rq_replay_list);
1506                         /*
1507                          * Requests with rq_replay stay on the list even if no
1508                          * commit is expected.
1509                          */
1510                         if (last->rq_transno > imp->imp_peer_committed_transno)
1511                                 ptlrpc_pinger_commit_expected(imp);
1512                 }
1513
1514                 spin_unlock(&imp->imp_lock);
1515         }
1516
1517         RETURN(rc);
1518 }
1519
1520 /**
1521  * Helper function to send request \a req over the network for the first time
1522  * Also adjusts request phase.
1523  * Returns 0 on success or error code.
1524  */
1525 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1526 {
1527         struct obd_import     *imp = req->rq_import;
1528         __u64                  min_xid = 0;
1529         int rc;
1530         ENTRY;
1531
1532         LASSERT(req->rq_phase == RQ_PHASE_NEW);
1533
1534         /* do not try to go further if there is not enough memory in enc_pool */
1535         if (req->rq_sent && req->rq_bulk != NULL)
1536                 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1537                     pool_is_at_full_capacity())
1538                         RETURN(-ENOMEM);
1539
1540         if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1541             (!req->rq_generation_set ||
1542              req->rq_import_generation == imp->imp_generation))
1543                 RETURN (0);
1544
1545         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1546
1547         spin_lock(&imp->imp_lock);
1548
1549         LASSERT(req->rq_xid != 0);
1550         LASSERT(!list_empty(&req->rq_unreplied_list));
1551
1552         if (!req->rq_generation_set)
1553                 req->rq_import_generation = imp->imp_generation;
1554
1555         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1556                 spin_lock(&req->rq_lock);
1557                 req->rq_waiting = 1;
1558                 spin_unlock(&req->rq_lock);
1559
1560                 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1561                           ptlrpc_import_state_name(req->rq_send_state),
1562                           ptlrpc_import_state_name(imp->imp_state));
1563                 LASSERT(list_empty(&req->rq_list));
1564                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1565                 atomic_inc(&req->rq_import->imp_inflight);
1566                 spin_unlock(&imp->imp_lock);
1567                 RETURN(0);
1568         }
1569
1570         if (rc != 0) {
1571                 spin_unlock(&imp->imp_lock);
1572                 req->rq_status = rc;
1573                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1574                 RETURN(rc);
1575         }
1576
1577         LASSERT(list_empty(&req->rq_list));
1578         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1579         atomic_inc(&req->rq_import->imp_inflight);
1580
1581         /* find the known replied XID from the unreplied list, CONNECT
1582          * and DISCONNECT requests are skipped to make the sanity check
1583          * on server side happy. see process_req_last_xid().
1584          *
1585          * For CONNECT: Because replay requests have lower XID, it'll
1586          * break the sanity check if CONNECT bump the exp_last_xid on
1587          * server.
1588          *
1589          * For DISCONNECT: Since client will abort inflight RPC before
1590          * sending DISCONNECT, DISCONNECT may carry an XID which higher
1591          * than the inflight RPC.
1592          */
1593         if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1594                 min_xid = ptlrpc_known_replied_xid(imp);
1595         spin_unlock(&imp->imp_lock);
1596
1597         lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1598
1599         lustre_msg_set_status(req->rq_reqmsg, current_pid());
1600
1601         rc = sptlrpc_req_refresh_ctx(req, -1);
1602         if (rc) {
1603                 if (req->rq_err) {
1604                         req->rq_status = rc;
1605                         RETURN(1);
1606                 } else {
1607                         spin_lock(&req->rq_lock);
1608                         req->rq_wait_ctx = 1;
1609                         spin_unlock(&req->rq_lock);
1610                         RETURN(0);
1611                 }
1612         }
1613
1614         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1615                " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1616                imp->imp_obd->obd_uuid.uuid,
1617                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1618                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1619                lustre_msg_get_opc(req->rq_reqmsg));
1620
1621         rc = ptl_send_rpc(req, 0);
1622         if (rc == -ENOMEM) {
1623                 spin_lock(&imp->imp_lock);
1624                 if (!list_empty(&req->rq_list)) {
1625                         list_del_init(&req->rq_list);
1626                         atomic_dec(&req->rq_import->imp_inflight);
1627                 }
1628                 spin_unlock(&imp->imp_lock);
1629                 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1630                 RETURN(rc);
1631         }
1632         if (rc) {
1633                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1634                 spin_lock(&req->rq_lock);
1635                 req->rq_net_err = 1;
1636                 spin_unlock(&req->rq_lock);
1637                 RETURN(rc);
1638         }
1639         RETURN(0);
1640 }
1641
1642 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1643 {
1644         int remaining, rc;
1645         ENTRY;
1646
1647         LASSERT(set->set_producer != NULL);
1648
1649         remaining = atomic_read(&set->set_remaining);
1650
1651         /* populate the ->set_requests list with requests until we
1652          * reach the maximum number of RPCs in flight for this set */
1653         while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1654                 rc = set->set_producer(set, set->set_producer_arg);
1655                 if (rc == -ENOENT) {
1656                         /* no more RPC to produce */
1657                         set->set_producer     = NULL;
1658                         set->set_producer_arg = NULL;
1659                         RETURN(0);
1660                 }
1661         }
1662
1663         RETURN((atomic_read(&set->set_remaining) - remaining));
1664 }
1665
1666 /**
1667  * this sends any unsent RPCs in \a set and returns 1 if all are sent
1668  * and no more replies are expected.
1669  * (it is possible to get less replies than requests sent e.g. due to timed out
1670  * requests or requests that we had trouble to send out)
1671  *
1672  * NOTE: This function contains a potential schedule point (cond_resched()).
1673  */
1674 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1675 {
1676         struct list_head *tmp, *next;
1677         struct list_head  comp_reqs;
1678         int force_timer_recalc = 0;
1679         ENTRY;
1680
1681         if (atomic_read(&set->set_remaining) == 0)
1682                 RETURN(1);
1683
1684         INIT_LIST_HEAD(&comp_reqs);
1685         list_for_each_safe(tmp, next, &set->set_requests) {
1686                 struct ptlrpc_request *req =
1687                         list_entry(tmp, struct ptlrpc_request,
1688                                    rq_set_chain);
1689                 struct obd_import *imp = req->rq_import;
1690                 int unregistered = 0;
1691                 int async = 1;
1692                 int rc = 0;
1693
1694                 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1695                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1696                         continue;
1697                 }
1698
1699                 /* This schedule point is mainly for the ptlrpcd caller of this
1700                  * function.  Most ptlrpc sets are not long-lived and unbounded
1701                  * in length, but at the least the set used by the ptlrpcd is.
1702                  * Since the processing time is unbounded, we need to insert an
1703                  * explicit schedule point to make the thread well-behaved.
1704                  */
1705                 cond_resched();
1706
1707                 /* If the caller requires to allow to be interpreted by force
1708                  * and it has really been interpreted, then move the request
1709                  * to RQ_PHASE_INTERPRET phase in spite of what the current
1710                  * phase is. */
1711                 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1712                         req->rq_status = -EINTR;
1713                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1714
1715                         /* Since it is interpreted and we have to wait for
1716                          * the reply to be unlinked, then use sync mode. */
1717                         async = 0;
1718
1719                         GOTO(interpret, req->rq_status);
1720                 }
1721
1722                 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1723                         force_timer_recalc = 1;
1724
1725                 /* delayed send - skip */
1726                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1727                         continue;
1728
1729                 /* delayed resend - skip */
1730                 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1731                     req->rq_sent > ktime_get_real_seconds())
1732                         continue;
1733
1734                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1735                       req->rq_phase == RQ_PHASE_BULK ||
1736                       req->rq_phase == RQ_PHASE_INTERPRET ||
1737                       req->rq_phase == RQ_PHASE_UNREG_RPC ||
1738                       req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1739                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1740                         LBUG();
1741                 }
1742
1743                 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1744                     req->rq_phase == RQ_PHASE_UNREG_BULK) {
1745                         LASSERT(req->rq_next_phase != req->rq_phase);
1746                         LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1747
1748                         if (req->rq_req_deadline &&
1749                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1750                                 req->rq_req_deadline = 0;
1751                         if (req->rq_reply_deadline &&
1752                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1753                                 req->rq_reply_deadline = 0;
1754                         if (req->rq_bulk_deadline &&
1755                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1756                                 req->rq_bulk_deadline = 0;
1757
1758                         /*
1759                          * Skip processing until reply is unlinked. We
1760                          * can't return to pool before that and we can't
1761                          * call interpret before that. We need to make
1762                          * sure that all rdma transfers finished and will
1763                          * not corrupt any data.
1764                          */
1765                         if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1766                             ptlrpc_client_recv_or_unlink(req))
1767                                 continue;
1768                         if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1769                             ptlrpc_client_bulk_active(req))
1770                                 continue;
1771
1772                         /*
1773                          * Turn fail_loc off to prevent it from looping
1774                          * forever.
1775                          */
1776                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1777                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1778                                                      OBD_FAIL_ONCE);
1779                         }
1780                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1781                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1782                                                      OBD_FAIL_ONCE);
1783                         }
1784
1785                         /*
1786                          * Move to next phase if reply was successfully
1787                          * unlinked.
1788                          */
1789                         ptlrpc_rqphase_move(req, req->rq_next_phase);
1790                 }
1791
1792                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1793                         GOTO(interpret, req->rq_status);
1794
1795                 /*
1796                  * Note that this also will start async reply unlink.
1797                  */
1798                 if (req->rq_net_err && !req->rq_timedout) {
1799                         ptlrpc_expire_one_request(req, 1);
1800
1801                         /*
1802                          * Check if we still need to wait for unlink.
1803                          */
1804                         if (ptlrpc_client_recv_or_unlink(req) ||
1805                             ptlrpc_client_bulk_active(req))
1806                                 continue;
1807                         /* If there is no need to resend, fail it now. */
1808                         if (req->rq_no_resend) {
1809                                 if (req->rq_status == 0)
1810                                         req->rq_status = -EIO;
1811                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1812                                 GOTO(interpret, req->rq_status);
1813                         } else {
1814                                 continue;
1815                         }
1816                 }
1817
1818                 if (req->rq_err) {
1819                         spin_lock(&req->rq_lock);
1820                         req->rq_replied = 0;
1821                         spin_unlock(&req->rq_lock);
1822                         if (req->rq_status == 0)
1823                                 req->rq_status = -EIO;
1824                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1825                         GOTO(interpret, req->rq_status);
1826                 }
1827
1828                 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1829                  * so it sets rq_intr regardless of individual rpc
1830                  * timeouts. The synchronous IO waiting path sets
1831                  * rq_intr irrespective of whether ptlrpcd
1832                  * has seen a timeout.  Our policy is to only interpret
1833                  * interrupted rpcs after they have timed out, so we
1834                  * need to enforce that here.
1835                  */
1836
1837                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1838                                      req->rq_wait_ctx)) {
1839                         req->rq_status = -EINTR;
1840                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1841                         GOTO(interpret, req->rq_status);
1842                 }
1843
1844                 if (req->rq_phase == RQ_PHASE_RPC) {
1845                         if (req->rq_timedout || req->rq_resend ||
1846                             req->rq_waiting || req->rq_wait_ctx) {
1847                                 int status;
1848
1849                                 if (!ptlrpc_unregister_reply(req, 1)) {
1850                                         ptlrpc_unregister_bulk(req, 1);
1851                                         continue;
1852                                 }
1853
1854                                 spin_lock(&imp->imp_lock);
1855                                 if (ptlrpc_import_delay_req(imp, req, &status)){
1856                                         /* put on delay list - only if we wait
1857                                          * recovery finished - before send */
1858                                         list_del_init(&req->rq_list);
1859                                         list_add_tail(&req->rq_list,
1860                                                           &imp->
1861                                                           imp_delayed_list);
1862                                         spin_unlock(&imp->imp_lock);
1863                                         continue;
1864                                 }
1865
1866                                 if (status != 0)  {
1867                                         req->rq_status = status;
1868                                         ptlrpc_rqphase_move(req,
1869                                                 RQ_PHASE_INTERPRET);
1870                                         spin_unlock(&imp->imp_lock);
1871                                         GOTO(interpret, req->rq_status);
1872                                 }
1873                                 if (ptlrpc_no_resend(req) &&
1874                                     !req->rq_wait_ctx) {
1875                                         req->rq_status = -ENOTCONN;
1876                                         ptlrpc_rqphase_move(req,
1877                                                             RQ_PHASE_INTERPRET);
1878                                         spin_unlock(&imp->imp_lock);
1879                                         GOTO(interpret, req->rq_status);
1880                                 }
1881
1882                                 list_del_init(&req->rq_list);
1883                                 list_add_tail(&req->rq_list,
1884                                                   &imp->imp_sending_list);
1885
1886                                 spin_unlock(&imp->imp_lock);
1887
1888                                 spin_lock(&req->rq_lock);
1889                                 req->rq_waiting = 0;
1890                                 spin_unlock(&req->rq_lock);
1891
1892                                 if (req->rq_timedout || req->rq_resend) {
1893                                         /* This is re-sending anyways,
1894                                          * let's mark req as resend. */
1895                                         spin_lock(&req->rq_lock);
1896                                         req->rq_resend = 1;
1897                                         spin_unlock(&req->rq_lock);
1898
1899                                         if (req->rq_bulk != NULL &&
1900                                             !ptlrpc_unregister_bulk(req, 1))
1901                                                 continue;
1902                                 }
1903                                 /*
1904                                  * rq_wait_ctx is only touched by ptlrpcd,
1905                                  * so no lock is needed here.
1906                                  */
1907                                 status = sptlrpc_req_refresh_ctx(req, -1);
1908                                 if (status) {
1909                                         if (req->rq_err) {
1910                                                 req->rq_status = status;
1911                                                 spin_lock(&req->rq_lock);
1912                                                 req->rq_wait_ctx = 0;
1913                                                 spin_unlock(&req->rq_lock);
1914                                                 force_timer_recalc = 1;
1915                                         } else {
1916                                                 spin_lock(&req->rq_lock);
1917                                                 req->rq_wait_ctx = 1;
1918                                                 spin_unlock(&req->rq_lock);
1919                                         }
1920
1921                                         continue;
1922                                 } else {
1923                                         spin_lock(&req->rq_lock);
1924                                         req->rq_wait_ctx = 0;
1925                                         spin_unlock(&req->rq_lock);
1926                                 }
1927
1928                                 rc = ptl_send_rpc(req, 0);
1929                                 if (rc == -ENOMEM) {
1930                                         spin_lock(&imp->imp_lock);
1931                                         if (!list_empty(&req->rq_list))
1932                                                 list_del_init(&req->rq_list);
1933                                         spin_unlock(&imp->imp_lock);
1934                                         ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1935                                         continue;
1936                                 }
1937                                 if (rc) {
1938                                         DEBUG_REQ(D_HA, req,
1939                                                   "send failed: rc = %d", rc);
1940                                         force_timer_recalc = 1;
1941                                         spin_lock(&req->rq_lock);
1942                                         req->rq_net_err = 1;
1943                                         spin_unlock(&req->rq_lock);
1944                                         continue;
1945                                 }
1946                                 /* need to reset the timeout */
1947                                 force_timer_recalc = 1;
1948                         }
1949
1950                         spin_lock(&req->rq_lock);
1951
1952                         if (ptlrpc_client_early(req)) {
1953                                 ptlrpc_at_recv_early_reply(req);
1954                                 spin_unlock(&req->rq_lock);
1955                                 continue;
1956                         }
1957
1958                         /* Still waiting for a reply? */
1959                         if (ptlrpc_client_recv(req)) {
1960                                 spin_unlock(&req->rq_lock);
1961                                 continue;
1962                         }
1963
1964                         /* Did we actually receive a reply? */
1965                         if (!ptlrpc_client_replied(req)) {
1966                                 spin_unlock(&req->rq_lock);
1967                                 continue;
1968                         }
1969
1970                         spin_unlock(&req->rq_lock);
1971
1972                         /* unlink from net because we are going to
1973                          * swab in-place of reply buffer */
1974                         unregistered = ptlrpc_unregister_reply(req, 1);
1975                         if (!unregistered)
1976                                 continue;
1977
1978                         req->rq_status = after_reply(req);
1979                         if (req->rq_resend)
1980                                 continue;
1981
1982                         /* If there is no bulk associated with this request,
1983                          * then we're done and should let the interpreter
1984                          * process the reply. Similarly if the RPC returned
1985                          * an error, and therefore the bulk will never arrive.
1986                          */
1987                         if (req->rq_bulk == NULL || req->rq_status < 0) {
1988                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1989                                 GOTO(interpret, req->rq_status);
1990                         }
1991
1992                         ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1993                 }
1994
1995                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1996                 if (ptlrpc_client_bulk_active(req))
1997                         continue;
1998
1999                 if (req->rq_bulk->bd_failure) {
2000                         /* The RPC reply arrived OK, but the bulk screwed
2001                          * up!  Dead weird since the server told us the RPC
2002                          * was good after getting the REPLY for her GET or
2003                          * the ACK for her PUT. */
2004                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2005                         req->rq_status = -EIO;
2006                 }
2007
2008                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2009
2010         interpret:
2011                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2012
2013                 /* This moves to "unregistering" phase we need to wait for
2014                  * reply unlink. */
2015                 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2016                         /* start async bulk unlink too */
2017                         ptlrpc_unregister_bulk(req, 1);
2018                         continue;
2019                 }
2020
2021                 if (!ptlrpc_unregister_bulk(req, async))
2022                         continue;
2023
2024                 /* When calling interpret receiving already should be
2025                  * finished. */
2026                 LASSERT(!req->rq_receiving_reply);
2027
2028                 ptlrpc_req_interpret(env, req, req->rq_status);
2029
2030                 if (ptlrpcd_check_work(req)) {
2031                         atomic_dec(&set->set_remaining);
2032                         continue;
2033                 }
2034                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2035
2036                 if (req->rq_reqmsg != NULL)
2037                         CDEBUG(D_RPCTRACE,
2038                                "Completed RPC pname:cluuid:pid:xid:nid:"
2039                                "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
2040                                imp->imp_obd->obd_uuid.uuid,
2041                                lustre_msg_get_status(req->rq_reqmsg),
2042                                req->rq_xid,
2043                                libcfs_nid2str(imp->imp_connection->c_peer.nid),
2044                                lustre_msg_get_opc(req->rq_reqmsg));
2045
2046                 spin_lock(&imp->imp_lock);
2047                 /* Request already may be not on sending or delaying list. This
2048                  * may happen in the case of marking it erroneous for the case
2049                  * ptlrpc_import_delay_req(req, status) find it impossible to
2050                  * allow sending this rpc and returns *status != 0. */
2051                 if (!list_empty(&req->rq_list)) {
2052                         list_del_init(&req->rq_list);
2053                         atomic_dec(&imp->imp_inflight);
2054                 }
2055                 list_del_init(&req->rq_unreplied_list);
2056                 spin_unlock(&imp->imp_lock);
2057
2058                 atomic_dec(&set->set_remaining);
2059                 wake_up_all(&imp->imp_recovery_waitq);
2060
2061                 if (set->set_producer) {
2062                         /* produce a new request if possible */
2063                         if (ptlrpc_set_producer(set) > 0)
2064                                 force_timer_recalc = 1;
2065
2066                         /* free the request that has just been completed
2067                          * in order not to pollute set->set_requests */
2068                         list_del_init(&req->rq_set_chain);
2069                         spin_lock(&req->rq_lock);
2070                         req->rq_set = NULL;
2071                         req->rq_invalid_rqset = 0;
2072                         spin_unlock(&req->rq_lock);
2073
2074                         /* record rq_status to compute the final status later */
2075                         if (req->rq_status != 0)
2076                                 set->set_rc = req->rq_status;
2077                         ptlrpc_req_finished(req);
2078                 } else {
2079                         list_move_tail(&req->rq_set_chain, &comp_reqs);
2080                 }
2081         }
2082
2083         /* move completed request at the head of list so it's easier for
2084          * caller to find them */
2085         list_splice(&comp_reqs, &set->set_requests);
2086
2087         /* If we hit an error, we want to recover promptly. */
2088         RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2089 }
2090 EXPORT_SYMBOL(ptlrpc_check_set);
2091
2092 /**
2093  * Time out request \a req. is \a async_unlink is set, that means do not wait
2094  * until LNet actually confirms network buffer unlinking.
2095  * Return 1 if we should give up further retrying attempts or 0 otherwise.
2096  */
2097 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2098 {
2099         struct obd_import *imp = req->rq_import;
2100         int rc = 0;
2101         ENTRY;
2102
2103         spin_lock(&req->rq_lock);
2104         req->rq_timedout = 1;
2105         spin_unlock(&req->rq_lock);
2106
2107         DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent %lld/real %lld]",
2108                   req->rq_net_err ? "failed due to network error" :
2109                      ((req->rq_real_sent == 0 ||
2110                        req->rq_real_sent < req->rq_sent ||
2111                        req->rq_real_sent >= req->rq_deadline) ?
2112                       "timed out for sent delay" : "timed out for slow reply"),
2113                   (s64)req->rq_sent, (s64)req->rq_real_sent);
2114
2115         if (imp != NULL && obd_debug_peer_on_timeout)
2116                 LNetDebugPeer(imp->imp_connection->c_peer);
2117
2118         ptlrpc_unregister_reply(req, async_unlink);
2119         ptlrpc_unregister_bulk(req, async_unlink);
2120
2121         if (obd_dump_on_timeout)
2122                 libcfs_debug_dumplog();
2123
2124         if (imp == NULL) {
2125                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2126                 RETURN(1);
2127         }
2128
2129         atomic_inc(&imp->imp_timeouts);
2130
2131         /* The DLM server doesn't want recovery run on its imports. */
2132         if (imp->imp_dlm_fake)
2133                 RETURN(1);
2134
2135         /* If this request is for recovery or other primordial tasks,
2136          * then error it out here. */
2137         if (req->rq_ctx_init || req->rq_ctx_fini ||
2138             req->rq_send_state != LUSTRE_IMP_FULL ||
2139             imp->imp_obd->obd_no_recov) {
2140                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2141                           ptlrpc_import_state_name(req->rq_send_state),
2142                           ptlrpc_import_state_name(imp->imp_state));
2143                 spin_lock(&req->rq_lock);
2144                 req->rq_status = -ETIMEDOUT;
2145                 req->rq_err = 1;
2146                 spin_unlock(&req->rq_lock);
2147                 RETURN(1);
2148         }
2149
2150         /* if a request can't be resent we can't wait for an answer after
2151            the timeout */
2152         if (ptlrpc_no_resend(req)) {
2153                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2154                 rc = 1;
2155         }
2156
2157         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2158
2159         RETURN(rc);
2160 }
2161
2162 /**
2163  * Time out all uncompleted requests in request set pointed by \a data
2164  * Callback used when waiting on sets with l_wait_event.
2165  * Always returns 1.
2166  */
2167 int ptlrpc_expired_set(void *data)
2168 {
2169         struct ptlrpc_request_set *set = data;
2170         struct list_head *tmp;
2171         time64_t now = ktime_get_real_seconds();
2172
2173         ENTRY;
2174         LASSERT(set != NULL);
2175
2176         /*
2177          * A timeout expired. See which reqs it applies to...
2178          */
2179         list_for_each(tmp, &set->set_requests) {
2180                 struct ptlrpc_request *req =
2181                         list_entry(tmp, struct ptlrpc_request,
2182                                    rq_set_chain);
2183
2184                 /* don't expire request waiting for context */
2185                 if (req->rq_wait_ctx)
2186                         continue;
2187
2188                 /* Request in-flight? */
2189                 if (!((req->rq_phase == RQ_PHASE_RPC &&
2190                        !req->rq_waiting && !req->rq_resend) ||
2191                       (req->rq_phase == RQ_PHASE_BULK)))
2192                         continue;
2193
2194                 if (req->rq_timedout ||     /* already dealt with */
2195                     req->rq_deadline > now) /* not expired */
2196                         continue;
2197
2198                 /* Deal with this guy. Do it asynchronously to not block
2199                  * ptlrpcd thread. */
2200                 ptlrpc_expire_one_request(req, 1);
2201         }
2202
2203         /*
2204          * When waiting for a whole set, we always break out of the
2205          * sleep so we can recalculate the timeout, or enable interrupts
2206          * if everyone's timed out.
2207          */
2208         RETURN(1);
2209 }
2210
2211 /**
2212  * Sets rq_intr flag in \a req under spinlock.
2213  */
2214 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2215 {
2216         spin_lock(&req->rq_lock);
2217         req->rq_intr = 1;
2218         spin_unlock(&req->rq_lock);
2219 }
2220 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2221
2222 /**
2223  * Interrupts (sets interrupted flag) all uncompleted requests in
2224  * a set \a data. Callback for l_wait_event for interruptible waits.
2225  */
2226 static void ptlrpc_interrupted_set(void *data)
2227 {
2228         struct ptlrpc_request_set *set = data;
2229         struct list_head *tmp;
2230
2231         LASSERT(set != NULL);
2232         CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2233
2234         list_for_each(tmp, &set->set_requests) {
2235                 struct ptlrpc_request *req =
2236                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2237
2238                 if (req->rq_intr)
2239                         continue;
2240
2241                 if (req->rq_phase != RQ_PHASE_RPC &&
2242                     req->rq_phase != RQ_PHASE_UNREG_RPC &&
2243                     !req->rq_allow_intr)
2244                         continue;
2245
2246                 ptlrpc_mark_interrupted(req);
2247         }
2248 }
2249
2250 /**
2251  * Get the smallest timeout in the set; this does NOT set a timeout.
2252  */
2253 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2254 {
2255         struct list_head *tmp;
2256         time64_t now = ktime_get_real_seconds();
2257         int timeout = 0;
2258         struct ptlrpc_request *req;
2259         time64_t deadline;
2260
2261         ENTRY;
2262         list_for_each(tmp, &set->set_requests) {
2263                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2264
2265                 /*
2266                  * Request in-flight?
2267                  */
2268                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2269                       (req->rq_phase == RQ_PHASE_BULK) ||
2270                       (req->rq_phase == RQ_PHASE_NEW)))
2271                         continue;
2272
2273                 /*
2274                  * Already timed out.
2275                  */
2276                 if (req->rq_timedout)
2277                         continue;
2278
2279                 /*
2280                  * Waiting for ctx.
2281                  */
2282                 if (req->rq_wait_ctx)
2283                         continue;
2284
2285                 if (req->rq_phase == RQ_PHASE_NEW)
2286                         deadline = req->rq_sent;
2287                 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2288                         deadline = req->rq_sent;
2289                 else
2290                         deadline = req->rq_sent + req->rq_timeout;
2291
2292                 if (deadline <= now)    /* actually expired already */
2293                         timeout = 1;    /* ASAP */
2294                 else if (timeout == 0 || timeout > deadline - now)
2295                         timeout = deadline - now;
2296         }
2297         RETURN(timeout);
2298 }
2299
2300 /**
2301  * Send all unset request from the set and then wait untill all
2302  * requests in the set complete (either get a reply, timeout, get an
2303  * error or otherwise be interrupted).
2304  * Returns 0 on success or error code otherwise.
2305  */
2306 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2307 {
2308         struct list_head            *tmp;
2309         struct ptlrpc_request *req;
2310         struct l_wait_info     lwi;
2311         int                    rc, timeout;
2312         ENTRY;
2313
2314         if (set->set_producer)
2315                 (void)ptlrpc_set_producer(set);
2316         else
2317                 list_for_each(tmp, &set->set_requests) {
2318                         req = list_entry(tmp, struct ptlrpc_request,
2319                                          rq_set_chain);
2320                         if (req->rq_phase == RQ_PHASE_NEW)
2321                                 (void)ptlrpc_send_new_req(req);
2322                 }
2323
2324         if (list_empty(&set->set_requests))
2325                 RETURN(0);
2326
2327         do {
2328                 timeout = ptlrpc_set_next_timeout(set);
2329
2330                 /* wait until all complete, interrupted, or an in-flight
2331                  * req times out */
2332                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2333                        set, timeout);
2334
2335                 if ((timeout == 0 && !signal_pending(current)) ||
2336                     set->set_allow_intr)
2337                         /* No requests are in-flight (ether timed out
2338                          * or delayed), so we can allow interrupts.
2339                          * We still want to block for a limited time,
2340                          * so we allow interrupts during the timeout. */
2341                         lwi = LWI_TIMEOUT_INTR_ALL(
2342                                         cfs_time_seconds(timeout ? timeout : 1),
2343                                         ptlrpc_expired_set,
2344                                         ptlrpc_interrupted_set, set);
2345                 else
2346                         /*
2347                          * At least one request is in flight, so no
2348                          * interrupts are allowed. Wait until all
2349                          * complete, or an in-flight req times out.
2350                          */
2351                         lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2352                                           ptlrpc_expired_set, set);
2353
2354                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2355
2356                 /* LU-769 - if we ignored the signal because it was already
2357                  * pending when we started, we need to handle it now or we risk
2358                  * it being ignored forever */
2359                 if (rc == -ETIMEDOUT &&
2360                     (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2361                     signal_pending(current)) {
2362                         sigset_t blocked_sigs =
2363                                            cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2364
2365                         /* In fact we only interrupt for the "fatal" signals
2366                          * like SIGINT or SIGKILL. We still ignore less
2367                          * important signals since ptlrpc set is not easily
2368                          * reentrant from userspace again */
2369                         if (signal_pending(current))
2370                                 ptlrpc_interrupted_set(set);
2371                         cfs_restore_sigs(blocked_sigs);
2372                 }
2373
2374                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2375
2376                 /* -EINTR => all requests have been flagged rq_intr so next
2377                  * check completes.
2378                  * -ETIMEDOUT => someone timed out.  When all reqs have
2379                  * timed out, signals are enabled allowing completion with
2380                  * EINTR.
2381                  * I don't really care if we go once more round the loop in
2382                  * the error cases -eeb. */
2383                 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2384                         list_for_each(tmp, &set->set_requests) {
2385                                 req = list_entry(tmp, struct ptlrpc_request,
2386                                                  rq_set_chain);
2387                                 spin_lock(&req->rq_lock);
2388                                 req->rq_invalid_rqset = 1;
2389                                 spin_unlock(&req->rq_lock);
2390                         }
2391                 }
2392         } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2393
2394         LASSERT(atomic_read(&set->set_remaining) == 0);
2395
2396         rc = set->set_rc; /* rq_status of already freed requests if any */
2397         list_for_each(tmp, &set->set_requests) {
2398                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2399
2400                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2401                 if (req->rq_status != 0)
2402                         rc = req->rq_status;
2403         }
2404
2405         if (set->set_interpret != NULL) {
2406                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2407                         set->set_interpret;
2408                 rc = interpreter (set, set->set_arg, rc);
2409         } else {
2410                 struct ptlrpc_set_cbdata *cbdata, *n;
2411                 int err;
2412
2413                 list_for_each_entry_safe(cbdata, n,
2414                                          &set->set_cblist, psc_item) {
2415                         list_del_init(&cbdata->psc_item);
2416                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2417                         if (err && !rc)
2418                                 rc = err;
2419                         OBD_FREE_PTR(cbdata);
2420                 }
2421         }
2422
2423         RETURN(rc);
2424 }
2425 EXPORT_SYMBOL(ptlrpc_set_wait);
2426
2427 /**
2428  * Helper fuction for request freeing.
2429  * Called when request count reached zero and request needs to be freed.
2430  * Removes request from all sorts of sending/replay lists it might be on,
2431  * frees network buffers if any are present.
2432  * If \a locked is set, that means caller is already holding import imp_lock
2433  * and so we no longer need to reobtain it (for certain lists manipulations)
2434  */
2435 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2436 {
2437         ENTRY;
2438
2439         if (request == NULL)
2440                 RETURN_EXIT;
2441
2442         LASSERT(!request->rq_srv_req);
2443         LASSERT(request->rq_export == NULL);
2444         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2445         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2446         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2447         LASSERTF(!request->rq_replay, "req %p\n", request);
2448
2449         req_capsule_fini(&request->rq_pill);
2450
2451         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
2452          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2453         if (request->rq_import != NULL) {
2454                 if (!locked)
2455                         spin_lock(&request->rq_import->imp_lock);
2456                 list_del_init(&request->rq_replay_list);
2457                 list_del_init(&request->rq_unreplied_list);
2458                 if (!locked)
2459                         spin_unlock(&request->rq_import->imp_lock);
2460         }
2461         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2462
2463         if (atomic_read(&request->rq_refcount) != 0) {
2464                 DEBUG_REQ(D_ERROR, request,
2465                           "freeing request with nonzero refcount");
2466                 LBUG();
2467         }
2468
2469         if (request->rq_repbuf != NULL)
2470                 sptlrpc_cli_free_repbuf(request);
2471
2472         if (request->rq_import != NULL) {
2473                 class_import_put(request->rq_import);
2474                 request->rq_import = NULL;
2475         }
2476         if (request->rq_bulk != NULL)
2477                 ptlrpc_free_bulk(request->rq_bulk);
2478
2479         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2480                 sptlrpc_cli_free_reqbuf(request);
2481
2482         if (request->rq_cli_ctx)
2483                 sptlrpc_req_put_ctx(request, !locked);
2484
2485         if (request->rq_pool)
2486                 __ptlrpc_free_req_to_pool(request);
2487         else
2488                 ptlrpc_request_cache_free(request);
2489         EXIT;
2490 }
2491
2492 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2493 /**
2494  * Drop one request reference. Must be called with import imp_lock held.
2495  * When reference count drops to zero, request is freed.
2496  */
2497 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2498 {
2499         assert_spin_locked(&request->rq_import->imp_lock);
2500         (void)__ptlrpc_req_finished(request, 1);
2501 }
2502
2503 /**
2504  * Helper function
2505  * Drops one reference count for request \a request.
2506  * \a locked set indicates that caller holds import imp_lock.
2507  * Frees the request whe reference count reaches zero.
2508  *
2509  * \retval 1    the request is freed
2510  * \retval 0    some others still hold references on the request
2511  */
2512 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2513 {
2514         int count;
2515         ENTRY;
2516
2517         if (!request)
2518                 RETURN(1);
2519
2520         LASSERT(request != LP_POISON);
2521         LASSERT(request->rq_reqmsg != LP_POISON);
2522
2523         DEBUG_REQ(D_INFO, request, "refcount now %u",
2524                   atomic_read(&request->rq_refcount) - 1);
2525
2526         spin_lock(&request->rq_lock);
2527         count = atomic_dec_return(&request->rq_refcount);
2528         LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2529
2530         /* For open RPC, the client does not know the EA size (LOV, ACL, and
2531          * so on) before replied, then the client has to reserve very large
2532          * reply buffer. Such buffer will not be released until the RPC freed.
2533          * Since The open RPC is replayable, we need to keep it in the replay
2534          * list until close. If there are a lot of files opened concurrently,
2535          * then the client may be OOM.
2536          *
2537          * If fact, it is unnecessary to keep reply buffer for open replay,
2538          * related EAs have already been saved via mdc_save_lovea() before
2539          * coming here. So it is safe to free the reply buffer some earlier
2540          * before releasing the RPC to avoid client OOM. LU-9514 */
2541         if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2542                 spin_lock(&request->rq_early_free_lock);
2543                 sptlrpc_cli_free_repbuf(request);
2544                 request->rq_repbuf = NULL;
2545                 request->rq_repbuf_len = 0;
2546                 request->rq_repdata = NULL;
2547                 request->rq_reqdata_len = 0;
2548                 spin_unlock(&request->rq_early_free_lock);
2549         }
2550         spin_unlock(&request->rq_lock);
2551
2552         if (!count)
2553                 __ptlrpc_free_req(request, locked);
2554
2555         RETURN(!count);
2556 }
2557
2558 /**
2559  * Drops one reference count for a request.
2560  */
2561 void ptlrpc_req_finished(struct ptlrpc_request *request)
2562 {
2563         __ptlrpc_req_finished(request, 0);
2564 }
2565 EXPORT_SYMBOL(ptlrpc_req_finished);
2566
2567 /**
2568  * Returns xid of a \a request
2569  */
2570 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2571 {
2572         return request->rq_xid;
2573 }
2574 EXPORT_SYMBOL(ptlrpc_req_xid);
2575
2576 /**
2577  * Disengage the client's reply buffer from the network
2578  * NB does _NOT_ unregister any client-side bulk.
2579  * IDEMPOTENT, but _not_ safe against concurrent callers.
2580  * The request owner (i.e. the thread doing the I/O) must call...
2581  * Returns 0 on success or 1 if unregistering cannot be made.
2582  */
2583 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2584 {
2585         int                rc;
2586         struct l_wait_info lwi;
2587
2588         /*
2589          * Might sleep.
2590          */
2591         LASSERT(!in_interrupt());
2592
2593         /* Let's setup deadline for reply unlink. */
2594         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2595             async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2596                 request->rq_reply_deadline = ktime_get_real_seconds() +
2597                                              LONG_UNLINK;
2598
2599         /*
2600          * Nothing left to do.
2601          */
2602         if (!ptlrpc_client_recv_or_unlink(request))
2603                 RETURN(1);
2604
2605         LNetMDUnlink(request->rq_reply_md_h);
2606
2607         /*
2608          * Let's check it once again.
2609          */
2610         if (!ptlrpc_client_recv_or_unlink(request))
2611                 RETURN(1);
2612
2613         /* Move to "Unregistering" phase as reply was not unlinked yet. */
2614         ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2615
2616         /*
2617          * Do not wait for unlink to finish.
2618          */
2619         if (async)
2620                 RETURN(0);
2621
2622         /*
2623          * We have to l_wait_event() whatever the result, to give liblustre
2624          * a chance to run reply_in_callback(), and to make sure we've
2625          * unlinked before returning a req to the pool.
2626          */
2627         for (;;) {
2628                 /* The wq argument is ignored by user-space wait_event macros */
2629                 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2630                                         &request->rq_set->set_waitq :
2631                                         &request->rq_reply_waitq;
2632                 /* Network access will complete in finite time but the HUGE
2633                  * timeout lets us CWARN for visibility of sluggish NALs */
2634                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2635                                            cfs_time_seconds(1), NULL, NULL);
2636                 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2637                                   &lwi);
2638                 if (rc == 0) {
2639                         ptlrpc_rqphase_move(request, request->rq_next_phase);
2640                         RETURN(1);
2641                 }
2642
2643                 LASSERT(rc == -ETIMEDOUT);
2644                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2645                           "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2646                           request->rq_receiving_reply,
2647                           request->rq_req_unlinked,
2648                           request->rq_reply_unlinked);
2649         }
2650         RETURN(0);
2651 }
2652
2653 static void ptlrpc_free_request(struct ptlrpc_request *req)
2654 {
2655         spin_lock(&req->rq_lock);
2656         req->rq_replay = 0;
2657         spin_unlock(&req->rq_lock);
2658
2659         if (req->rq_commit_cb != NULL)
2660                 req->rq_commit_cb(req);
2661         list_del_init(&req->rq_replay_list);
2662
2663         __ptlrpc_req_finished(req, 1);
2664 }
2665
2666 /**
2667  * the request is committed and dropped from the replay list of its import
2668  */
2669 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2670 {
2671         struct obd_import       *imp = req->rq_import;
2672
2673         spin_lock(&imp->imp_lock);
2674         if (list_empty(&req->rq_replay_list)) {
2675                 spin_unlock(&imp->imp_lock);
2676                 return;
2677         }
2678
2679         if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2680                 ptlrpc_free_request(req);
2681
2682         spin_unlock(&imp->imp_lock);
2683 }
2684 EXPORT_SYMBOL(ptlrpc_request_committed);
2685
2686 /**
2687  * Iterates through replay_list on import and prunes
2688  * all requests have transno smaller than last_committed for the
2689  * import and don't have rq_replay set.
2690  * Since requests are sorted in transno order, stops when meetign first
2691  * transno bigger than last_committed.
2692  * caller must hold imp->imp_lock
2693  */
2694 void ptlrpc_free_committed(struct obd_import *imp)
2695 {
2696         struct ptlrpc_request   *req, *saved;
2697         struct ptlrpc_request   *last_req = NULL; /* temporary fire escape */
2698         bool                     skip_committed_list = true;
2699         ENTRY;
2700
2701         LASSERT(imp != NULL);
2702         assert_spin_locked(&imp->imp_lock);
2703
2704         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2705             imp->imp_generation == imp->imp_last_generation_checked) {
2706                 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2707                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2708                 RETURN_EXIT;
2709         }
2710         CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2711                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2712                imp->imp_generation);
2713
2714         if (imp->imp_generation != imp->imp_last_generation_checked ||
2715             imp->imp_last_transno_checked == 0)
2716                 skip_committed_list = false;
2717
2718         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2719         imp->imp_last_generation_checked = imp->imp_generation;
2720
2721         list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2722                                      rq_replay_list) {
2723                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
2724                 LASSERT(req != last_req);
2725                 last_req = req;
2726
2727                 if (req->rq_transno == 0) {
2728                         DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2729                         LBUG();
2730                 }
2731                 if (req->rq_import_generation < imp->imp_generation) {
2732                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2733                         GOTO(free_req, 0);
2734                 }
2735
2736                 /* not yet committed */
2737                 if (req->rq_transno > imp->imp_peer_committed_transno) {
2738                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2739                         break;
2740                 }
2741
2742                 if (req->rq_replay) {
2743                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2744                         list_move_tail(&req->rq_replay_list,
2745                                            &imp->imp_committed_list);
2746                         continue;
2747                 }
2748
2749                 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2750                           imp->imp_peer_committed_transno);
2751 free_req:
2752                 ptlrpc_free_request(req);
2753         }
2754
2755         if (skip_committed_list)
2756                 GOTO(out, 0);
2757
2758         list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2759                                  rq_replay_list) {
2760                 LASSERT(req->rq_transno != 0);
2761                 if (req->rq_import_generation < imp->imp_generation ||
2762                     !req->rq_replay) {
2763                         DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2764                                   req->rq_import_generation <
2765                                   imp->imp_generation ? "stale" : "closed");
2766
2767                         if (imp->imp_replay_cursor == &req->rq_replay_list)
2768                                 imp->imp_replay_cursor =
2769                                         req->rq_replay_list.next;
2770
2771                         ptlrpc_free_request(req);
2772                 }
2773         }
2774 out:
2775         EXIT;
2776 }
2777
2778 void ptlrpc_cleanup_client(struct obd_import *imp)
2779 {
2780         ENTRY;
2781         EXIT;
2782 }
2783
2784 /**
2785  * Schedule previously sent request for resend.
2786  * For bulk requests we assign new xid (to avoid problems with
2787  * lost replies and therefore several transfers landing into same buffer
2788  * from different sending attempts).
2789  */
2790 void ptlrpc_resend_req(struct ptlrpc_request *req)
2791 {
2792         DEBUG_REQ(D_HA, req, "going to resend");
2793         spin_lock(&req->rq_lock);
2794
2795         /* Request got reply but linked to the import list still.
2796            Let ptlrpc_check_set() to process it. */
2797         if (ptlrpc_client_replied(req)) {
2798                 spin_unlock(&req->rq_lock);
2799                 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2800                 return;
2801         }
2802
2803         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2804         req->rq_status = -EAGAIN;
2805
2806         req->rq_resend = 1;
2807         req->rq_net_err = 0;
2808         req->rq_timedout = 0;
2809
2810         ptlrpc_client_wake_req(req);
2811         spin_unlock(&req->rq_lock);
2812 }
2813
2814 /* XXX: this function and rq_status are currently unused */
2815 void ptlrpc_restart_req(struct ptlrpc_request *req)
2816 {
2817         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2818         req->rq_status = -ERESTARTSYS;
2819
2820         spin_lock(&req->rq_lock);
2821         req->rq_restart = 1;
2822         req->rq_timedout = 0;
2823         ptlrpc_client_wake_req(req);
2824         spin_unlock(&req->rq_lock);
2825 }
2826
2827 /**
2828  * Grab additional reference on a request \a req
2829  */
2830 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2831 {
2832         ENTRY;
2833         atomic_inc(&req->rq_refcount);
2834         RETURN(req);
2835 }
2836 EXPORT_SYMBOL(ptlrpc_request_addref);
2837
2838 /**
2839  * Add a request to import replay_list.
2840  * Must be called under imp_lock
2841  */
2842 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2843                                       struct obd_import *imp)
2844 {
2845         struct list_head *tmp;
2846
2847         assert_spin_locked(&imp->imp_lock);
2848
2849         if (req->rq_transno == 0) {
2850                 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2851                 LBUG();
2852         }
2853
2854         /* clear this for new requests that were resent as well
2855            as resent replayed requests. */
2856         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2857
2858         /* don't re-add requests that have been replayed */
2859         if (!list_empty(&req->rq_replay_list))
2860                 return;
2861
2862         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2863
2864         spin_lock(&req->rq_lock);
2865         req->rq_resend = 0;
2866         spin_unlock(&req->rq_lock);
2867
2868         LASSERT(imp->imp_replayable);
2869         /* Balanced in ptlrpc_free_committed, usually. */
2870         ptlrpc_request_addref(req);
2871         list_for_each_prev(tmp, &imp->imp_replay_list) {
2872                 struct ptlrpc_request *iter = list_entry(tmp,
2873                                                          struct ptlrpc_request,
2874                                                          rq_replay_list);
2875
2876                 /* We may have duplicate transnos if we create and then
2877                  * open a file, or for closes retained if to match creating
2878                  * opens, so use req->rq_xid as a secondary key.
2879                  * (See bugs 684, 685, and 428.)
2880                  * XXX no longer needed, but all opens need transnos!
2881                  */
2882                 if (iter->rq_transno > req->rq_transno)
2883                         continue;
2884
2885                 if (iter->rq_transno == req->rq_transno) {
2886                         LASSERT(iter->rq_xid != req->rq_xid);
2887                         if (iter->rq_xid > req->rq_xid)
2888                                 continue;
2889                 }
2890
2891                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2892                 return;
2893         }
2894
2895         list_add(&req->rq_replay_list, &imp->imp_replay_list);
2896 }
2897
2898 /**
2899  * Send request and wait until it completes.
2900  * Returns request processing status.
2901  */
2902 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2903 {
2904         struct ptlrpc_request_set *set;
2905         int rc;
2906         ENTRY;
2907
2908         LASSERT(req->rq_set == NULL);
2909         LASSERT(!req->rq_receiving_reply);
2910
2911         set = ptlrpc_prep_set();
2912         if (set == NULL) {
2913                 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2914                 RETURN(-ENOMEM);
2915         }
2916
2917         /* for distributed debugging */
2918         lustre_msg_set_status(req->rq_reqmsg, current_pid());
2919
2920         /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2921         ptlrpc_request_addref(req);
2922         ptlrpc_set_add_req(set, req);
2923         rc = ptlrpc_set_wait(set);
2924         ptlrpc_set_destroy(set);
2925
2926         RETURN(rc);
2927 }
2928 EXPORT_SYMBOL(ptlrpc_queue_wait);
2929
2930 /**
2931  * Callback used for replayed requests reply processing.
2932  * In case of successful reply calls registered request replay callback.
2933  * In case of error restart replay process.
2934  */
2935 static int ptlrpc_replay_interpret(const struct lu_env *env,
2936                                    struct ptlrpc_request *req,
2937                                    void * data, int rc)
2938 {
2939         struct ptlrpc_replay_async_args *aa = data;
2940         struct obd_import *imp = req->rq_import;
2941
2942         ENTRY;
2943         atomic_dec(&imp->imp_replay_inflight);
2944
2945         /* Note: if it is bulk replay (MDS-MDS replay), then even if
2946          * server got the request, but bulk transfer timeout, let's
2947          * replay the bulk req again */
2948         if (!ptlrpc_client_replied(req) ||
2949             (req->rq_bulk != NULL &&
2950              lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2951                 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2952                 GOTO(out, rc = -ETIMEDOUT);
2953         }
2954
2955         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2956             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2957              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2958                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2959
2960         /** VBR: check version failure */
2961         if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2962                 /** replay was failed due to version mismatch */
2963                 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2964                 spin_lock(&imp->imp_lock);
2965                 imp->imp_vbr_failed = 1;
2966                 imp->imp_no_lock_replay = 1;
2967                 spin_unlock(&imp->imp_lock);
2968                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2969         } else {
2970                 /** The transno had better not change over replay. */
2971                 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2972                          lustre_msg_get_transno(req->rq_repmsg) ||
2973                          lustre_msg_get_transno(req->rq_repmsg) == 0,
2974                          "%#llx/%#llx\n",
2975                          lustre_msg_get_transno(req->rq_reqmsg),
2976                          lustre_msg_get_transno(req->rq_repmsg));
2977         }
2978
2979         spin_lock(&imp->imp_lock);
2980         /** if replays by version then gap occur on server, no trust to locks */
2981         if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2982                 imp->imp_no_lock_replay = 1;
2983         imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2984         spin_unlock(&imp->imp_lock);
2985         LASSERT(imp->imp_last_replay_transno);
2986
2987         /* transaction number shouldn't be bigger than the latest replayed */
2988         if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2989                 DEBUG_REQ(D_ERROR, req,
2990                           "Reported transno %llu is bigger than the "
2991                           "replayed one: %llu", req->rq_transno,
2992                           lustre_msg_get_transno(req->rq_reqmsg));
2993                 GOTO(out, rc = -EINVAL);
2994         }
2995
2996         DEBUG_REQ(D_HA, req, "got rep");
2997
2998         /* let the callback do fixups, possibly including in the request */
2999         if (req->rq_replay_cb)
3000                 req->rq_replay_cb(req);
3001
3002         if (ptlrpc_client_replied(req) &&
3003             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3004                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3005                           lustre_msg_get_status(req->rq_repmsg),
3006                           aa->praa_old_status);
3007
3008                 /* Note: If the replay fails for MDT-MDT recovery, let's
3009                  * abort all of the following requests in the replay
3010                  * and sending list, because MDT-MDT update requests
3011                  * are dependent on each other, see LU-7039 */
3012                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3013                         struct ptlrpc_request *free_req;
3014                         struct ptlrpc_request *tmp;
3015
3016                         spin_lock(&imp->imp_lock);
3017                         list_for_each_entry_safe(free_req, tmp,
3018                                                  &imp->imp_replay_list,
3019                                                  rq_replay_list) {
3020                                 ptlrpc_free_request(free_req);
3021                         }
3022
3023                         list_for_each_entry_safe(free_req, tmp,
3024                                                  &imp->imp_committed_list,
3025                                                  rq_replay_list) {
3026                                 ptlrpc_free_request(free_req);
3027                         }
3028
3029                         list_for_each_entry_safe(free_req, tmp,
3030                                                 &imp->imp_delayed_list,
3031                                                 rq_list) {
3032                                 spin_lock(&free_req->rq_lock);
3033                                 free_req->rq_err = 1;
3034                                 free_req->rq_status = -EIO;
3035                                 ptlrpc_client_wake_req(free_req);
3036                                 spin_unlock(&free_req->rq_lock);
3037                         }
3038
3039                         list_for_each_entry_safe(free_req, tmp,
3040                                                 &imp->imp_sending_list,
3041                                                 rq_list) {
3042                                 spin_lock(&free_req->rq_lock);
3043                                 free_req->rq_err = 1;
3044                                 free_req->rq_status = -EIO;
3045                                 ptlrpc_client_wake_req(free_req);
3046                                 spin_unlock(&free_req->rq_lock);
3047                         }
3048                         spin_unlock(&imp->imp_lock);
3049                 }
3050         } else {
3051                 /* Put it back for re-replay. */
3052                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3053         }
3054
3055         /*
3056          * Errors while replay can set transno to 0, but
3057          * imp_last_replay_transno shouldn't be set to 0 anyway
3058          */
3059         if (req->rq_transno == 0)
3060                 CERROR("Transno is 0 during replay!\n");
3061
3062         /* continue with recovery */
3063         rc = ptlrpc_import_recovery_state_machine(imp);
3064  out:
3065         req->rq_send_state = aa->praa_old_state;
3066
3067         if (rc != 0)
3068                 /* this replay failed, so restart recovery */
3069                 ptlrpc_connect_import(imp);
3070
3071         RETURN(rc);
3072 }
3073
3074 /**
3075  * Prepares and queues request for replay.
3076  * Adds it to ptlrpcd queue for actual sending.
3077  * Returns 0 on success.
3078  */
3079 int ptlrpc_replay_req(struct ptlrpc_request *req)
3080 {
3081         struct ptlrpc_replay_async_args *aa;
3082         ENTRY;
3083
3084         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3085
3086         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
3087         aa = ptlrpc_req_async_args(req);
3088         memset(aa, 0, sizeof *aa);
3089
3090         /* Prepare request to be resent with ptlrpcd */
3091         aa->praa_old_state = req->rq_send_state;
3092         req->rq_send_state = LUSTRE_IMP_REPLAY;
3093         req->rq_phase = RQ_PHASE_NEW;
3094         req->rq_next_phase = RQ_PHASE_UNDEFINED;
3095         if (req->rq_repmsg)
3096                 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3097         req->rq_status = 0;
3098         req->rq_interpret_reply = ptlrpc_replay_interpret;
3099         /* Readjust the timeout for current conditions */
3100         ptlrpc_at_set_req_timeout(req);
3101
3102         /* Tell server the net_latency, so the server can calculate how long
3103          * it should wait for next replay */
3104         lustre_msg_set_service_time(req->rq_reqmsg,
3105                                     ptlrpc_at_get_net_latency(req));
3106         DEBUG_REQ(D_HA, req, "REPLAY");
3107
3108         atomic_inc(&req->rq_import->imp_replay_inflight);
3109         spin_lock(&req->rq_lock);
3110         req->rq_early_free_repbuf = 0;
3111         spin_unlock(&req->rq_lock);
3112         ptlrpc_request_addref(req);     /* ptlrpcd needs a ref */
3113
3114         ptlrpcd_add_req(req);
3115         RETURN(0);
3116 }
3117
3118 /**
3119  * Aborts all in-flight request on import \a imp sending and delayed lists
3120  */
3121 void ptlrpc_abort_inflight(struct obd_import *imp)
3122 {
3123         struct list_head *tmp, *n;
3124         ENTRY;
3125
3126         /* Make sure that no new requests get processed for this import.
3127          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3128          * this flag and then putting requests on sending_list or delayed_list.
3129          */
3130         spin_lock(&imp->imp_lock);
3131
3132         /* XXX locking?  Maybe we should remove each request with the list
3133          * locked?  Also, how do we know if the requests on the list are
3134          * being freed at this time?
3135          */
3136         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3137                 struct ptlrpc_request *req = list_entry(tmp,
3138                                                         struct ptlrpc_request,
3139                                                         rq_list);
3140
3141                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3142
3143                 spin_lock(&req->rq_lock);
3144                 if (req->rq_import_generation < imp->imp_generation) {
3145                         req->rq_err = 1;
3146                         req->rq_status = -EIO;
3147                         ptlrpc_client_wake_req(req);
3148                 }
3149                 spin_unlock(&req->rq_lock);
3150         }
3151
3152         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3153                 struct ptlrpc_request *req =
3154                         list_entry(tmp, struct ptlrpc_request, rq_list);
3155
3156                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3157
3158                 spin_lock(&req->rq_lock);
3159                 if (req->rq_import_generation < imp->imp_generation) {
3160                         req->rq_err = 1;
3161                         req->rq_status = -EIO;
3162                         ptlrpc_client_wake_req(req);
3163                 }
3164                 spin_unlock(&req->rq_lock);
3165         }
3166
3167         /* Last chance to free reqs left on the replay list, but we
3168          * will still leak reqs that haven't committed.  */
3169         if (imp->imp_replayable)
3170                 ptlrpc_free_committed(imp);
3171
3172         spin_unlock(&imp->imp_lock);
3173
3174         EXIT;
3175 }
3176
3177 /**
3178  * Abort all uncompleted requests in request set \a set
3179  */
3180 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3181 {
3182         struct list_head *tmp, *pos;
3183
3184         LASSERT(set != NULL);
3185
3186         list_for_each_safe(pos, tmp, &set->set_requests) {
3187                 struct ptlrpc_request *req =
3188                         list_entry(pos, struct ptlrpc_request,
3189                                    rq_set_chain);
3190
3191                 spin_lock(&req->rq_lock);
3192                 if (req->rq_phase != RQ_PHASE_RPC) {
3193                         spin_unlock(&req->rq_lock);
3194                         continue;
3195                 }
3196
3197                 req->rq_err = 1;
3198                 req->rq_status = -EINTR;
3199                 ptlrpc_client_wake_req(req);
3200                 spin_unlock(&req->rq_lock);
3201         }
3202 }
3203
3204 static __u64 ptlrpc_last_xid;
3205 static spinlock_t ptlrpc_last_xid_lock;
3206
3207 /**
3208  * Initialize the XID for the node.  This is common among all requests on
3209  * this node, and only requires the property that it is monotonically
3210  * increasing.  It does not need to be sequential.  Since this is also used
3211  * as the RDMA match bits, it is important that a single client NOT have
3212  * the same match bits for two different in-flight requests, hence we do
3213  * NOT want to have an XID per target or similar.
3214  *