Whamcloud - gitweb
LU-7809 lod: stop recovery before destory dtrq list
[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.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2015, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 /** Implementation of client-side PortalRPC interfaces */
38
39 #define DEBUG_SUBSYSTEM S_RPC
40
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
47
48 #include "ptlrpc_internal.h"
49
50 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
51         .add_kiov_frag  = ptlrpc_prep_bulk_page_pin,
52         .release_frags  = ptlrpc_release_bulk_page_pin,
53 };
54 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
55
56 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
57         .add_kiov_frag  = ptlrpc_prep_bulk_page_nopin,
58         .release_frags  = ptlrpc_release_bulk_noop,
59 };
60 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
61
62 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
63         .add_iov_frag = ptlrpc_prep_bulk_frag,
64 };
65 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
66
67 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
68 static int ptlrpcd_check_work(struct ptlrpc_request *req);
69 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
70
71 /**
72  * Initialize passed in client structure \a cl.
73  */
74 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
75                         struct ptlrpc_client *cl)
76 {
77         cl->cli_request_portal = req_portal;
78         cl->cli_reply_portal   = rep_portal;
79         cl->cli_name           = name;
80 }
81 EXPORT_SYMBOL(ptlrpc_init_client);
82
83 /**
84  * Return PortalRPC connection for remore uud \a uuid
85  */
86 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
87 {
88         struct ptlrpc_connection *c;
89         lnet_nid_t                self;
90         lnet_process_id_t         peer;
91         int                       err;
92
93         /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
94          * before accessing its values. */
95         /* coverity[uninit_use_in_call] */
96         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
97         if (err != 0) {
98                 CNETERR("cannot find peer %s!\n", uuid->uuid);
99                 return NULL;
100         }
101
102         c = ptlrpc_connection_get(peer, self, uuid);
103         if (c) {
104                 memcpy(c->c_remote_uuid.uuid,
105                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
106         }
107
108         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
109
110         return c;
111 }
112
113 /**
114  * Allocate and initialize new bulk descriptor on the sender.
115  * Returns pointer to the descriptor or NULL on error.
116  */
117 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
118                                          enum ptlrpc_bulk_op_type type,
119                                          unsigned portal,
120                                          const struct ptlrpc_bulk_frag_ops *ops)
121 {
122         struct ptlrpc_bulk_desc *desc;
123         int i;
124
125         /* ensure that only one of KIOV or IOVEC is set but not both */
126         LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
127                  ops->add_kiov_frag != NULL) ||
128                 (ptlrpc_is_bulk_desc_kvec(type) &&
129                  ops->add_iov_frag != NULL));
130
131         OBD_ALLOC_PTR(desc);
132         if (desc == NULL)
133                 return NULL;
134         if (type & PTLRPC_BULK_BUF_KIOV) {
135                 OBD_ALLOC_LARGE(GET_KIOV(desc),
136                                 nfrags * sizeof(*GET_KIOV(desc)));
137                 if (GET_KIOV(desc) == NULL)
138                         goto out;
139         } else {
140                 OBD_ALLOC_LARGE(GET_KVEC(desc),
141                                 nfrags * sizeof(*GET_KVEC(desc)));
142                 if (GET_KVEC(desc) == NULL)
143                         goto out;
144         }
145
146         spin_lock_init(&desc->bd_lock);
147         init_waitqueue_head(&desc->bd_waitq);
148         desc->bd_max_iov = nfrags;
149         desc->bd_iov_count = 0;
150         desc->bd_portal = portal;
151         desc->bd_type = type;
152         desc->bd_md_count = 0;
153         desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
154         LASSERT(max_brw > 0);
155         desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
156         /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
157          * node. Negotiated ocd_brw_size will always be <= this number. */
158         for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
159                 LNetInvalidateHandle(&desc->bd_mds[i]);
160
161         return desc;
162 out:
163         OBD_FREE_PTR(desc);
164         return NULL;
165 }
166
167 /**
168  * Prepare bulk descriptor for specified outgoing request \a req that
169  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
170  * the bulk to be sent. Used on client-side.
171  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
172  * error.
173  */
174 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
175                                               unsigned nfrags, unsigned max_brw,
176                                               unsigned int type,
177                                               unsigned portal,
178                                               const struct ptlrpc_bulk_frag_ops
179                                                 *ops)
180 {
181         struct obd_import *imp = req->rq_import;
182         struct ptlrpc_bulk_desc *desc;
183
184         ENTRY;
185         LASSERT(ptlrpc_is_bulk_op_passive(type));
186
187         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
188         if (desc == NULL)
189                 RETURN(NULL);
190
191         desc->bd_import_generation = req->rq_import_generation;
192         desc->bd_import = class_import_get(imp);
193         desc->bd_req = req;
194
195         desc->bd_cbid.cbid_fn  = client_bulk_callback;
196         desc->bd_cbid.cbid_arg = desc;
197
198         /* This makes req own desc, and free it when she frees herself */
199         req->rq_bulk = desc;
200
201         return desc;
202 }
203 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
204
205 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
206                              struct page *page, int pageoffset, int len,
207                              int pin)
208 {
209         lnet_kiov_t *kiov;
210
211         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
212         LASSERT(page != NULL);
213         LASSERT(pageoffset >= 0);
214         LASSERT(len > 0);
215         LASSERT(pageoffset + len <= PAGE_SIZE);
216         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
217
218         kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
219
220         desc->bd_nob += len;
221
222         if (pin)
223                 page_cache_get(page);
224
225         kiov->kiov_page = page;
226         kiov->kiov_offset = pageoffset;
227         kiov->kiov_len = len;
228
229         desc->bd_iov_count++;
230 }
231 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
232
233 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
234                           void *frag, int len)
235 {
236         struct kvec *iovec;
237         ENTRY;
238
239         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
240         LASSERT(frag != NULL);
241         LASSERT(len > 0);
242         LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
243
244         iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
245
246         desc->bd_nob += len;
247
248         iovec->iov_base = frag;
249         iovec->iov_len = len;
250
251         desc->bd_iov_count++;
252
253         RETURN(desc->bd_nob);
254 }
255 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
256
257 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
258 {
259         ENTRY;
260
261         LASSERT(desc != NULL);
262         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
263         LASSERT(desc->bd_md_count == 0);         /* network hands off */
264         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
265         LASSERT(desc->bd_frag_ops != NULL);
266
267         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
268                 sptlrpc_enc_pool_put_pages(desc);
269
270         if (desc->bd_export)
271                 class_export_put(desc->bd_export);
272         else
273                 class_import_put(desc->bd_import);
274
275         if (desc->bd_frag_ops->release_frags != NULL)
276                 desc->bd_frag_ops->release_frags(desc);
277
278         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
279                 OBD_FREE_LARGE(GET_KIOV(desc),
280                         desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
281         else
282                 OBD_FREE_LARGE(GET_KVEC(desc),
283                         desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
284         OBD_FREE_PTR(desc);
285         EXIT;
286 }
287 EXPORT_SYMBOL(ptlrpc_free_bulk);
288
289 /**
290  * Set server timelimit for this req, i.e. how long are we willing to wait
291  * for reply before timing out this request.
292  */
293 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
294 {
295         __u32 serv_est;
296         int idx;
297         struct imp_at *at;
298
299         LASSERT(req->rq_import);
300
301         if (AT_OFF) {
302                 /* non-AT settings */
303                 /**
304                  * \a imp_server_timeout means this is reverse import and
305                  * we send (currently only) ASTs to the client and cannot afford
306                  * to wait too long for the reply, otherwise the other client
307                  * (because of which we are sending this request) would
308                  * timeout waiting for us
309                  */
310                 req->rq_timeout = req->rq_import->imp_server_timeout ?
311                                   obd_timeout / 2 : obd_timeout;
312         } else {
313                 at = &req->rq_import->imp_at;
314                 idx = import_at_get_index(req->rq_import,
315                                           req->rq_request_portal);
316                 serv_est = at_get(&at->iat_service_estimate[idx]);
317                 req->rq_timeout = at_est2timeout(serv_est);
318         }
319         /* We could get even fancier here, using history to predict increased
320            loading... */
321
322         /* Let the server know what this RPC timeout is by putting it in the
323            reqmsg*/
324         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
325 }
326 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
327
328 /* Adjust max service estimate based on server value */
329 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
330                                   unsigned int serv_est)
331 {
332         int idx;
333         unsigned int oldse;
334         struct imp_at *at;
335
336         LASSERT(req->rq_import);
337         at = &req->rq_import->imp_at;
338
339         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
340         /* max service estimates are tracked on the server side,
341            so just keep minimal history here */
342         oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
343         if (oldse != 0)
344                 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
345                        "has changed from %d to %d\n",
346                        req->rq_import->imp_obd->obd_name,req->rq_request_portal,
347                        oldse, at_get(&at->iat_service_estimate[idx]));
348 }
349
350 /* Expected network latency per remote node (secs) */
351 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
352 {
353         return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
354 }
355
356 /* Adjust expected network latency */
357 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
358                                unsigned int service_time)
359 {
360         unsigned int nl, oldnl;
361         struct imp_at *at;
362         time_t now = cfs_time_current_sec();
363
364         LASSERT(req->rq_import);
365
366         if (service_time > now - req->rq_sent + 3) {
367                 /* bz16408, however, this can also happen if early reply
368                  * is lost and client RPC is expired and resent, early reply
369                  * or reply of original RPC can still be fit in reply buffer
370                  * of resent RPC, now client is measuring time from the
371                  * resent time, but server sent back service time of original
372                  * RPC.
373                  */
374                 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
375                        D_ADAPTTO : D_WARNING,
376                        "Reported service time %u > total measured time "
377                        CFS_DURATION_T"\n", service_time,
378                        cfs_time_sub(now, req->rq_sent));
379                 return;
380         }
381
382         /* Network latency is total time less server processing time */
383         nl = max_t(int, now - req->rq_sent -
384                         service_time, 0) + 1; /* st rounding */
385         at = &req->rq_import->imp_at;
386
387         oldnl = at_measured(&at->iat_net_latency, nl);
388         if (oldnl != 0)
389                 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
390                        "has changed from %d to %d\n",
391                        req->rq_import->imp_obd->obd_name,
392                        obd_uuid2str(
393                                &req->rq_import->imp_connection->c_remote_uuid),
394                        oldnl, at_get(&at->iat_net_latency));
395 }
396
397 static int unpack_reply(struct ptlrpc_request *req)
398 {
399         int rc;
400
401         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
402                 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
403                 if (rc) {
404                         DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
405                         return(-EPROTO);
406                 }
407         }
408
409         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
410         if (rc) {
411                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
412                 return(-EPROTO);
413         }
414         return 0;
415 }
416
417 /**
418  * Handle an early reply message, called with the rq_lock held.
419  * If anything goes wrong just ignore it - same as if it never happened
420  */
421 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
422 __must_hold(&req->rq_lock)
423 {
424         struct ptlrpc_request *early_req;
425         time_t                 olddl;
426         int                    rc;
427         ENTRY;
428
429         req->rq_early = 0;
430         spin_unlock(&req->rq_lock);
431
432         rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
433         if (rc) {
434                 spin_lock(&req->rq_lock);
435                 RETURN(rc);
436         }
437
438         rc = unpack_reply(early_req);
439         if (rc != 0) {
440                 sptlrpc_cli_finish_early_reply(early_req);
441                 spin_lock(&req->rq_lock);
442                 RETURN(rc);
443         }
444
445         /* Use new timeout value just to adjust the local value for this
446          * request, don't include it into at_history. It is unclear yet why
447          * service time increased and should it be counted or skipped, e.g.
448          * that can be recovery case or some error or server, the real reply
449          * will add all new data if it is worth to add. */
450         req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
451         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
452
453         /* Network latency can be adjusted, it is pure network delays */
454         ptlrpc_at_adj_net_latency(req,
455                         lustre_msg_get_service_time(early_req->rq_repmsg));
456
457         sptlrpc_cli_finish_early_reply(early_req);
458
459         spin_lock(&req->rq_lock);
460         olddl = req->rq_deadline;
461         /* server assumes it now has rq_timeout from when the request
462          * arrived, so the client should give it at least that long.
463          * since we don't know the arrival time we'll use the original
464          * sent time */
465         req->rq_deadline = req->rq_sent + req->rq_timeout +
466                            ptlrpc_at_get_net_latency(req);
467
468         DEBUG_REQ(D_ADAPTTO, req,
469                   "Early reply #%d, new deadline in "CFS_DURATION_T"s "
470                   "("CFS_DURATION_T"s)", req->rq_early_count,
471                   cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
472                   cfs_time_sub(req->rq_deadline, olddl));
473
474         RETURN(rc);
475 }
476
477 static struct kmem_cache *request_cache;
478
479 int ptlrpc_request_cache_init(void)
480 {
481         request_cache = kmem_cache_create("ptlrpc_cache",
482                                           sizeof(struct ptlrpc_request),
483                                           0, SLAB_HWCACHE_ALIGN, NULL);
484         return request_cache == NULL ? -ENOMEM : 0;
485 }
486
487 void ptlrpc_request_cache_fini(void)
488 {
489         kmem_cache_destroy(request_cache);
490 }
491
492 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
493 {
494         struct ptlrpc_request *req;
495
496         OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
497         return req;
498 }
499
500 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
501 {
502         OBD_SLAB_FREE_PTR(req, request_cache);
503 }
504
505 /**
506  * Wind down request pool \a pool.
507  * Frees all requests from the pool too
508  */
509 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
510 {
511         struct list_head *l, *tmp;
512         struct ptlrpc_request *req;
513
514         LASSERT(pool != NULL);
515
516         spin_lock(&pool->prp_lock);
517         list_for_each_safe(l, tmp, &pool->prp_req_list) {
518                 req = list_entry(l, struct ptlrpc_request, rq_list);
519                 list_del(&req->rq_list);
520                 LASSERT(req->rq_reqbuf);
521                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
522                 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
523                 ptlrpc_request_cache_free(req);
524         }
525         spin_unlock(&pool->prp_lock);
526         OBD_FREE(pool, sizeof(*pool));
527 }
528 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
529
530 /**
531  * Allocates, initializes and adds \a num_rq requests to the pool \a pool
532  */
533 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
534 {
535         int i;
536         int size = 1;
537
538         while (size < pool->prp_rq_size)
539                 size <<= 1;
540
541         LASSERTF(list_empty(&pool->prp_req_list) ||
542                  size == pool->prp_rq_size,
543                  "Trying to change pool size with nonempty pool "
544                  "from %d to %d bytes\n", pool->prp_rq_size, size);
545
546         spin_lock(&pool->prp_lock);
547         pool->prp_rq_size = size;
548         for (i = 0; i < num_rq; i++) {
549                 struct ptlrpc_request *req;
550                 struct lustre_msg *msg;
551
552                 spin_unlock(&pool->prp_lock);
553                 req = ptlrpc_request_cache_alloc(GFP_NOFS);
554                 if (!req)
555                         return i;
556                 OBD_ALLOC_LARGE(msg, size);
557                 if (!msg) {
558                         ptlrpc_request_cache_free(req);
559                         return i;
560                 }
561                 req->rq_reqbuf = msg;
562                 req->rq_reqbuf_len = size;
563                 req->rq_pool = pool;
564                 spin_lock(&pool->prp_lock);
565                 list_add_tail(&req->rq_list, &pool->prp_req_list);
566         }
567         spin_unlock(&pool->prp_lock);
568         return num_rq;
569 }
570 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
571
572 /**
573  * Create and initialize new request pool with given attributes:
574  * \a num_rq - initial number of requests to create for the pool
575  * \a msgsize - maximum message size possible for requests in thid pool
576  * \a populate_pool - function to be called when more requests need to be added
577  *                    to the pool
578  * Returns pointer to newly created pool or NULL on error.
579  */
580 struct ptlrpc_request_pool *
581 ptlrpc_init_rq_pool(int num_rq, int msgsize,
582                     int (*populate_pool)(struct ptlrpc_request_pool *, int))
583 {
584         struct ptlrpc_request_pool *pool;
585
586         OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
587         if (!pool)
588                 return NULL;
589
590         /* Request next power of two for the allocation, because internally
591            kernel would do exactly this */
592
593         spin_lock_init(&pool->prp_lock);
594         INIT_LIST_HEAD(&pool->prp_req_list);
595         pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
596         pool->prp_populate = populate_pool;
597
598         populate_pool(pool, num_rq);
599
600         return pool;
601 }
602 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
603
604 /**
605  * Fetches one request from pool \a pool
606  */
607 static struct ptlrpc_request *
608 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
609 {
610         struct ptlrpc_request *request;
611         struct lustre_msg *reqbuf;
612
613         if (!pool)
614                 return NULL;
615
616         spin_lock(&pool->prp_lock);
617
618         /* See if we have anything in a pool, and bail out if nothing,
619          * in writeout path, where this matters, this is safe to do, because
620          * nothing is lost in this case, and when some in-flight requests
621          * complete, this code will be called again. */
622         if (unlikely(list_empty(&pool->prp_req_list))) {
623                 spin_unlock(&pool->prp_lock);
624                 return NULL;
625         }
626
627         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
628                              rq_list);
629         list_del_init(&request->rq_list);
630         spin_unlock(&pool->prp_lock);
631
632         LASSERT(request->rq_reqbuf);
633         LASSERT(request->rq_pool);
634
635         reqbuf = request->rq_reqbuf;
636         memset(request, 0, sizeof(*request));
637         request->rq_reqbuf = reqbuf;
638         request->rq_reqbuf_len = pool->prp_rq_size;
639         request->rq_pool = pool;
640
641         return request;
642 }
643
644 /**
645  * Returns freed \a request to pool.
646  */
647 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
648 {
649         struct ptlrpc_request_pool *pool = request->rq_pool;
650
651         spin_lock(&pool->prp_lock);
652         LASSERT(list_empty(&request->rq_list));
653         LASSERT(!request->rq_receiving_reply);
654         list_add_tail(&request->rq_list, &pool->prp_req_list);
655         spin_unlock(&pool->prp_lock);
656 }
657
658 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
659 {
660         struct obd_import       *imp = req->rq_import;
661         struct list_head        *tmp;
662         struct ptlrpc_request   *iter;
663
664         assert_spin_locked(&imp->imp_lock);
665         LASSERT(list_empty(&req->rq_unreplied_list));
666
667         /* unreplied list is sorted by xid in ascending order */
668         list_for_each_prev(tmp, &imp->imp_unreplied_list) {
669                 iter = list_entry(tmp, struct ptlrpc_request,
670                                   rq_unreplied_list);
671
672                 LASSERT(req->rq_xid != iter->rq_xid);
673                 if (req->rq_xid < iter->rq_xid)
674                         continue;
675                 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
676                 return;
677         }
678         list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
679 }
680
681 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
682 {
683         req->rq_xid = ptlrpc_next_xid();
684         ptlrpc_add_unreplied(req);
685 }
686
687 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
688 {
689         spin_lock(&req->rq_import->imp_lock);
690         ptlrpc_assign_next_xid_nolock(req);
691         spin_unlock(&req->rq_import->imp_lock);
692 }
693
694 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
695                              __u32 version, int opcode, char **bufs,
696                              struct ptlrpc_cli_ctx *ctx)
697 {
698         int count;
699         struct obd_import *imp;
700         __u32 *lengths;
701         int rc;
702
703         ENTRY;
704
705         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
706         imp = request->rq_import;
707         lengths = request->rq_pill.rc_area[RCL_CLIENT];
708
709         if (ctx != NULL) {
710                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
711         } else {
712                 rc = sptlrpc_req_get_ctx(request);
713                 if (rc)
714                         GOTO(out_free, rc);
715         }
716         sptlrpc_req_set_flavor(request, opcode);
717
718         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
719                                  lengths, bufs);
720         if (rc)
721                 GOTO(out_ctx, rc);
722
723         lustre_msg_add_version(request->rq_reqmsg, version);
724         request->rq_send_state = LUSTRE_IMP_FULL;
725         request->rq_type = PTL_RPC_MSG_REQUEST;
726
727         request->rq_req_cbid.cbid_fn  = request_out_callback;
728         request->rq_req_cbid.cbid_arg = request;
729
730         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
731         request->rq_reply_cbid.cbid_arg = request;
732
733         request->rq_reply_deadline = 0;
734         request->rq_bulk_deadline = 0;
735         request->rq_req_deadline = 0;
736         request->rq_phase = RQ_PHASE_NEW;
737         request->rq_next_phase = RQ_PHASE_UNDEFINED;
738
739         request->rq_request_portal = imp->imp_client->cli_request_portal;
740         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
741
742         ptlrpc_at_set_req_timeout(request);
743
744         lustre_msg_set_opc(request->rq_reqmsg, opcode);
745         ptlrpc_assign_next_xid(request);
746
747         /* Let's setup deadline for req/reply/bulk unlink for opcode. */
748         if (cfs_fail_val == opcode) {
749                 time_t *fail_t = NULL, *fail2_t = NULL;
750
751                 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
752                         fail_t = &request->rq_bulk_deadline;
753                 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
754                         fail_t = &request->rq_reply_deadline;
755                 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
756                         fail_t = &request->rq_req_deadline;
757                 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
758                         fail_t = &request->rq_reply_deadline;
759                         fail2_t = &request->rq_bulk_deadline;
760                 }
761
762                 if (fail_t) {
763                         *fail_t = cfs_time_current_sec() + LONG_UNLINK;
764
765                         if (fail2_t)
766                                 *fail2_t = cfs_time_current_sec() + LONG_UNLINK;
767
768                         /* The RPC is infected, let the test to change the
769                          * fail_loc */
770                         set_current_state(TASK_UNINTERRUPTIBLE);
771                         schedule_timeout(cfs_time_seconds(2));
772                         set_current_state(TASK_RUNNING);
773                 }
774         }
775
776         RETURN(0);
777
778 out_ctx:
779         LASSERT(!request->rq_pool);
780         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
781 out_free:
782         class_import_put(imp);
783
784         return rc;
785
786 }
787 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
788
789 /**
790  * Pack request buffers for network transfer, performing necessary encryption
791  * steps if necessary.
792  */
793 int ptlrpc_request_pack(struct ptlrpc_request *request,
794                         __u32 version, int opcode)
795 {
796         int rc;
797         rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
798         if (rc)
799                 return rc;
800
801         /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
802          * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
803          * have to send old ptlrpc_body to keep interoprability with these
804          * clients.
805          *
806          * Only three kinds of server->client RPCs so far:
807          *  - LDLM_BL_CALLBACK
808          *  - LDLM_CP_CALLBACK
809          *  - LDLM_GL_CALLBACK
810          *
811          * XXX This should be removed whenever we drop the interoprability with
812          *     the these old clients.
813          */
814         if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
815             opcode == LDLM_GL_CALLBACK)
816                 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
817                                    sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
818
819         return rc;
820 }
821 EXPORT_SYMBOL(ptlrpc_request_pack);
822
823 /**
824  * Helper function to allocate new request on import \a imp
825  * and possibly using existing request from pool \a pool if provided.
826  * Returns allocated request structure with import field filled or
827  * NULL on error.
828  */
829 static inline
830 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
831                                               struct ptlrpc_request_pool *pool)
832 {
833         struct ptlrpc_request *request = NULL;
834
835         request = ptlrpc_request_cache_alloc(GFP_NOFS);
836
837         if (!request && pool)
838                 request = ptlrpc_prep_req_from_pool(pool);
839
840         if (request) {
841                 ptlrpc_cli_req_init(request);
842
843                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
844                 LASSERT(imp != LP_POISON);
845                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
846                         imp->imp_client);
847                 LASSERT(imp->imp_client != LP_POISON);
848
849                 request->rq_import = class_import_get(imp);
850         } else {
851                 CERROR("request allocation out of memory\n");
852         }
853
854         return request;
855 }
856
857 /**
858  * Helper function for creating a request.
859  * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
860  * buffer structures according to capsule template \a format.
861  * Returns allocated request structure pointer or NULL on error.
862  */
863 static struct ptlrpc_request *
864 ptlrpc_request_alloc_internal(struct obd_import *imp,
865                               struct ptlrpc_request_pool * pool,
866                               const struct req_format *format)
867 {
868         struct ptlrpc_request *request;
869
870         request = __ptlrpc_request_alloc(imp, pool);
871         if (request == NULL)
872                 return NULL;
873
874         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
875         req_capsule_set(&request->rq_pill, format);
876         return request;
877 }
878
879 /**
880  * Allocate new request structure for import \a imp and initialize its
881  * buffer structure according to capsule template \a format.
882  */
883 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
884                                             const struct req_format *format)
885 {
886         return ptlrpc_request_alloc_internal(imp, NULL, format);
887 }
888 EXPORT_SYMBOL(ptlrpc_request_alloc);
889
890 /**
891  * Allocate new request structure for import \a imp from pool \a pool and
892  * initialize its buffer structure according to capsule template \a format.
893  */
894 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
895                                             struct ptlrpc_request_pool * pool,
896                                             const struct req_format *format)
897 {
898         return ptlrpc_request_alloc_internal(imp, pool, format);
899 }
900 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
901
902 /**
903  * For requests not from pool, free memory of the request structure.
904  * For requests obtained from a pool earlier, return request back to pool.
905  */
906 void ptlrpc_request_free(struct ptlrpc_request *request)
907 {
908         if (request->rq_pool)
909                 __ptlrpc_free_req_to_pool(request);
910         else
911                 ptlrpc_request_cache_free(request);
912 }
913 EXPORT_SYMBOL(ptlrpc_request_free);
914
915 /**
916  * Allocate new request for operatione \a opcode and immediatelly pack it for
917  * network transfer.
918  * Only used for simple requests like OBD_PING where the only important
919  * part of the request is operation itself.
920  * Returns allocated request or NULL on error.
921  */
922 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
923                                                 const struct req_format *format,
924                                                 __u32 version, int opcode)
925 {
926         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
927         int                    rc;
928
929         if (req) {
930                 rc = ptlrpc_request_pack(req, version, opcode);
931                 if (rc) {
932                         ptlrpc_request_free(req);
933                         req = NULL;
934                 }
935         }
936         return req;
937 }
938 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
939
940 /**
941  * Allocate and initialize new request set structure on the current CPT.
942  * Returns a pointer to the newly allocated set structure or NULL on error.
943  */
944 struct ptlrpc_request_set *ptlrpc_prep_set(void)
945 {
946         struct ptlrpc_request_set       *set;
947         int                             cpt;
948
949         ENTRY;
950         cpt = cfs_cpt_current(cfs_cpt_table, 0);
951         OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
952         if (!set)
953                 RETURN(NULL);
954         atomic_set(&set->set_refcount, 1);
955         INIT_LIST_HEAD(&set->set_requests);
956         init_waitqueue_head(&set->set_waitq);
957         atomic_set(&set->set_new_count, 0);
958         atomic_set(&set->set_remaining, 0);
959         spin_lock_init(&set->set_new_req_lock);
960         INIT_LIST_HEAD(&set->set_new_requests);
961         INIT_LIST_HEAD(&set->set_cblist);
962         set->set_max_inflight = UINT_MAX;
963         set->set_producer     = NULL;
964         set->set_producer_arg = NULL;
965         set->set_rc           = 0;
966
967         RETURN(set);
968 }
969 EXPORT_SYMBOL(ptlrpc_prep_set);
970
971 /**
972  * Allocate and initialize new request set structure with flow control
973  * extension. This extension allows to control the number of requests in-flight
974  * for the whole set. A callback function to generate requests must be provided
975  * and the request set will keep the number of requests sent over the wire to
976  * @max_inflight.
977  * Returns a pointer to the newly allocated set structure or NULL on error.
978  */
979 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
980                                              void *arg)
981
982 {
983         struct ptlrpc_request_set *set;
984
985         set = ptlrpc_prep_set();
986         if (!set)
987                 RETURN(NULL);
988
989         set->set_max_inflight  = max;
990         set->set_producer      = func;
991         set->set_producer_arg  = arg;
992
993         RETURN(set);
994 }
995
996 /**
997  * Wind down and free request set structure previously allocated with
998  * ptlrpc_prep_set.
999  * Ensures that all requests on the set have completed and removes
1000  * all requests from the request list in a set.
1001  * If any unsent request happen to be on the list, pretends that they got
1002  * an error in flight and calls their completion handler.
1003  */
1004 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1005 {
1006         struct list_head        *tmp;
1007         struct list_head        *next;
1008         int                      expected_phase;
1009         int                      n = 0;
1010         ENTRY;
1011
1012         /* Requests on the set should either all be completed, or all be new */
1013         expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1014                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1015         list_for_each(tmp, &set->set_requests) {
1016                 struct ptlrpc_request *req =
1017                         list_entry(tmp, struct ptlrpc_request,
1018                                    rq_set_chain);
1019
1020                 LASSERT(req->rq_phase == expected_phase);
1021                 n++;
1022         }
1023
1024         LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1025                  atomic_read(&set->set_remaining) == n, "%d / %d\n",
1026                  atomic_read(&set->set_remaining), n);
1027
1028         list_for_each_safe(tmp, next, &set->set_requests) {
1029                 struct ptlrpc_request *req =
1030                         list_entry(tmp, struct ptlrpc_request,
1031                                    rq_set_chain);
1032                 list_del_init(&req->rq_set_chain);
1033
1034                 LASSERT(req->rq_phase == expected_phase);
1035
1036                 if (req->rq_phase == RQ_PHASE_NEW) {
1037                         ptlrpc_req_interpret(NULL, req, -EBADR);
1038                         atomic_dec(&set->set_remaining);
1039                 }
1040
1041                 spin_lock(&req->rq_lock);
1042                 req->rq_set = NULL;
1043                 req->rq_invalid_rqset = 0;
1044                 spin_unlock(&req->rq_lock);
1045
1046                 ptlrpc_req_finished (req);
1047         }
1048
1049         LASSERT(atomic_read(&set->set_remaining) == 0);
1050
1051         ptlrpc_reqset_put(set);
1052         EXIT;
1053 }
1054 EXPORT_SYMBOL(ptlrpc_set_destroy);
1055
1056 /**
1057  * Add a callback function \a fn to the set.
1058  * This function would be called when all requests on this set are completed.
1059  * The function will be passed \a data argument.
1060  */
1061 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1062                       set_interpreter_func fn, void *data)
1063 {
1064         struct ptlrpc_set_cbdata *cbdata;
1065
1066         OBD_ALLOC_PTR(cbdata);
1067         if (cbdata == NULL)
1068                 RETURN(-ENOMEM);
1069
1070         cbdata->psc_interpret = fn;
1071         cbdata->psc_data = data;
1072         list_add_tail(&cbdata->psc_item, &set->set_cblist);
1073
1074         RETURN(0);
1075 }
1076
1077 /**
1078  * Add a new request to the general purpose request set.
1079  * Assumes request reference from the caller.
1080  */
1081 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1082                         struct ptlrpc_request *req)
1083 {
1084         LASSERT(list_empty(&req->rq_set_chain));
1085
1086         if (req->rq_allow_intr)
1087                 set->set_allow_intr = 1;
1088
1089         /* The set takes over the caller's request reference */
1090         list_add_tail(&req->rq_set_chain, &set->set_requests);
1091         req->rq_set = set;
1092         atomic_inc(&set->set_remaining);
1093         req->rq_queued_time = cfs_time_current();
1094
1095         if (req->rq_reqmsg != NULL)
1096                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1097
1098         if (set->set_producer != NULL)
1099                 /* If the request set has a producer callback, the RPC must be
1100                  * sent straight away */
1101                 ptlrpc_send_new_req(req);
1102 }
1103 EXPORT_SYMBOL(ptlrpc_set_add_req);
1104
1105 /**
1106  * Add a request to a request with dedicated server thread
1107  * and wake the thread to make any necessary processing.
1108  * Currently only used for ptlrpcd.
1109  */
1110 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1111                            struct ptlrpc_request *req)
1112 {
1113         struct ptlrpc_request_set *set = pc->pc_set;
1114         int count, i;
1115
1116         LASSERT(req->rq_set == NULL);
1117         LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1118
1119         spin_lock(&set->set_new_req_lock);
1120         /*
1121          * The set takes over the caller's request reference.
1122          */
1123         req->rq_set = set;
1124         req->rq_queued_time = cfs_time_current();
1125         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1126         count = atomic_inc_return(&set->set_new_count);
1127         spin_unlock(&set->set_new_req_lock);
1128
1129         /* Only need to call wakeup once for the first entry. */
1130         if (count == 1) {
1131                 wake_up(&set->set_waitq);
1132
1133                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1134                  *      guarantee the async RPC can be processed ASAP, we have
1135                  *      no other better choice. It maybe fixed in future. */
1136                 for (i = 0; i < pc->pc_npartners; i++)
1137                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1138         }
1139 }
1140
1141 /**
1142  * Based on the current state of the import, determine if the request
1143  * can be sent, is an error, or should be delayed.
1144  *
1145  * Returns true if this request should be delayed. If false, and
1146  * *status is set, then the request can not be sent and *status is the
1147  * error code.  If false and status is 0, then request can be sent.
1148  *
1149  * The imp->imp_lock must be held.
1150  */
1151 static int ptlrpc_import_delay_req(struct obd_import *imp,
1152                                    struct ptlrpc_request *req, int *status)
1153 {
1154         int delay = 0;
1155         ENTRY;
1156
1157         LASSERT (status != NULL);
1158         *status = 0;
1159
1160         if (req->rq_ctx_init || req->rq_ctx_fini) {
1161                 /* always allow ctx init/fini rpc go through */
1162         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1163                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1164                 *status = -EIO;
1165         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1166                 /* pings may safely race with umount */
1167                 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1168                           D_HA : D_ERROR, req, "IMP_CLOSED ");
1169                 *status = -EIO;
1170         } else if (ptlrpc_send_limit_expired(req)) {
1171                 /* probably doesn't need to be a D_ERROR after initial testing*/
1172                 DEBUG_REQ(D_HA, req, "send limit expired ");
1173                 *status = -ETIMEDOUT;
1174         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1175                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
1176                 /* allow CONNECT even if import is invalid */ ;
1177                 if (atomic_read(&imp->imp_inval_count) != 0) {
1178                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1179                         *status = -EIO;
1180                 }
1181         } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1182                 if (!imp->imp_deactive)
1183                         DEBUG_REQ(D_NET, req, "IMP_INVALID");
1184                 *status = -ESHUTDOWN; /* bz 12940 */
1185         } else if (req->rq_import_generation != imp->imp_generation) {
1186                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1187                 *status = -EIO;
1188         } else if (req->rq_send_state != imp->imp_state) {
1189                 /* invalidate in progress - any requests should be drop */
1190                 if (atomic_read(&imp->imp_inval_count) != 0) {
1191                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1192                         *status = -EIO;
1193                 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1194                         *status = -EWOULDBLOCK;
1195                 } else if (req->rq_allow_replay &&
1196                           (imp->imp_state == LUSTRE_IMP_REPLAY ||
1197                            imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1198                            imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1199                            imp->imp_state == LUSTRE_IMP_RECOVER)) {
1200                         DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1201                 } else {
1202                         delay = 1;
1203                 }
1204         }
1205
1206         RETURN(delay);
1207 }
1208
1209 /**
1210  * Decide if the error message should be printed to the console or not.
1211  * Makes its decision based on request type, status, and failure frequency.
1212  *
1213  * \param[in] req  request that failed and may need a console message
1214  *
1215  * \retval false if no message should be printed
1216  * \retval true  if console message should be printed
1217  */
1218 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1219 {
1220         __u32 opc;
1221
1222         LASSERT(req->rq_reqmsg != NULL);
1223         opc = lustre_msg_get_opc(req->rq_reqmsg);
1224
1225         /* Suppress particular reconnect errors which are to be expected. */
1226         if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1227                 int err;
1228
1229                 /* Suppress timed out reconnect requests */
1230                 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1231                     req->rq_timedout)
1232                         return false;
1233
1234                 /* Suppress most unavailable/again reconnect requests, but
1235                  * print occasionally so it is clear client is trying to
1236                  * connect to a server where no target is running. */
1237                 err = lustre_msg_get_status(req->rq_repmsg);
1238                 if ((err == -ENODEV || err == -EAGAIN) &&
1239                     req->rq_import->imp_conn_cnt % 30 != 20)
1240                         return false;
1241         }
1242
1243         return true;
1244 }
1245
1246 /**
1247  * Check request processing status.
1248  * Returns the status.
1249  */
1250 static int ptlrpc_check_status(struct ptlrpc_request *req)
1251 {
1252         int err;
1253         ENTRY;
1254
1255         err = lustre_msg_get_status(req->rq_repmsg);
1256         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1257                 struct obd_import *imp = req->rq_import;
1258                 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1259                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1260
1261                 if (ptlrpc_console_allow(req))
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         int rc;
1333         struct timeval work_start;
1334         __u64 committed;
1335         long timediff;
1336         ENTRY;
1337
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         do_gettimeofday(&work_start);
1364         timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
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                 time_t  now = cfs_time_current_sec();
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 > cfs_time_current_sec()) &&
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 from PID %d waiting for recovery: "
1561                           "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1562                           ptlrpc_import_state_name(req->rq_send_state),
1563                           ptlrpc_import_state_name(imp->imp_state));
1564                 LASSERT(list_empty(&req->rq_list));
1565                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1566                 atomic_inc(&req->rq_import->imp_inflight);
1567                 spin_unlock(&imp->imp_lock);
1568                 RETURN(0);
1569         }
1570
1571         if (rc != 0) {
1572                 spin_unlock(&imp->imp_lock);
1573                 req->rq_status = rc;
1574                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1575                 RETURN(rc);
1576         }
1577
1578         LASSERT(list_empty(&req->rq_list));
1579         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1580         atomic_inc(&req->rq_import->imp_inflight);
1581
1582         /* find the known replied XID from the unreplied list, CONNECT
1583          * and DISCONNECT requests are skipped to make the sanity check
1584          * on server side happy. see process_req_last_xid().
1585          *
1586          * For CONNECT: Because replay requests have lower XID, it'll
1587          * break the sanity check if CONNECT bump the exp_last_xid on
1588          * server.
1589          *
1590          * For DISCONNECT: Since client will abort inflight RPC before
1591          * sending DISCONNECT, DISCONNECT may carry an XID which higher
1592          * than the inflight RPC.
1593          */
1594         if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1595                 min_xid = ptlrpc_known_replied_xid(imp);
1596         spin_unlock(&imp->imp_lock);
1597
1598         lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1599
1600         lustre_msg_set_status(req->rq_reqmsg, current_pid());
1601
1602         rc = sptlrpc_req_refresh_ctx(req, -1);
1603         if (rc) {
1604                 if (req->rq_err) {
1605                         req->rq_status = rc;
1606                         RETURN(1);
1607                 } else {
1608                         spin_lock(&req->rq_lock);
1609                         req->rq_wait_ctx = 1;
1610                         spin_unlock(&req->rq_lock);
1611                         RETURN(0);
1612                 }
1613         }
1614
1615         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1616                " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1617                imp->imp_obd->obd_uuid.uuid,
1618                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1619                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1620                lustre_msg_get_opc(req->rq_reqmsg));
1621
1622         rc = ptl_send_rpc(req, 0);
1623         if (rc == -ENOMEM) {
1624                 spin_lock(&imp->imp_lock);
1625                 if (!list_empty(&req->rq_list)) {
1626                         list_del_init(&req->rq_list);
1627                         atomic_dec(&req->rq_import->imp_inflight);
1628                 }
1629                 spin_unlock(&imp->imp_lock);
1630                 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1631                 RETURN(rc);
1632         }
1633         if (rc) {
1634                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1635                 spin_lock(&req->rq_lock);
1636                 req->rq_net_err = 1;
1637                 spin_unlock(&req->rq_lock);
1638                 RETURN(rc);
1639         }
1640         RETURN(0);
1641 }
1642
1643 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1644 {
1645         int remaining, rc;
1646         ENTRY;
1647
1648         LASSERT(set->set_producer != NULL);
1649
1650         remaining = atomic_read(&set->set_remaining);
1651
1652         /* populate the ->set_requests list with requests until we
1653          * reach the maximum number of RPCs in flight for this set */
1654         while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1655                 rc = set->set_producer(set, set->set_producer_arg);
1656                 if (rc == -ENOENT) {
1657                         /* no more RPC to produce */
1658                         set->set_producer     = NULL;
1659                         set->set_producer_arg = NULL;
1660                         RETURN(0);
1661                 }
1662         }
1663
1664         RETURN((atomic_read(&set->set_remaining) - remaining));
1665 }
1666
1667 /**
1668  * this sends any unsent RPCs in \a set and returns 1 if all are sent
1669  * and no more replies are expected.
1670  * (it is possible to get less replies than requests sent e.g. due to timed out
1671  * requests or requests that we had trouble to send out)
1672  *
1673  * NOTE: This function contains a potential schedule point (cond_resched()).
1674  */
1675 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1676 {
1677         struct list_head *tmp, *next;
1678         struct list_head  comp_reqs;
1679         int force_timer_recalc = 0;
1680         ENTRY;
1681
1682         if (atomic_read(&set->set_remaining) == 0)
1683                 RETURN(1);
1684
1685         INIT_LIST_HEAD(&comp_reqs);
1686         list_for_each_safe(tmp, next, &set->set_requests) {
1687                 struct ptlrpc_request *req =
1688                         list_entry(tmp, struct ptlrpc_request,
1689                                    rq_set_chain);
1690                 struct obd_import *imp = req->rq_import;
1691                 int unregistered = 0;
1692                 int async = 1;
1693                 int rc = 0;
1694
1695                 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1696                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1697                         continue;
1698                 }
1699
1700                 /* This schedule point is mainly for the ptlrpcd caller of this
1701                  * function.  Most ptlrpc sets are not long-lived and unbounded
1702                  * in length, but at the least the set used by the ptlrpcd is.
1703                  * Since the processing time is unbounded, we need to insert an
1704                  * explicit schedule point to make the thread well-behaved.
1705                  */
1706                 cond_resched();
1707
1708                 /* If the caller requires to allow to be interpreted by force
1709                  * and it has really been interpreted, then move the request
1710                  * to RQ_PHASE_INTERPRET phase in spite of what the current
1711                  * phase is. */
1712                 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1713                         req->rq_status = -EINTR;
1714                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1715
1716                         /* Since it is interpreted and we have to wait for
1717                          * the reply to be unlinked, then use sync mode. */
1718                         async = 0;
1719
1720                         GOTO(interpret, req->rq_status);
1721                 }
1722
1723                 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1724                         force_timer_recalc = 1;
1725
1726                 /* delayed send - skip */
1727                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1728                         continue;
1729
1730                 /* delayed resend - skip */
1731                 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1732                     req->rq_sent > cfs_time_current_sec())
1733                         continue;
1734
1735                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1736                       req->rq_phase == RQ_PHASE_BULK ||
1737                       req->rq_phase == RQ_PHASE_INTERPRET ||
1738                       req->rq_phase == RQ_PHASE_UNREG_RPC ||
1739                       req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1740                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1741                         LBUG();
1742                 }
1743
1744                 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1745                     req->rq_phase == RQ_PHASE_UNREG_BULK) {
1746                         LASSERT(req->rq_next_phase != req->rq_phase);
1747                         LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1748
1749                         if (req->rq_req_deadline &&
1750                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1751                                 req->rq_req_deadline = 0;
1752                         if (req->rq_reply_deadline &&
1753                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1754                                 req->rq_reply_deadline = 0;
1755                         if (req->rq_bulk_deadline &&
1756                             !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1757                                 req->rq_bulk_deadline = 0;
1758
1759                         /*
1760                          * Skip processing until reply is unlinked. We
1761                          * can't return to pool before that and we can't
1762                          * call interpret before that. We need to make
1763                          * sure that all rdma transfers finished and will
1764                          * not corrupt any data.
1765                          */
1766                         if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1767                             ptlrpc_client_recv_or_unlink(req))
1768                                 continue;
1769                         if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1770                             ptlrpc_client_bulk_active(req))
1771                                 continue;
1772
1773                         /*
1774                          * Turn fail_loc off to prevent it from looping
1775                          * forever.
1776                          */
1777                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1778                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1779                                                      OBD_FAIL_ONCE);
1780                         }
1781                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1782                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1783                                                      OBD_FAIL_ONCE);
1784                         }
1785
1786                         /*
1787                          * Move to next phase if reply was successfully
1788                          * unlinked.
1789                          */
1790                         ptlrpc_rqphase_move(req, req->rq_next_phase);
1791                 }
1792
1793                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1794                         GOTO(interpret, req->rq_status);
1795
1796                 /*
1797                  * Note that this also will start async reply unlink.
1798                  */
1799                 if (req->rq_net_err && !req->rq_timedout) {
1800                         ptlrpc_expire_one_request(req, 1);
1801
1802                         /*
1803                          * Check if we still need to wait for unlink.
1804                          */
1805                         if (ptlrpc_client_recv_or_unlink(req) ||
1806                             ptlrpc_client_bulk_active(req))
1807                                 continue;
1808                         /* If there is no need to resend, fail it now. */
1809                         if (req->rq_no_resend) {
1810                                 if (req->rq_status == 0)
1811                                         req->rq_status = -EIO;
1812                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1813                                 GOTO(interpret, req->rq_status);
1814                         } else {
1815                                 continue;
1816                         }
1817                 }
1818
1819                 if (req->rq_err) {
1820                         spin_lock(&req->rq_lock);
1821                         req->rq_replied = 0;
1822                         spin_unlock(&req->rq_lock);
1823                         if (req->rq_status == 0)
1824                                 req->rq_status = -EIO;
1825                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1826                         GOTO(interpret, req->rq_status);
1827                 }
1828
1829                 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1830                  * so it sets rq_intr regardless of individual rpc
1831                  * timeouts. The synchronous IO waiting path sets
1832                  * rq_intr irrespective of whether ptlrpcd
1833                  * has seen a timeout.  Our policy is to only interpret
1834                  * interrupted rpcs after they have timed out, so we
1835                  * need to enforce that here.
1836                  */
1837
1838                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1839                                      req->rq_wait_ctx)) {
1840                         req->rq_status = -EINTR;
1841                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1842                         GOTO(interpret, req->rq_status);
1843                 }
1844
1845                 if (req->rq_phase == RQ_PHASE_RPC) {
1846                         if (req->rq_timedout || req->rq_resend ||
1847                             req->rq_waiting || req->rq_wait_ctx) {
1848                                 int status;
1849
1850                                 if (!ptlrpc_unregister_reply(req, 1)) {
1851                                         ptlrpc_unregister_bulk(req, 1);
1852                                         continue;
1853                                 }
1854
1855                                 spin_lock(&imp->imp_lock);
1856                                 if (ptlrpc_import_delay_req(imp, req, &status)){
1857                                         /* put on delay list - only if we wait
1858                                          * recovery finished - before send */
1859                                         list_del_init(&req->rq_list);
1860                                         list_add_tail(&req->rq_list,
1861                                                           &imp->
1862                                                           imp_delayed_list);
1863                                         spin_unlock(&imp->imp_lock);
1864                                         continue;
1865                                 }
1866
1867                                 if (status != 0)  {
1868                                         req->rq_status = status;
1869                                         ptlrpc_rqphase_move(req,
1870                                                 RQ_PHASE_INTERPRET);
1871                                         spin_unlock(&imp->imp_lock);
1872                                         GOTO(interpret, req->rq_status);
1873                                 }
1874                                 if (ptlrpc_no_resend(req) &&
1875                                     !req->rq_wait_ctx) {
1876                                         req->rq_status = -ENOTCONN;
1877                                         ptlrpc_rqphase_move(req,
1878                                                             RQ_PHASE_INTERPRET);
1879                                         spin_unlock(&imp->imp_lock);
1880                                         GOTO(interpret, req->rq_status);
1881                                 }
1882
1883                                 list_del_init(&req->rq_list);
1884                                 list_add_tail(&req->rq_list,
1885                                                   &imp->imp_sending_list);
1886
1887                                 spin_unlock(&imp->imp_lock);
1888
1889                                 spin_lock(&req->rq_lock);
1890                                 req->rq_waiting = 0;
1891                                 spin_unlock(&req->rq_lock);
1892
1893                                 if (req->rq_timedout || req->rq_resend) {
1894                                         /* This is re-sending anyways,
1895                                          * let's mark req as resend. */
1896                                         spin_lock(&req->rq_lock);
1897                                         req->rq_resend = 1;
1898                                         spin_unlock(&req->rq_lock);
1899
1900                                         if (req->rq_bulk != NULL &&
1901                                             !ptlrpc_unregister_bulk(req, 1))
1902                                                 continue;
1903                                 }
1904                                 /*
1905                                  * rq_wait_ctx is only touched by ptlrpcd,
1906                                  * so no lock is needed here.
1907                                  */
1908                                 status = sptlrpc_req_refresh_ctx(req, -1);
1909                                 if (status) {
1910                                         if (req->rq_err) {
1911                                                 req->rq_status = status;
1912                                                 spin_lock(&req->rq_lock);
1913                                                 req->rq_wait_ctx = 0;
1914                                                 spin_unlock(&req->rq_lock);
1915                                                 force_timer_recalc = 1;
1916                                         } else {
1917                                                 spin_lock(&req->rq_lock);
1918                                                 req->rq_wait_ctx = 1;
1919                                                 spin_unlock(&req->rq_lock);
1920                                         }
1921
1922                                         continue;
1923                                 } else {
1924                                         spin_lock(&req->rq_lock);
1925                                         req->rq_wait_ctx = 0;
1926                                         spin_unlock(&req->rq_lock);
1927                                 }
1928
1929                                 rc = ptl_send_rpc(req, 0);
1930                                 if (rc == -ENOMEM) {
1931                                         spin_lock(&imp->imp_lock);
1932                                         if (!list_empty(&req->rq_list))
1933                                                 list_del_init(&req->rq_list);
1934                                         spin_unlock(&imp->imp_lock);
1935                                         ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1936                                         continue;
1937                                 }
1938                                 if (rc) {
1939                                         DEBUG_REQ(D_HA, req,
1940                                                   "send failed: rc = %d", rc);
1941                                         force_timer_recalc = 1;
1942                                         spin_lock(&req->rq_lock);
1943                                         req->rq_net_err = 1;
1944                                         spin_unlock(&req->rq_lock);
1945                                         continue;
1946                                 }
1947                                 /* need to reset the timeout */
1948                                 force_timer_recalc = 1;
1949                         }
1950
1951                         spin_lock(&req->rq_lock);
1952
1953                         if (ptlrpc_client_early(req)) {
1954                                 ptlrpc_at_recv_early_reply(req);
1955                                 spin_unlock(&req->rq_lock);
1956                                 continue;
1957                         }
1958
1959                         /* Still waiting for a reply? */
1960                         if (ptlrpc_client_recv(req)) {
1961                                 spin_unlock(&req->rq_lock);
1962                                 continue;
1963                         }
1964
1965                         /* Did we actually receive a reply? */
1966                         if (!ptlrpc_client_replied(req)) {
1967                                 spin_unlock(&req->rq_lock);
1968                                 continue;
1969                         }
1970
1971                         spin_unlock(&req->rq_lock);
1972
1973                         /* unlink from net because we are going to
1974                          * swab in-place of reply buffer */
1975                         unregistered = ptlrpc_unregister_reply(req, 1);
1976                         if (!unregistered)
1977                                 continue;
1978
1979                         req->rq_status = after_reply(req);
1980                         if (req->rq_resend)
1981                                 continue;
1982
1983                         /* If there is no bulk associated with this request,
1984                          * then we're done and should let the interpreter
1985                          * process the reply. Similarly if the RPC returned
1986                          * an error, and therefore the bulk will never arrive.
1987                          */
1988                         if (req->rq_bulk == NULL || req->rq_status < 0) {
1989                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1990                                 GOTO(interpret, req->rq_status);
1991                         }
1992
1993                         ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1994                 }
1995
1996                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1997                 if (ptlrpc_client_bulk_active(req))
1998                         continue;
1999
2000                 if (req->rq_bulk->bd_failure) {
2001                         /* The RPC reply arrived OK, but the bulk screwed
2002                          * up!  Dead weird since the server told us the RPC
2003                          * was good after getting the REPLY for her GET or
2004                          * the ACK for her PUT. */
2005                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2006                         req->rq_status = -EIO;
2007                 }
2008
2009                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2010
2011         interpret:
2012                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2013
2014                 /* This moves to "unregistering" phase we need to wait for
2015                  * reply unlink. */
2016                 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2017                         /* start async bulk unlink too */
2018                         ptlrpc_unregister_bulk(req, 1);
2019                         continue;
2020                 }
2021
2022                 if (!ptlrpc_unregister_bulk(req, async))
2023                         continue;
2024
2025                 /* When calling interpret receiving already should be
2026                  * finished. */
2027                 LASSERT(!req->rq_receiving_reply);
2028
2029                 ptlrpc_req_interpret(env, req, req->rq_status);
2030
2031                 if (ptlrpcd_check_work(req)) {
2032                         atomic_dec(&set->set_remaining);
2033                         continue;
2034                 }
2035                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2036
2037                 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
2038                         "Completed RPC pname:cluuid:pid:xid:nid:"
2039                         "opc %s:%s:%d:%llu:%s:%d\n",
2040                         current_comm(), imp->imp_obd->obd_uuid.uuid,
2041                         lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2042                         libcfs_nid2str(imp->imp_connection->c_peer.nid),
2043                         lustre_msg_get_opc(req->rq_reqmsg));
2044
2045                 spin_lock(&imp->imp_lock);
2046                 /* Request already may be not on sending or delaying list. This
2047                  * may happen in the case of marking it erroneous for the case
2048                  * ptlrpc_import_delay_req(req, status) find it impossible to
2049                  * allow sending this rpc and returns *status != 0. */
2050                 if (!list_empty(&req->rq_list)) {
2051                         list_del_init(&req->rq_list);
2052                         atomic_dec(&imp->imp_inflight);
2053                 }
2054                 list_del_init(&req->rq_unreplied_list);
2055                 spin_unlock(&imp->imp_lock);
2056
2057                 atomic_dec(&set->set_remaining);
2058                 wake_up_all(&imp->imp_recovery_waitq);
2059
2060                 if (set->set_producer) {
2061                         /* produce a new request if possible */
2062                         if (ptlrpc_set_producer(set) > 0)
2063                                 force_timer_recalc = 1;
2064
2065                         /* free the request that has just been completed
2066                          * in order not to pollute set->set_requests */
2067                         list_del_init(&req->rq_set_chain);
2068                         spin_lock(&req->rq_lock);
2069                         req->rq_set = NULL;
2070                         req->rq_invalid_rqset = 0;
2071                         spin_unlock(&req->rq_lock);
2072
2073                         /* record rq_status to compute the final status later */
2074                         if (req->rq_status != 0)
2075                                 set->set_rc = req->rq_status;
2076                         ptlrpc_req_finished(req);
2077                 } else {
2078                         list_move_tail(&req->rq_set_chain, &comp_reqs);
2079                 }
2080         }
2081
2082         /* move completed request at the head of list so it's easier for
2083          * caller to find them */
2084         list_splice(&comp_reqs, &set->set_requests);
2085
2086         /* If we hit an error, we want to recover promptly. */
2087         RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2088 }
2089 EXPORT_SYMBOL(ptlrpc_check_set);
2090
2091 /**
2092  * Time out request \a req. is \a async_unlink is set, that means do not wait
2093  * until LNet actually confirms network buffer unlinking.
2094  * Return 1 if we should give up further retrying attempts or 0 otherwise.
2095  */
2096 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2097 {
2098         struct obd_import *imp = req->rq_import;
2099         int rc = 0;
2100         ENTRY;
2101
2102         spin_lock(&req->rq_lock);
2103         req->rq_timedout = 1;
2104         spin_unlock(&req->rq_lock);
2105
2106         DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
2107                   "/real "CFS_DURATION_T"]",
2108                   req->rq_net_err ? "failed due to network error" :
2109                      ((req->rq_real_sent == 0 ||
2110                        cfs_time_before(req->rq_real_sent, req->rq_sent) ||
2111                        cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
2112                       "timed out for sent delay" : "timed out for slow reply"),
2113                   req->rq_sent, 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         time_t                          now = cfs_time_current_sec();
2172         ENTRY;
2173
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         time_t                   now = cfs_time_current_sec();
2257         int                      timeout = 0;
2258         struct ptlrpc_request   *req;
2259         int                      deadline;
2260         ENTRY;
2261
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 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2510 {
2511         ENTRY;
2512         if (request == NULL)
2513                 RETURN(1);
2514
2515         if (request == LP_POISON ||
2516             request->rq_reqmsg == LP_POISON) {
2517                 CERROR("dereferencing freed request (bug 575)\n");
2518                 LBUG();
2519                 RETURN(1);
2520         }
2521
2522         DEBUG_REQ(D_INFO, request, "refcount now %u",
2523                   atomic_read(&request->rq_refcount) - 1);
2524
2525         if (atomic_dec_and_test(&request->rq_refcount)) {
2526                 __ptlrpc_free_req(request, locked);
2527                 RETURN(1);
2528         }
2529
2530         RETURN(0);
2531 }
2532
2533 /**
2534  * Drops one reference count for a request.
2535  */
2536 void ptlrpc_req_finished(struct ptlrpc_request *request)
2537 {
2538         __ptlrpc_req_finished(request, 0);
2539 }
2540 EXPORT_SYMBOL(ptlrpc_req_finished);
2541
2542 /**
2543  * Returns xid of a \a request
2544  */
2545 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2546 {
2547         return request->rq_xid;
2548 }
2549 EXPORT_SYMBOL(ptlrpc_req_xid);
2550
2551 /**
2552  * Disengage the client's reply buffer from the network
2553  * NB does _NOT_ unregister any client-side bulk.
2554  * IDEMPOTENT, but _not_ safe against concurrent callers.
2555  * The request owner (i.e. the thread doing the I/O) must call...
2556  * Returns 0 on success or 1 if unregistering cannot be made.
2557  */
2558 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2559 {
2560         int                rc;
2561         struct l_wait_info lwi;
2562
2563         /*
2564          * Might sleep.
2565          */
2566         LASSERT(!in_interrupt());
2567
2568         /* Let's setup deadline for reply unlink. */
2569         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2570             async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2571                 request->rq_reply_deadline =
2572                         cfs_time_current_sec() + LONG_UNLINK;
2573
2574         /*
2575          * Nothing left to do.
2576          */
2577         if (!ptlrpc_client_recv_or_unlink(request))
2578                 RETURN(1);
2579
2580         LNetMDUnlink(request->rq_reply_md_h);
2581
2582         /*
2583          * Let's check it once again.
2584          */
2585         if (!ptlrpc_client_recv_or_unlink(request))
2586                 RETURN(1);
2587
2588         /* Move to "Unregistering" phase as reply was not unlinked yet. */
2589         ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2590
2591         /*
2592          * Do not wait for unlink to finish.
2593          */
2594         if (async)
2595                 RETURN(0);
2596
2597         /*
2598          * We have to l_wait_event() whatever the result, to give liblustre
2599          * a chance to run reply_in_callback(), and to make sure we've
2600          * unlinked before returning a req to the pool.
2601          */
2602         for (;;) {
2603                 /* The wq argument is ignored by user-space wait_event macros */
2604                 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2605                                         &request->rq_set->set_waitq :
2606                                         &request->rq_reply_waitq;
2607                 /* Network access will complete in finite time but the HUGE
2608                  * timeout lets us CWARN for visibility of sluggish NALs */
2609                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2610                                            cfs_time_seconds(1), NULL, NULL);
2611                 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2612                                   &lwi);
2613                 if (rc == 0) {
2614                         ptlrpc_rqphase_move(request, request->rq_next_phase);
2615                         RETURN(1);
2616                 }
2617
2618                 LASSERT(rc == -ETIMEDOUT);
2619                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2620                           "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2621                           request->rq_receiving_reply,
2622                           request->rq_req_unlinked,
2623                           request->rq_reply_unlinked);
2624         }
2625         RETURN(0);
2626 }
2627
2628 static void ptlrpc_free_request(struct ptlrpc_request *req)
2629 {
2630         spin_lock(&req->rq_lock);
2631         req->rq_replay = 0;
2632         spin_unlock(&req->rq_lock);
2633
2634         if (req->rq_commit_cb != NULL)
2635                 req->rq_commit_cb(req);
2636         list_del_init(&req->rq_replay_list);
2637
2638         __ptlrpc_req_finished(req, 1);
2639 }
2640
2641 /**
2642  * the request is committed and dropped from the replay list of its import
2643  */
2644 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2645 {
2646         struct obd_import       *imp = req->rq_import;
2647
2648         spin_lock(&imp->imp_lock);
2649         if (list_empty(&req->rq_replay_list)) {
2650                 spin_unlock(&imp->imp_lock);
2651                 return;
2652         }
2653
2654         if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2655                 ptlrpc_free_request(req);
2656
2657         spin_unlock(&imp->imp_lock);
2658 }
2659 EXPORT_SYMBOL(ptlrpc_request_committed);
2660
2661 /**
2662  * Iterates through replay_list on import and prunes
2663  * all requests have transno smaller than last_committed for the
2664  * import and don't have rq_replay set.
2665  * Since requests are sorted in transno order, stops when meetign first
2666  * transno bigger than last_committed.
2667  * caller must hold imp->imp_lock
2668  */
2669 void ptlrpc_free_committed(struct obd_import *imp)
2670 {
2671         struct ptlrpc_request   *req, *saved;
2672         struct ptlrpc_request   *last_req = NULL; /* temporary fire escape */
2673         bool                     skip_committed_list = true;
2674         ENTRY;
2675
2676         LASSERT(imp != NULL);
2677         assert_spin_locked(&imp->imp_lock);
2678
2679         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2680             imp->imp_generation == imp->imp_last_generation_checked) {
2681                 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2682                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2683                 RETURN_EXIT;
2684         }
2685         CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2686                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2687                imp->imp_generation);
2688
2689         if (imp->imp_generation != imp->imp_last_generation_checked ||
2690             imp->imp_last_transno_checked == 0)
2691                 skip_committed_list = false;
2692
2693         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2694         imp->imp_last_generation_checked = imp->imp_generation;
2695
2696         list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2697                                      rq_replay_list) {
2698                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
2699                 LASSERT(req != last_req);
2700                 last_req = req;
2701
2702                 if (req->rq_transno == 0) {
2703                         DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2704                         LBUG();
2705                 }
2706                 if (req->rq_import_generation < imp->imp_generation) {
2707                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2708                         GOTO(free_req, 0);
2709                 }
2710
2711                 /* not yet committed */
2712                 if (req->rq_transno > imp->imp_peer_committed_transno) {
2713                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2714                         break;
2715                 }
2716
2717                 if (req->rq_replay) {
2718                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2719                         list_move_tail(&req->rq_replay_list,
2720                                            &imp->imp_committed_list);
2721                         continue;
2722                 }
2723
2724                 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2725                           imp->imp_peer_committed_transno);
2726 free_req:
2727                 ptlrpc_free_request(req);
2728         }
2729
2730         if (skip_committed_list)
2731                 GOTO(out, 0);
2732
2733         list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2734                                      rq_replay_list) {
2735                 LASSERT(req->rq_transno != 0);
2736                 if (req->rq_import_generation < imp->imp_generation) {
2737                         DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2738                         ptlrpc_free_request(req);
2739                 } else if (!req->rq_replay) {
2740                         DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2741                         ptlrpc_free_request(req);
2742                 }
2743         }
2744 out:
2745         EXIT;
2746 }
2747
2748 void ptlrpc_cleanup_client(struct obd_import *imp)
2749 {
2750         ENTRY;
2751         EXIT;
2752 }
2753
2754 /**
2755  * Schedule previously sent request for resend.
2756  * For bulk requests we assign new xid (to avoid problems with
2757  * lost replies and therefore several transfers landing into same buffer
2758  * from different sending attempts).
2759  */
2760 void ptlrpc_resend_req(struct ptlrpc_request *req)
2761 {
2762         DEBUG_REQ(D_HA, req, "going to resend");
2763         spin_lock(&req->rq_lock);
2764
2765         /* Request got reply but linked to the import list still.
2766            Let ptlrpc_check_set() to process it. */
2767         if (ptlrpc_client_replied(req)) {
2768                 spin_unlock(&req->rq_lock);
2769                 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2770                 return;
2771         }
2772
2773         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2774         req->rq_status = -EAGAIN;
2775
2776         req->rq_resend = 1;
2777         req->rq_net_err = 0;
2778         req->rq_timedout = 0;
2779
2780         ptlrpc_client_wake_req(req);
2781         spin_unlock(&req->rq_lock);
2782 }
2783
2784 /* XXX: this function and rq_status are currently unused */
2785 void ptlrpc_restart_req(struct ptlrpc_request *req)
2786 {
2787         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2788         req->rq_status = -ERESTARTSYS;
2789
2790         spin_lock(&req->rq_lock);
2791         req->rq_restart = 1;
2792         req->rq_timedout = 0;
2793         ptlrpc_client_wake_req(req);
2794         spin_unlock(&req->rq_lock);
2795 }
2796
2797 /**
2798  * Grab additional reference on a request \a req
2799  */
2800 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2801 {
2802         ENTRY;
2803         atomic_inc(&req->rq_refcount);
2804         RETURN(req);
2805 }
2806 EXPORT_SYMBOL(ptlrpc_request_addref);
2807
2808 /**
2809  * Add a request to import replay_list.
2810  * Must be called under imp_lock
2811  */
2812 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2813                                       struct obd_import *imp)
2814 {
2815         struct list_head *tmp;
2816
2817         assert_spin_locked(&imp->imp_lock);
2818
2819         if (req->rq_transno == 0) {
2820                 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2821                 LBUG();
2822         }
2823
2824         /* clear this for new requests that were resent as well
2825            as resent replayed requests. */
2826         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2827
2828         /* don't re-add requests that have been replayed */
2829         if (!list_empty(&req->rq_replay_list))
2830                 return;
2831
2832         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2833
2834         spin_lock(&req->rq_lock);
2835         req->rq_resend = 0;
2836         spin_unlock(&req->rq_lock);
2837
2838         LASSERT(imp->imp_replayable);
2839         /* Balanced in ptlrpc_free_committed, usually. */
2840         ptlrpc_request_addref(req);
2841         list_for_each_prev(tmp, &imp->imp_replay_list) {
2842                 struct ptlrpc_request *iter = list_entry(tmp,
2843                                                          struct ptlrpc_request,
2844                                                          rq_replay_list);
2845
2846                 /* We may have duplicate transnos if we create and then
2847                  * open a file, or for closes retained if to match creating
2848                  * opens, so use req->rq_xid as a secondary key.
2849                  * (See bugs 684, 685, and 428.)
2850                  * XXX no longer needed, but all opens need transnos!
2851                  */
2852                 if (iter->rq_transno > req->rq_transno)
2853                         continue;
2854
2855                 if (iter->rq_transno == req->rq_transno) {
2856                         LASSERT(iter->rq_xid != req->rq_xid);
2857                         if (iter->rq_xid > req->rq_xid)
2858                                 continue;
2859                 }
2860
2861                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2862                 return;
2863         }
2864
2865         list_add(&req->rq_replay_list, &imp->imp_replay_list);
2866 }
2867
2868 /**
2869  * Send request and wait until it completes.
2870  * Returns request processing status.
2871  */
2872 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2873 {
2874         struct ptlrpc_request_set *set;
2875         int rc;
2876         ENTRY;
2877
2878         LASSERT(req->rq_set == NULL);
2879         LASSERT(!req->rq_receiving_reply);
2880
2881         set = ptlrpc_prep_set();
2882         if (set == NULL) {
2883                 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2884                 RETURN(-ENOMEM);
2885         }
2886
2887         /* for distributed debugging */
2888         lustre_msg_set_status(req->rq_reqmsg, current_pid());
2889
2890         /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2891         ptlrpc_request_addref(req);
2892         ptlrpc_set_add_req(set, req);
2893         rc = ptlrpc_set_wait(set);
2894         ptlrpc_set_destroy(set);
2895
2896         RETURN(rc);
2897 }
2898 EXPORT_SYMBOL(ptlrpc_queue_wait);
2899
2900 /**
2901  * Callback used for replayed requests reply processing.
2902  * In case of successful reply calls registered request replay callback.
2903  * In case of error restart replay process.
2904  */
2905 static int ptlrpc_replay_interpret(const struct lu_env *env,
2906                                    struct ptlrpc_request *req,
2907                                    void * data, int rc)
2908 {
2909         struct ptlrpc_replay_async_args *aa = data;
2910         struct obd_import *imp = req->rq_import;
2911
2912         ENTRY;
2913         atomic_dec(&imp->imp_replay_inflight);
2914
2915         /* Note: if it is bulk replay (MDS-MDS replay), then even if
2916          * server got the request, but bulk transfer timeout, let's
2917          * replay the bulk req again */
2918         if (!ptlrpc_client_replied(req) ||
2919             (req->rq_bulk != NULL &&
2920              lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2921                 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2922                 GOTO(out, rc = -ETIMEDOUT);
2923         }
2924
2925         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2926             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2927              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2928                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2929
2930         /** VBR: check version failure */
2931         if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2932                 /** replay was failed due to version mismatch */
2933                 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2934                 spin_lock(&imp->imp_lock);
2935                 imp->imp_vbr_failed = 1;
2936                 imp->imp_no_lock_replay = 1;
2937                 spin_unlock(&imp->imp_lock);
2938                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2939         } else {
2940                 /** The transno had better not change over replay. */
2941                 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2942                          lustre_msg_get_transno(req->rq_repmsg) ||
2943                          lustre_msg_get_transno(req->rq_repmsg) == 0,
2944                          "%#llx/%#llx\n",
2945                          lustre_msg_get_transno(req->rq_reqmsg),
2946                          lustre_msg_get_transno(req->rq_repmsg));
2947         }
2948
2949         spin_lock(&imp->imp_lock);
2950         /** if replays by version then gap occur on server, no trust to locks */
2951         if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2952                 imp->imp_no_lock_replay = 1;
2953         imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2954         spin_unlock(&imp->imp_lock);
2955         LASSERT(imp->imp_last_replay_transno);
2956
2957         /* transaction number shouldn't be bigger than the latest replayed */
2958         if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2959                 DEBUG_REQ(D_ERROR, req,
2960                           "Reported transno %llu is bigger than the "
2961                           "replayed one: %llu", req->rq_transno,
2962                           lustre_msg_get_transno(req->rq_reqmsg));
2963                 GOTO(out, rc = -EINVAL);
2964         }
2965
2966         DEBUG_REQ(D_HA, req, "got rep");
2967
2968         /* let the callback do fixups, possibly including in the request */
2969         if (req->rq_replay_cb)
2970                 req->rq_replay_cb(req);
2971
2972         if (ptlrpc_client_replied(req) &&
2973             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2974                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2975                           lustre_msg_get_status(req->rq_repmsg),
2976                           aa->praa_old_status);
2977
2978                 /* Note: If the replay fails for MDT-MDT recovery, let's
2979                  * abort all of the following requests in the replay
2980                  * and sending list, because MDT-MDT update requests
2981                  * are dependent on each other, see LU-7039 */
2982                 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
2983                         struct ptlrpc_request *free_req;
2984                         struct ptlrpc_request *tmp;
2985
2986                         spin_lock(&imp->imp_lock);
2987                         list_for_each_entry_safe(free_req, tmp,
2988                                                  &imp->imp_replay_list,
2989                                                  rq_replay_list) {
2990                                 ptlrpc_free_request(free_req);
2991                         }
2992
2993                         list_for_each_entry_safe(free_req, tmp,
2994                                                  &imp->imp_committed_list,
2995                                                  rq_replay_list) {
2996                                 ptlrpc_free_request(free_req);
2997                         }
2998
2999                         list_for_each_entry_safe(free_req, tmp,
3000                                                 &imp->imp_delayed_list,
3001                                                 rq_list) {
3002                                 spin_lock(&free_req->rq_lock);
3003                                 free_req->rq_err = 1;
3004                                 free_req->rq_status = -EIO;
3005                                 ptlrpc_client_wake_req(free_req);
3006                                 spin_unlock(&free_req->rq_lock);
3007                         }
3008
3009                         list_for_each_entry_safe(free_req, tmp,
3010                                                 &imp->imp_sending_list,
3011                                                 rq_list) {
3012                                 spin_lock(&free_req->rq_lock);
3013                                 free_req->rq_err = 1;
3014                                 free_req->rq_status = -EIO;
3015                                 ptlrpc_client_wake_req(free_req);
3016                                 spin_unlock(&free_req->rq_lock);
3017                         }
3018                         spin_unlock(&imp->imp_lock);
3019                 }
3020         } else {
3021                 /* Put it back for re-replay. */
3022                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3023         }
3024
3025         /*
3026          * Errors while replay can set transno to 0, but
3027          * imp_last_replay_transno shouldn't be set to 0 anyway
3028          */
3029         if (req->rq_transno == 0)
3030                 CERROR("Transno is 0 during replay!\n");
3031
3032         /* continue with recovery */
3033         rc = ptlrpc_import_recovery_state_machine(imp);
3034  out:
3035         req->rq_send_state = aa->praa_old_state;
3036
3037         if (rc != 0)
3038                 /* this replay failed, so restart recovery */
3039                 ptlrpc_connect_import(imp);
3040
3041         RETURN(rc);
3042 }
3043
3044 /**
3045  * Prepares and queues request for replay.
3046  * Adds it to ptlrpcd queue for actual sending.
3047  * Returns 0 on success.
3048  */
3049 int ptlrpc_replay_req(struct ptlrpc_request *req)
3050 {
3051         struct ptlrpc_replay_async_args *aa;
3052         ENTRY;
3053
3054         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3055
3056         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
3057         aa = ptlrpc_req_async_args(req);
3058         memset(aa, 0, sizeof *aa);
3059
3060         /* Prepare request to be resent with ptlrpcd */
3061         aa->praa_old_state = req->rq_send_state;
3062         req->rq_send_state = LUSTRE_IMP_REPLAY;
3063         req->rq_phase = RQ_PHASE_NEW;
3064         req->rq_next_phase = RQ_PHASE_UNDEFINED;
3065         if (req->rq_repmsg)
3066                 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3067         req->rq_status = 0;
3068         req->rq_interpret_reply = ptlrpc_replay_interpret;
3069         /* Readjust the timeout for current conditions */
3070         ptlrpc_at_set_req_timeout(req);
3071
3072         /* Tell server the net_latency, so the server can calculate how long
3073          * it should wait for next replay */
3074         lustre_msg_set_service_time(req->rq_reqmsg,
3075                                     ptlrpc_at_get_net_latency(req));
3076         DEBUG_REQ(D_HA, req, "REPLAY");
3077
3078         atomic_inc(&req->rq_import->imp_replay_inflight);
3079         ptlrpc_request_addref(req);     /* ptlrpcd needs a ref */
3080
3081         ptlrpcd_add_req(req);
3082         RETURN(0);
3083 }
3084
3085 /**
3086  * Aborts all in-flight request on import \a imp sending and delayed lists
3087  */
3088 void ptlrpc_abort_inflight(struct obd_import *imp)
3089 {
3090         struct list_head *tmp, *n;
3091         ENTRY;
3092
3093         /* Make sure that no new requests get processed for this import.
3094          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3095          * this flag and then putting requests on sending_list or delayed_list.
3096          */
3097         spin_lock(&imp->imp_lock);
3098
3099         /* XXX locking?  Maybe we should remove each request with the list
3100          * locked?  Also, how do we know if the requests on the list are
3101          * being freed at this time?
3102          */
3103         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3104                 struct ptlrpc_request *req = list_entry(tmp,
3105                                                         struct ptlrpc_request,
3106                                                         rq_list);
3107
3108                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3109
3110                 spin_lock(&req->rq_lock);
3111                 if (req->rq_import_generation < imp->imp_generation) {
3112                         req->rq_err = 1;
3113                         req->rq_status = -EIO;
3114                         ptlrpc_client_wake_req(req);
3115                 }
3116                 spin_unlock(&req->rq_lock);
3117         }
3118
3119         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3120                 struct ptlrpc_request *req =
3121                         list_entry(tmp, struct ptlrpc_request, rq_list);
3122
3123                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3124
3125                 spin_lock(&req->rq_lock);
3126                 if (req->rq_import_generation < imp->imp_generation) {
3127                         req->rq_err = 1;
3128                         req->rq_status = -EIO;
3129                         ptlrpc_client_wake_req(req);
3130                 }
3131                 spin_unlock(&req->rq_lock);
3132         }
3133
3134         /* Last chance to free reqs left on the replay list, but we
3135          * will still leak reqs that haven't committed.  */
3136         if (imp->imp_replayable)
3137                 ptlrpc_free_committed(imp);
3138
3139         spin_unlock(&imp->imp_lock);
3140
3141         EXIT;
3142 }
3143
3144 /**
3145  * Abort all uncompleted requests in request set \a set
3146  */
3147 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3148 {
3149         struct list_head *tmp, *pos;
3150
3151         LASSERT(set != NULL);
3152
3153         list_for_each_safe(pos, tmp, &set->set_requests) {
3154                 struct ptlrpc_request *req =
3155                         list_entry(pos, struct ptlrpc_request,
3156                                    rq_set_chain);
3157
3158                 spin_lock(&req->rq_lock);
3159                 if (req->rq_phase != RQ_PHASE_RPC) {
3160                         spin_unlock(&req->rq_lock);
3161                         continue;
3162                 }
3163
3164                 req->rq_err = 1;
3165                 req->rq_status = -EINTR;
3166                 ptlrpc_client_wake_req(req);
3167                 spin_unlock(&req->rq_lock);
3168         }
3169 }
3170
3171 static __u64 ptlrpc_last_xid;
3172 static spinlock_t ptlrpc_last_xid_lock;
3173
3174 /**
3175  * Initialize the XID for the node.  This is common among all requests on
3176  * this node, and only requires the property that it is monotonically
3177  * increasing.  It does not need to be sequential.  Since this is also used
3178  * as the RDMA match bits, it is important that a single client NOT have
3179  * the same match bits for two different in-flight requests, hence we do
3180  * NOT want to have an XID per target or similar.
3181  *
3182  * To avoid an unlikely collision between match bits after a client reboot
3183  * (which would deliver old data into the wrong RDMA buffer) initialize
3184  * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3185  * If the time is clearly incorrect, we instead use a 62-bit random number.
3186  * In the worst case the random number will overflow 1M RPCs per second in
3187  * 9133 years, or permutations thereof.
3188  */
3189 #define YEAR_2004 (1ULL << 30)
3190 void ptlrpc_init_xid(void)
3191 {
3192         time_t now = cfs_time_current_sec();
3193
3194         spin_lock_init(&ptlrpc_last_xid_lock);
3195         if (now < YEAR_2004) {
3196                 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3197                 ptlrpc_last_xid >>= 2;
3198                 ptlrpc_last_xid |= (1ULL << 61);
3199         } else {
3200                 ptlrpc_last_xid = (__u64)now << 20;
3201         }
3202
3203         /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3204         CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3205         ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3206 }
3207
3208 /**
3209  * Increase xid and returns resulting new value to the caller.
3210  *
3211  * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3212  * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3213  * itself uses the last bulk xid needed, so the server can determine the
3214  * the number of bulk transfers from the RPC XID and a bitmask.  The starting
3215  * xid must align to a power-of-two value.
3216  *
3217  * This is assumed to be true due to the initial