Whamcloud - gitweb
LU-5710 all: second batch of corrected typos and grammar errors
[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, 2014, 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 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
51 static int ptlrpcd_check_work(struct ptlrpc_request *req);
52 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
53
54 /**
55  * Initialize passed in client structure \a cl.
56  */
57 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
58                         struct ptlrpc_client *cl)
59 {
60         cl->cli_request_portal = req_portal;
61         cl->cli_reply_portal   = rep_portal;
62         cl->cli_name           = name;
63 }
64 EXPORT_SYMBOL(ptlrpc_init_client);
65
66 /**
67  * Return PortalRPC connection for remore uud \a uuid
68  */
69 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
70 {
71         struct ptlrpc_connection *c;
72         lnet_nid_t                self;
73         lnet_process_id_t         peer;
74         int                       err;
75
76         /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
77          * before accessing its values. */
78         /* coverity[uninit_use_in_call] */
79         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
80         if (err != 0) {
81                 CNETERR("cannot find peer %s!\n", uuid->uuid);
82                 return NULL;
83         }
84
85         c = ptlrpc_connection_get(peer, self, uuid);
86         if (c) {
87                 memcpy(c->c_remote_uuid.uuid,
88                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
89         }
90
91         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
92
93         return c;
94 }
95
96 /**
97  * Allocate and initialize new bulk descriptor on the sender.
98  * Returns pointer to the descriptor or NULL on error.
99  */
100 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned npages, unsigned max_brw,
101                                          unsigned type, unsigned portal)
102 {
103         struct ptlrpc_bulk_desc *desc;
104         int i;
105
106         OBD_ALLOC(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[npages]));
107         if (!desc)
108                 return NULL;
109
110         spin_lock_init(&desc->bd_lock);
111         init_waitqueue_head(&desc->bd_waitq);
112         desc->bd_max_iov = npages;
113         desc->bd_iov_count = 0;
114         desc->bd_portal = portal;
115         desc->bd_type = type;
116         desc->bd_md_count = 0;
117         LASSERT(max_brw > 0);
118         desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
119         /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
120          * node. Negotiated ocd_brw_size will always be <= this number. */
121         for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
122                 LNetInvalidateHandle(&desc->bd_mds[i]);
123
124         return desc;
125 }
126
127 /**
128  * Prepare bulk descriptor for specified outgoing request \a req that
129  * can fit \a npages * pages. \a type is bulk type. \a portal is where
130  * the bulk to be sent. Used on client-side.
131  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
132  * error.
133  */
134 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
135                                               unsigned npages, unsigned max_brw,
136                                               unsigned type, unsigned portal)
137 {
138         struct obd_import *imp = req->rq_import;
139         struct ptlrpc_bulk_desc *desc;
140
141         ENTRY;
142         LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
143         desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
144         if (desc == NULL)
145                 RETURN(NULL);
146
147         desc->bd_import_generation = req->rq_import_generation;
148         desc->bd_import = class_import_get(imp);
149         desc->bd_req = req;
150
151         desc->bd_cbid.cbid_fn  = client_bulk_callback;
152         desc->bd_cbid.cbid_arg = desc;
153
154         /* This makes req own desc, and free it when she frees herself */
155         req->rq_bulk = desc;
156
157         return desc;
158 }
159 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
160
161 /*
162  * Add a page \a page to the bulk descriptor \a desc.
163  * Data to transfer in the page starts at offset \a pageoffset and
164  * amount of data to transfer from the page is \a len
165  */
166 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
167                              struct page *page, int pageoffset, int len, int pin)
168 {
169         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
170         LASSERT(page != NULL);
171         LASSERT(pageoffset >= 0);
172         LASSERT(len > 0);
173         LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
174
175         desc->bd_nob += len;
176
177         if (pin)
178                 page_cache_get(page);
179
180         ptlrpc_add_bulk_page(desc, page, pageoffset, len);
181 }
182 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
183
184 /**
185  * Uninitialize and free bulk descriptor \a desc.
186  * Works on bulk descriptors both from server and client side.
187  */
188 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin)
189 {
190         int i;
191         ENTRY;
192
193         LASSERT(desc != NULL);
194         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
195         LASSERT(desc->bd_md_count == 0);         /* network hands off */
196         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
197
198         sptlrpc_enc_pool_put_pages(desc);
199
200         if (desc->bd_export)
201                 class_export_put(desc->bd_export);
202         else
203                 class_import_put(desc->bd_import);
204
205         if (unpin) {
206                 for (i = 0; i < desc->bd_iov_count ; i++)
207                         page_cache_release(desc->bd_iov[i].kiov_page);
208         }
209
210         OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
211                                 bd_iov[desc->bd_max_iov]));
212         EXIT;
213 }
214 EXPORT_SYMBOL(__ptlrpc_free_bulk);
215
216 /**
217  * Set server timelimit for this req, i.e. how long are we willing to wait
218  * for reply before timing out this request.
219  */
220 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
221 {
222         __u32 serv_est;
223         int idx;
224         struct imp_at *at;
225
226         LASSERT(req->rq_import);
227
228         if (AT_OFF) {
229                 /* non-AT settings */
230                 /**
231                  * \a imp_server_timeout means this is reverse import and
232                  * we send (currently only) ASTs to the client and cannot afford
233                  * to wait too long for the reply, otherwise the other client
234                  * (because of which we are sending this request) would
235                  * timeout waiting for us
236                  */
237                 req->rq_timeout = req->rq_import->imp_server_timeout ?
238                                   obd_timeout / 2 : obd_timeout;
239         } else {
240                 at = &req->rq_import->imp_at;
241                 idx = import_at_get_index(req->rq_import,
242                                           req->rq_request_portal);
243                 serv_est = at_get(&at->iat_service_estimate[idx]);
244                 req->rq_timeout = at_est2timeout(serv_est);
245         }
246         /* We could get even fancier here, using history to predict increased
247            loading... */
248
249         /* Let the server know what this RPC timeout is by putting it in the
250            reqmsg*/
251         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
252 }
253 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
254
255 /* Adjust max service estimate based on server value */
256 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
257                                   unsigned int serv_est)
258 {
259         int idx;
260         unsigned int oldse;
261         struct imp_at *at;
262
263         LASSERT(req->rq_import);
264         at = &req->rq_import->imp_at;
265
266         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
267         /* max service estimates are tracked on the server side,
268            so just keep minimal history here */
269         oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
270         if (oldse != 0)
271                 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
272                        "has changed from %d to %d\n",
273                        req->rq_import->imp_obd->obd_name,req->rq_request_portal,
274                        oldse, at_get(&at->iat_service_estimate[idx]));
275 }
276
277 /* Expected network latency per remote node (secs) */
278 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
279 {
280         return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
281 }
282
283 /* Adjust expected network latency */
284 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
285                                unsigned int service_time)
286 {
287         unsigned int nl, oldnl;
288         struct imp_at *at;
289         time_t now = cfs_time_current_sec();
290
291         LASSERT(req->rq_import);
292
293         if (service_time > now - req->rq_sent + 3) {
294                 /* bz16408, however, this can also happen if early reply
295                  * is lost and client RPC is expired and resent, early reply
296                  * or reply of original RPC can still be fit in reply buffer
297                  * of resent RPC, now client is measuring time from the
298                  * resent time, but server sent back service time of original
299                  * RPC.
300                  */
301                 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
302                        D_ADAPTTO : D_WARNING,
303                        "Reported service time %u > total measured time "
304                        CFS_DURATION_T"\n", service_time,
305                        cfs_time_sub(now, req->rq_sent));
306                 return;
307         }
308
309         /* Network latency is total time less server processing time */
310         nl = max_t(int, now - req->rq_sent -
311                         service_time, 0) + 1; /* st rounding */
312         at = &req->rq_import->imp_at;
313
314         oldnl = at_measured(&at->iat_net_latency, nl);
315         if (oldnl != 0)
316                 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
317                        "has changed from %d to %d\n",
318                        req->rq_import->imp_obd->obd_name,
319                        obd_uuid2str(
320                                &req->rq_import->imp_connection->c_remote_uuid),
321                        oldnl, at_get(&at->iat_net_latency));
322 }
323
324 static int unpack_reply(struct ptlrpc_request *req)
325 {
326         int rc;
327
328         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
329                 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
330                 if (rc) {
331                         DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
332                         return(-EPROTO);
333                 }
334         }
335
336         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
337         if (rc) {
338                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
339                 return(-EPROTO);
340         }
341         return 0;
342 }
343
344 /**
345  * Handle an early reply message, called with the rq_lock held.
346  * If anything goes wrong just ignore it - same as if it never happened
347  */
348 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
349 __must_hold(&req->rq_lock)
350 {
351         struct ptlrpc_request *early_req;
352         time_t                 olddl;
353         int                    rc;
354         ENTRY;
355
356         req->rq_early = 0;
357         spin_unlock(&req->rq_lock);
358
359         rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
360         if (rc) {
361                 spin_lock(&req->rq_lock);
362                 RETURN(rc);
363         }
364
365         rc = unpack_reply(early_req);
366         if (rc != 0) {
367                 sptlrpc_cli_finish_early_reply(early_req);
368                 spin_lock(&req->rq_lock);
369                 RETURN(rc);
370         }
371
372         /* Use new timeout value just to adjust the local value for this
373          * request, don't include it into at_history. It is unclear yet why
374          * service time increased and should it be counted or skipped, e.g.
375          * that can be recovery case or some error or server, the real reply
376          * will add all new data if it is worth to add. */
377         req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
378         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
379
380         /* Network latency can be adjusted, it is pure network delays */
381         ptlrpc_at_adj_net_latency(req,
382                         lustre_msg_get_service_time(early_req->rq_repmsg));
383
384         sptlrpc_cli_finish_early_reply(early_req);
385
386         spin_lock(&req->rq_lock);
387         olddl = req->rq_deadline;
388         /* server assumes it now has rq_timeout from when the request
389          * arrived, so the client should give it at least that long.
390          * since we don't know the arrival time we'll use the original
391          * sent time */
392         req->rq_deadline = req->rq_sent + req->rq_timeout +
393                            ptlrpc_at_get_net_latency(req);
394
395         DEBUG_REQ(D_ADAPTTO, req,
396                   "Early reply #%d, new deadline in "CFS_DURATION_T"s "
397                   "("CFS_DURATION_T"s)", req->rq_early_count,
398                   cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
399                   cfs_time_sub(req->rq_deadline, olddl));
400
401         RETURN(rc);
402 }
403
404 static struct kmem_cache *request_cache;
405
406 int ptlrpc_request_cache_init(void)
407 {
408         request_cache = kmem_cache_create("ptlrpc_cache",
409                                           sizeof(struct ptlrpc_request),
410                                           0, SLAB_HWCACHE_ALIGN, NULL);
411         return request_cache == NULL ? -ENOMEM : 0;
412 }
413
414 void ptlrpc_request_cache_fini(void)
415 {
416         kmem_cache_destroy(request_cache);
417 }
418
419 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
420 {
421         struct ptlrpc_request *req;
422
423         OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
424         return req;
425 }
426
427 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
428 {
429         OBD_SLAB_FREE_PTR(req, request_cache);
430 }
431
432 /**
433  * Wind down request pool \a pool.
434  * Frees all requests from the pool too
435  */
436 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
437 {
438         struct list_head *l, *tmp;
439         struct ptlrpc_request *req;
440
441         LASSERT(pool != NULL);
442
443         spin_lock(&pool->prp_lock);
444         list_for_each_safe(l, tmp, &pool->prp_req_list) {
445                 req = list_entry(l, struct ptlrpc_request, rq_list);
446                 list_del(&req->rq_list);
447                 LASSERT(req->rq_reqbuf);
448                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
449                 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
450                 ptlrpc_request_cache_free(req);
451         }
452         spin_unlock(&pool->prp_lock);
453         OBD_FREE(pool, sizeof(*pool));
454 }
455 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
456
457 /**
458  * Allocates, initializes and adds \a num_rq requests to the pool \a pool
459  */
460 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
461 {
462         int i;
463         int size = 1;
464
465         while (size < pool->prp_rq_size)
466                 size <<= 1;
467
468         LASSERTF(list_empty(&pool->prp_req_list) ||
469                  size == pool->prp_rq_size,
470                  "Trying to change pool size with nonempty pool "
471                  "from %d to %d bytes\n", pool->prp_rq_size, size);
472
473         spin_lock(&pool->prp_lock);
474         pool->prp_rq_size = size;
475         for (i = 0; i < num_rq; i++) {
476                 struct ptlrpc_request *req;
477                 struct lustre_msg *msg;
478
479                 spin_unlock(&pool->prp_lock);
480                 req = ptlrpc_request_cache_alloc(GFP_NOFS);
481                 if (!req)
482                         return;
483                 OBD_ALLOC_LARGE(msg, size);
484                 if (!msg) {
485                         ptlrpc_request_cache_free(req);
486                         return;
487                 }
488                 req->rq_reqbuf = msg;
489                 req->rq_reqbuf_len = size;
490                 req->rq_pool = pool;
491                 spin_lock(&pool->prp_lock);
492                 list_add_tail(&req->rq_list, &pool->prp_req_list);
493         }
494         spin_unlock(&pool->prp_lock);
495         return;
496 }
497 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
498
499 /**
500  * Create and initialize new request pool with given attributes:
501  * \a num_rq - initial number of requests to create for the pool
502  * \a msgsize - maximum message size possible for requests in thid pool
503  * \a populate_pool - function to be called when more requests need to be added
504  *                    to the pool
505  * Returns pointer to newly created pool or NULL on error.
506  */
507 struct ptlrpc_request_pool *
508 ptlrpc_init_rq_pool(int num_rq, int msgsize,
509                     void (*populate_pool)(struct ptlrpc_request_pool *, int))
510 {
511         struct ptlrpc_request_pool *pool;
512
513         OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
514         if (!pool)
515                 return NULL;
516
517         /* Request next power of two for the allocation, because internally
518            kernel would do exactly this */
519
520         spin_lock_init(&pool->prp_lock);
521         INIT_LIST_HEAD(&pool->prp_req_list);
522         pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
523         pool->prp_populate = populate_pool;
524
525         populate_pool(pool, num_rq);
526
527         if (list_empty(&pool->prp_req_list)) {
528                 /* have not allocated a single request for the pool */
529                 OBD_FREE(pool, sizeof(struct ptlrpc_request_pool));
530                 pool = NULL;
531         }
532         return pool;
533 }
534 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
535
536 /**
537  * Fetches one request from pool \a pool
538  */
539 static struct ptlrpc_request *
540 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
541 {
542         struct ptlrpc_request *request;
543         struct lustre_msg *reqbuf;
544
545         if (!pool)
546                 return NULL;
547
548         spin_lock(&pool->prp_lock);
549
550         /* See if we have anything in a pool, and bail out if nothing,
551          * in writeout path, where this matters, this is safe to do, because
552          * nothing is lost in this case, and when some in-flight requests
553          * complete, this code will be called again. */
554         if (unlikely(list_empty(&pool->prp_req_list))) {
555                 spin_unlock(&pool->prp_lock);
556                 return NULL;
557         }
558
559         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
560                              rq_list);
561         list_del_init(&request->rq_list);
562         spin_unlock(&pool->prp_lock);
563
564         LASSERT(request->rq_reqbuf);
565         LASSERT(request->rq_pool);
566
567         reqbuf = request->rq_reqbuf;
568         memset(request, 0, sizeof(*request));
569         request->rq_reqbuf = reqbuf;
570         request->rq_reqbuf_len = pool->prp_rq_size;
571         request->rq_pool = pool;
572
573         return request;
574 }
575
576 /**
577  * Returns freed \a request to pool.
578  */
579 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
580 {
581         struct ptlrpc_request_pool *pool = request->rq_pool;
582
583         spin_lock(&pool->prp_lock);
584         LASSERT(list_empty(&request->rq_list));
585         LASSERT(!request->rq_receiving_reply);
586         list_add_tail(&request->rq_list, &pool->prp_req_list);
587         spin_unlock(&pool->prp_lock);
588 }
589
590 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
591                                       __u32 version, int opcode,
592                                       int count, __u32 *lengths, char **bufs,
593                                       struct ptlrpc_cli_ctx *ctx)
594 {
595         struct obd_import  *imp = request->rq_import;
596         int                 rc;
597         ENTRY;
598
599         if (unlikely(ctx))
600                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
601         else {
602                 rc = sptlrpc_req_get_ctx(request);
603                 if (rc)
604                         GOTO(out_free, rc);
605         }
606
607         sptlrpc_req_set_flavor(request, opcode);
608
609         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
610                                  lengths, bufs);
611         if (rc) {
612                 LASSERT(!request->rq_pool);
613                 GOTO(out_ctx, rc);
614         }
615
616         lustre_msg_add_version(request->rq_reqmsg, version);
617         request->rq_send_state = LUSTRE_IMP_FULL;
618         request->rq_type = PTL_RPC_MSG_REQUEST;
619
620         request->rq_req_cbid.cbid_fn  = request_out_callback;
621         request->rq_req_cbid.cbid_arg = request;
622
623         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
624         request->rq_reply_cbid.cbid_arg = request;
625
626         request->rq_reply_deadline = 0;
627         request->rq_phase = RQ_PHASE_NEW;
628         request->rq_next_phase = RQ_PHASE_UNDEFINED;
629
630         request->rq_request_portal = imp->imp_client->cli_request_portal;
631         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
632
633         ptlrpc_at_set_req_timeout(request);
634
635         request->rq_xid = ptlrpc_next_xid();
636         lustre_msg_set_opc(request->rq_reqmsg, opcode);
637
638         RETURN(0);
639 out_ctx:
640         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
641 out_free:
642         class_import_put(imp);
643         return rc;
644 }
645
646 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
647                              __u32 version, int opcode, char **bufs,
648                              struct ptlrpc_cli_ctx *ctx)
649 {
650         int count;
651
652         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
653         return __ptlrpc_request_bufs_pack(request, version, opcode, count,
654                                           request->rq_pill.rc_area[RCL_CLIENT],
655                                           bufs, ctx);
656 }
657 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
658
659 /**
660  * Pack request buffers for network transfer, performing necessary encryption
661  * steps if necessary.
662  */
663 int ptlrpc_request_pack(struct ptlrpc_request *request,
664                         __u32 version, int opcode)
665 {
666         int rc;
667         rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
668         if (rc)
669                 return rc;
670
671         /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
672          * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
673          * have to send old ptlrpc_body to keep interoprability with these
674          * clients.
675          *
676          * Only three kinds of server->client RPCs so far:
677          *  - LDLM_BL_CALLBACK
678          *  - LDLM_CP_CALLBACK
679          *  - LDLM_GL_CALLBACK
680          *
681          * XXX This should be removed whenever we drop the interoprability with
682          *     the these old clients.
683          */
684         if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
685             opcode == LDLM_GL_CALLBACK)
686                 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
687                                    sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
688
689         return rc;
690 }
691 EXPORT_SYMBOL(ptlrpc_request_pack);
692
693 /**
694  * Helper function to allocate new request on import \a imp
695  * and possibly using existing request from pool \a pool if provided.
696  * Returns allocated request structure with import field filled or
697  * NULL on error.
698  */
699 static inline
700 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
701                                               struct ptlrpc_request_pool *pool)
702 {
703         struct ptlrpc_request *request = NULL;
704
705         if (pool)
706                 request = ptlrpc_prep_req_from_pool(pool);
707
708         if (!request)
709                 request = ptlrpc_request_cache_alloc(GFP_NOFS);
710
711         if (request) {
712                 ptlrpc_cli_req_init(request);
713
714                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
715                 LASSERT(imp != LP_POISON);
716                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
717                         imp->imp_client);
718                 LASSERT(imp->imp_client != LP_POISON);
719
720                 request->rq_import = class_import_get(imp);
721         } else {
722                 CERROR("request allocation out of memory\n");
723         }
724
725         return request;
726 }
727
728 /**
729  * Helper function for creating a request.
730  * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
731  * buffer structures according to capsule template \a format.
732  * Returns allocated request structure pointer or NULL on error.
733  */
734 static struct ptlrpc_request *
735 ptlrpc_request_alloc_internal(struct obd_import *imp,
736                               struct ptlrpc_request_pool * pool,
737                               const struct req_format *format)
738 {
739         struct ptlrpc_request *request;
740
741         request = __ptlrpc_request_alloc(imp, pool);
742         if (request == NULL)
743                 return NULL;
744
745         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
746         req_capsule_set(&request->rq_pill, format);
747         return request;
748 }
749
750 /**
751  * Allocate new request structure for import \a imp and initialize its
752  * buffer structure according to capsule template \a format.
753  */
754 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
755                                             const struct req_format *format)
756 {
757         return ptlrpc_request_alloc_internal(imp, NULL, format);
758 }
759 EXPORT_SYMBOL(ptlrpc_request_alloc);
760
761 /**
762  * Allocate new request structure for import \a imp from pool \a pool and
763  * initialize its buffer structure according to capsule template \a format.
764  */
765 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
766                                             struct ptlrpc_request_pool * pool,
767                                             const struct req_format *format)
768 {
769         return ptlrpc_request_alloc_internal(imp, pool, format);
770 }
771 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
772
773 /**
774  * For requests not from pool, free memory of the request structure.
775  * For requests obtained from a pool earlier, return request back to pool.
776  */
777 void ptlrpc_request_free(struct ptlrpc_request *request)
778 {
779         if (request->rq_pool)
780                 __ptlrpc_free_req_to_pool(request);
781         else
782                 ptlrpc_request_cache_free(request);
783 }
784 EXPORT_SYMBOL(ptlrpc_request_free);
785
786 /**
787  * Allocate new request for operatione \a opcode and immediatelly pack it for
788  * network transfer.
789  * Only used for simple requests like OBD_PING where the only important
790  * part of the request is operation itself.
791  * Returns allocated request or NULL on error.
792  */
793 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
794                                                 const struct req_format *format,
795                                                 __u32 version, int opcode)
796 {
797         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
798         int                    rc;
799
800         if (req) {
801                 rc = ptlrpc_request_pack(req, version, opcode);
802                 if (rc) {
803                         ptlrpc_request_free(req);
804                         req = NULL;
805                 }
806         }
807         return req;
808 }
809 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
810
811 /**
812  * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
813  * for operation \a opcode. Request would contain \a count buffers.
814  * Sizes of buffers are described in array \a lengths and buffers themselves
815  * are provided by a pointer \a bufs.
816  * Returns prepared request structure pointer or NULL on error.
817  */
818 struct ptlrpc_request *
819 ptlrpc_prep_req_pool(struct obd_import *imp,
820                      __u32 version, int opcode,
821                      int count, __u32 *lengths, char **bufs,
822                      struct ptlrpc_request_pool *pool)
823 {
824         struct ptlrpc_request *request;
825         int                    rc;
826
827         request = __ptlrpc_request_alloc(imp, pool);
828         if (!request)
829                 return NULL;
830
831         rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
832                                         lengths, bufs, NULL);
833         if (rc) {
834                 ptlrpc_request_free(request);
835                 request = NULL;
836         }
837         return request;
838 }
839
840 /**
841  * Same as ptlrpc_prep_req_pool, but without pool
842  */
843 struct ptlrpc_request *
844 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
845                 __u32 *lengths, char **bufs)
846 {
847         return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
848                                     NULL);
849 }
850
851 /**
852  * Allocate and initialize new request set structure.
853  * Returns a pointer to the newly allocated set structure or NULL on error.
854  */
855 struct ptlrpc_request_set *ptlrpc_prep_set(void)
856 {
857         struct ptlrpc_request_set *set;
858
859         ENTRY;
860         OBD_ALLOC(set, sizeof *set);
861         if (!set)
862                 RETURN(NULL);
863         atomic_set(&set->set_refcount, 1);
864         INIT_LIST_HEAD(&set->set_requests);
865         init_waitqueue_head(&set->set_waitq);
866         atomic_set(&set->set_new_count, 0);
867         atomic_set(&set->set_remaining, 0);
868         spin_lock_init(&set->set_new_req_lock);
869         INIT_LIST_HEAD(&set->set_new_requests);
870         INIT_LIST_HEAD(&set->set_cblist);
871         set->set_max_inflight = UINT_MAX;
872         set->set_producer     = NULL;
873         set->set_producer_arg = NULL;
874         set->set_rc           = 0;
875
876         RETURN(set);
877 }
878 EXPORT_SYMBOL(ptlrpc_prep_set);
879
880 /**
881  * Allocate and initialize new request set structure with flow control
882  * extension. This extension allows to control the number of requests in-flight
883  * for the whole set. A callback function to generate requests must be provided
884  * and the request set will keep the number of requests sent over the wire to
885  * @max_inflight.
886  * Returns a pointer to the newly allocated set structure or NULL on error.
887  */
888 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
889                                              void *arg)
890
891 {
892         struct ptlrpc_request_set *set;
893
894         set = ptlrpc_prep_set();
895         if (!set)
896                 RETURN(NULL);
897
898         set->set_max_inflight  = max;
899         set->set_producer      = func;
900         set->set_producer_arg  = arg;
901
902         RETURN(set);
903 }
904
905 /**
906  * Wind down and free request set structure previously allocated with
907  * ptlrpc_prep_set.
908  * Ensures that all requests on the set have completed and removes
909  * all requests from the request list in a set.
910  * If any unsent request happen to be on the list, pretends that they got
911  * an error in flight and calls their completion handler.
912  */
913 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
914 {
915         struct list_head        *tmp;
916         struct list_head        *next;
917         int                      expected_phase;
918         int                      n = 0;
919         ENTRY;
920
921         /* Requests on the set should either all be completed, or all be new */
922         expected_phase = (atomic_read(&set->set_remaining) == 0) ?
923                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
924         list_for_each(tmp, &set->set_requests) {
925                 struct ptlrpc_request *req =
926                         list_entry(tmp, struct ptlrpc_request,
927                                    rq_set_chain);
928
929                 LASSERT(req->rq_phase == expected_phase);
930                 n++;
931         }
932
933         LASSERTF(atomic_read(&set->set_remaining) == 0 ||
934                  atomic_read(&set->set_remaining) == n, "%d / %d\n",
935                  atomic_read(&set->set_remaining), n);
936
937         list_for_each_safe(tmp, next, &set->set_requests) {
938                 struct ptlrpc_request *req =
939                         list_entry(tmp, struct ptlrpc_request,
940                                    rq_set_chain);
941                 list_del_init(&req->rq_set_chain);
942
943                 LASSERT(req->rq_phase == expected_phase);
944
945                 if (req->rq_phase == RQ_PHASE_NEW) {
946                         ptlrpc_req_interpret(NULL, req, -EBADR);
947                         atomic_dec(&set->set_remaining);
948                 }
949
950                 spin_lock(&req->rq_lock);
951                 req->rq_set = NULL;
952                 req->rq_invalid_rqset = 0;
953                 spin_unlock(&req->rq_lock);
954
955                 ptlrpc_req_finished (req);
956         }
957
958         LASSERT(atomic_read(&set->set_remaining) == 0);
959
960         ptlrpc_reqset_put(set);
961         EXIT;
962 }
963 EXPORT_SYMBOL(ptlrpc_set_destroy);
964
965 /**
966  * Add a callback function \a fn to the set.
967  * This function would be called when all requests on this set are completed.
968  * The function will be passed \a data argument.
969  */
970 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
971                       set_interpreter_func fn, void *data)
972 {
973         struct ptlrpc_set_cbdata *cbdata;
974
975         OBD_ALLOC_PTR(cbdata);
976         if (cbdata == NULL)
977                 RETURN(-ENOMEM);
978
979         cbdata->psc_interpret = fn;
980         cbdata->psc_data = data;
981         list_add_tail(&cbdata->psc_item, &set->set_cblist);
982
983         RETURN(0);
984 }
985
986 /**
987  * Add a new request to the general purpose request set.
988  * Assumes request reference from the caller.
989  */
990 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
991                         struct ptlrpc_request *req)
992 {
993         LASSERT(list_empty(&req->rq_set_chain));
994
995         /* The set takes over the caller's request reference */
996         list_add_tail(&req->rq_set_chain, &set->set_requests);
997         req->rq_set = set;
998         atomic_inc(&set->set_remaining);
999         req->rq_queued_time = cfs_time_current();
1000
1001         if (req->rq_reqmsg != NULL)
1002                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1003
1004         if (set->set_producer != NULL)
1005                 /* If the request set has a producer callback, the RPC must be
1006                  * sent straight away */
1007                 ptlrpc_send_new_req(req);
1008 }
1009 EXPORT_SYMBOL(ptlrpc_set_add_req);
1010
1011 /**
1012  * Add a request to a request with dedicated server thread
1013  * and wake the thread to make any necessary processing.
1014  * Currently only used for ptlrpcd.
1015  */
1016 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1017                            struct ptlrpc_request *req)
1018 {
1019         struct ptlrpc_request_set *set = pc->pc_set;
1020         int count, i;
1021
1022         LASSERT(req->rq_set == NULL);
1023         LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1024
1025         spin_lock(&set->set_new_req_lock);
1026         /*
1027          * The set takes over the caller's request reference.
1028          */
1029         req->rq_set = set;
1030         req->rq_queued_time = cfs_time_current();
1031         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1032         count = atomic_inc_return(&set->set_new_count);
1033         spin_unlock(&set->set_new_req_lock);
1034
1035         /* Only need to call wakeup once for the first entry. */
1036         if (count == 1) {
1037                 wake_up(&set->set_waitq);
1038
1039                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1040                  *      guarantee the async RPC can be processed ASAP, we have
1041                  *      no other better choice. It maybe fixed in future. */
1042                 for (i = 0; i < pc->pc_npartners; i++)
1043                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1044         }
1045 }
1046
1047 /**
1048  * Based on the current state of the import, determine if the request
1049  * can be sent, is an error, or should be delayed.
1050  *
1051  * Returns true if this request should be delayed. If false, and
1052  * *status is set, then the request can not be sent and *status is the
1053  * error code.  If false and status is 0, then request can be sent.
1054  *
1055  * The imp->imp_lock must be held.
1056  */
1057 static int ptlrpc_import_delay_req(struct obd_import *imp,
1058                                    struct ptlrpc_request *req, int *status)
1059 {
1060         int delay = 0;
1061         ENTRY;
1062
1063         LASSERT (status != NULL);
1064         *status = 0;
1065
1066         if (req->rq_ctx_init || req->rq_ctx_fini) {
1067                 /* always allow ctx init/fini rpc go through */
1068         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1069                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1070                 *status = -EIO;
1071         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1072                 /* pings may safely race with umount */
1073                 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1074                           D_HA : D_ERROR, req, "IMP_CLOSED ");
1075                 *status = -EIO;
1076         } else if (ptlrpc_send_limit_expired(req)) {
1077                 /* probably doesn't need to be a D_ERROR after initial testing*/
1078                 DEBUG_REQ(D_HA, req, "send limit expired ");
1079                 *status = -ETIMEDOUT;
1080         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1081                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
1082                 /* allow CONNECT even if import is invalid */ ;
1083                 if (atomic_read(&imp->imp_inval_count) != 0) {
1084                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1085                         *status = -EIO;
1086                 }
1087         } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1088                 if (!imp->imp_deactive)
1089                         DEBUG_REQ(D_NET, req, "IMP_INVALID");
1090                 *status = -ESHUTDOWN; /* bz 12940 */
1091         } else if (req->rq_import_generation != imp->imp_generation) {
1092                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1093                 *status = -EIO;
1094         } else if (req->rq_send_state != imp->imp_state) {
1095                 /* invalidate in progress - any requests should be drop */
1096                 if (atomic_read(&imp->imp_inval_count) != 0) {
1097                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1098                         *status = -EIO;
1099                 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1100                         *status = -EWOULDBLOCK;
1101                 } else if (req->rq_allow_replay &&
1102                           (imp->imp_state == LUSTRE_IMP_REPLAY ||
1103                            imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1104                            imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1105                            imp->imp_state == LUSTRE_IMP_RECOVER)) {
1106                         DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1107                 } else {
1108                         delay = 1;
1109                 }
1110         }
1111
1112         RETURN(delay);
1113 }
1114
1115 /**
1116  * Decide if the error message should be printed to the console or not.
1117  * Makes its decision based on request type, status, and failure frequency.
1118  *
1119  * \param[in] req  request that failed and may need a console message
1120  *
1121  * \retval false if no message should be printed
1122  * \retval true  if console message should be printed
1123  */
1124 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1125 {
1126         __u32 opc;
1127
1128         LASSERT(req->rq_reqmsg != NULL);
1129         opc = lustre_msg_get_opc(req->rq_reqmsg);
1130
1131         /* Suppress particular reconnect errors which are to be expected. */
1132         if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1133                 int err;
1134
1135                 /* Suppress timed out reconnect requests */
1136                 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1137                     req->rq_timedout)
1138                         return false;
1139
1140                 /* Suppress most unavailable/again reconnect requests, but
1141                  * print occasionally so it is clear client is trying to
1142                  * connect to a server where no target is running. */
1143                 err = lustre_msg_get_status(req->rq_repmsg);
1144                 if ((err == -ENODEV || err == -EAGAIN) &&
1145                     req->rq_import->imp_conn_cnt % 30 != 20)
1146                         return false;
1147         }
1148
1149         return true;
1150 }
1151
1152 /**
1153  * Check request processing status.
1154  * Returns the status.
1155  */
1156 static int ptlrpc_check_status(struct ptlrpc_request *req)
1157 {
1158         int err;
1159         ENTRY;
1160
1161         err = lustre_msg_get_status(req->rq_repmsg);
1162         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1163                 struct obd_import *imp = req->rq_import;
1164                 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1165                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1166
1167                 if (ptlrpc_console_allow(req))
1168                         LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1169                                            "failed: rc = %d\n",
1170                                            imp->imp_obd->obd_name,
1171                                            ll_opcode2str(opc),
1172                                            libcfs_nid2str(nid), err);
1173                 RETURN(err < 0 ? err : -EINVAL);
1174         }
1175
1176         if (err < 0) {
1177                 DEBUG_REQ(D_INFO, req, "status is %d", err);
1178         } else if (err > 0) {
1179                 /* XXX: translate this error from net to host */
1180                 DEBUG_REQ(D_INFO, req, "status is %d", err);
1181         }
1182
1183         RETURN(err);
1184 }
1185
1186 /**
1187  * save pre-versions of objects into request for replay.
1188  * Versions are obtained from server reply.
1189  * used for VBR.
1190  */
1191 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1192 {
1193         struct lustre_msg *repmsg = req->rq_repmsg;
1194         struct lustre_msg *reqmsg = req->rq_reqmsg;
1195         __u64 *versions = lustre_msg_get_versions(repmsg);
1196         ENTRY;
1197
1198         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1199                 return;
1200
1201         LASSERT(versions);
1202         lustre_msg_set_versions(reqmsg, versions);
1203         CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1204                versions[0], versions[1]);
1205
1206         EXIT;
1207 }
1208
1209 /**
1210  * Callback function called when client receives RPC reply for \a req.
1211  * Returns 0 on success or error code.
1212  * The return alue would be assigned to req->rq_status by the caller
1213  * as request processing status.
1214  * This function also decides if the request needs to be saved for later replay.
1215  */
1216 static int after_reply(struct ptlrpc_request *req)
1217 {
1218         struct obd_import *imp = req->rq_import;
1219         struct obd_device *obd = req->rq_import->imp_obd;
1220         int rc;
1221         struct timeval work_start;
1222         long timediff;
1223         ENTRY;
1224
1225         LASSERT(obd != NULL);
1226         /* repbuf must be unlinked */
1227         LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1228
1229         if (req->rq_reply_truncated) {
1230                 if (ptlrpc_no_resend(req)) {
1231                         DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1232                                   " expected: %d, actual size: %d",
1233                                   req->rq_nob_received, req->rq_repbuf_len);
1234                         RETURN(-EOVERFLOW);
1235                 }
1236
1237                 sptlrpc_cli_free_repbuf(req);
1238                 /* Pass the required reply buffer size (include
1239                  * space for early reply).
1240                  * NB: no need to roundup because alloc_repbuf
1241                  * will roundup it */
1242                 req->rq_replen       = req->rq_nob_received;
1243                 req->rq_nob_received = 0;
1244                 spin_lock(&req->rq_lock);
1245                 req->rq_resend       = 1;
1246                 spin_unlock(&req->rq_lock);
1247                 RETURN(0);
1248         }
1249
1250         /*
1251          * NB Until this point, the whole of the incoming message,
1252          * including buflens, status etc is in the sender's byte order.
1253          */
1254         rc = sptlrpc_cli_unwrap_reply(req);
1255         if (rc) {
1256                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1257                 RETURN(rc);
1258         }
1259
1260         /*
1261          * Security layer unwrap might ask resend this request.
1262          */
1263         if (req->rq_resend)
1264                 RETURN(0);
1265
1266         rc = unpack_reply(req);
1267         if (rc)
1268                 RETURN(rc);
1269
1270         /* retry indefinitely on EINPROGRESS */
1271         if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1272             ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1273                 time_t  now = cfs_time_current_sec();
1274
1275                 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1276                 spin_lock(&req->rq_lock);
1277                 req->rq_resend = 1;
1278                 spin_unlock(&req->rq_lock);
1279                 req->rq_nr_resend++;
1280
1281                 /* allocate new xid to avoid reply reconstruction */
1282                 if (!req->rq_bulk) {
1283                         /* new xid is already allocated for bulk in
1284                          * ptlrpc_check_set() */
1285                         req->rq_xid = ptlrpc_next_xid();
1286                         DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1287                                   "resend on EINPROGRESS");
1288                 }
1289
1290                 /* Readjust the timeout for current conditions */
1291                 ptlrpc_at_set_req_timeout(req);
1292                 /* delay resend to give a chance to the server to get ready.
1293                  * The delay is increased by 1s on every resend and is capped to
1294                  * the current request timeout (i.e. obd_timeout if AT is off,
1295                  * or AT service time x 125% + 5s, see at_est2timeout) */
1296                 if (req->rq_nr_resend > req->rq_timeout)
1297                         req->rq_sent = now + req->rq_timeout;
1298                 else
1299                         req->rq_sent = now + req->rq_nr_resend;
1300
1301                 RETURN(0);
1302         }
1303
1304         do_gettimeofday(&work_start);
1305         timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1306         if (obd->obd_svc_stats != NULL) {
1307                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1308                                     timediff);
1309                 ptlrpc_lprocfs_rpc_sent(req, timediff);
1310         }
1311
1312         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1313             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1314                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1315                           lustre_msg_get_type(req->rq_repmsg));
1316                 RETURN(-EPROTO);
1317         }
1318
1319         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1320                 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1321         ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1322         ptlrpc_at_adj_net_latency(req,
1323                                   lustre_msg_get_service_time(req->rq_repmsg));
1324
1325         rc = ptlrpc_check_status(req);
1326         imp->imp_connect_error = rc;
1327
1328         if (rc) {
1329                 /*
1330                  * Either we've been evicted, or the server has failed for
1331                  * some reason. Try to reconnect, and if that fails, punt to
1332                  * the upcall.
1333                  */
1334                 if (ptlrpc_recoverable_error(rc)) {
1335                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
1336                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1337                                 RETURN(rc);
1338                         }
1339                         ptlrpc_request_handle_notconn(req);
1340                         RETURN(rc);
1341                 }
1342         } else {
1343                 /*
1344                  * Let's look if server sent slv. Do it only for RPC with
1345                  * rc == 0.
1346                  */
1347                 ldlm_cli_update_pool(req);
1348         }
1349
1350         /*
1351          * Store transno in reqmsg for replay.
1352          */
1353         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1354                 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1355                 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1356         }
1357
1358         if (imp->imp_replayable) {
1359                 spin_lock(&imp->imp_lock);
1360                 /*
1361                  * No point in adding already-committed requests to the replay
1362                  * list, we will just remove them immediately. b=9829
1363                  */
1364                 if (req->rq_transno != 0 &&
1365                     (req->rq_transno >
1366                      lustre_msg_get_last_committed(req->rq_repmsg) ||
1367                      req->rq_replay)) {
1368                         /** version recovery */
1369                         ptlrpc_save_versions(req);
1370                         ptlrpc_retain_replayable_request(req, imp);
1371                 } else if (req->rq_commit_cb != NULL &&
1372                            list_empty(&req->rq_replay_list)) {
1373                         /* NB: don't call rq_commit_cb if it's already on
1374                          * rq_replay_list, ptlrpc_free_committed() will call
1375                          * it later, see LU-3618 for details */
1376                         spin_unlock(&imp->imp_lock);
1377                         req->rq_commit_cb(req);
1378                         spin_lock(&imp->imp_lock);
1379                 }
1380
1381                 /*
1382                  * Replay-enabled imports return commit-status information.
1383                  */
1384                 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1385                         imp->imp_peer_committed_transno =
1386                                 lustre_msg_get_last_committed(req->rq_repmsg);
1387                 }
1388
1389                 ptlrpc_free_committed(imp);
1390
1391                 if (!list_empty(&imp->imp_replay_list)) {
1392                         struct ptlrpc_request *last;
1393
1394                         last = list_entry(imp->imp_replay_list.prev,
1395                                           struct ptlrpc_request,
1396                                           rq_replay_list);
1397                         /*
1398                          * Requests with rq_replay stay on the list even if no
1399                          * commit is expected.
1400                          */
1401                         if (last->rq_transno > imp->imp_peer_committed_transno)
1402                                 ptlrpc_pinger_commit_expected(imp);
1403                 }
1404
1405                 spin_unlock(&imp->imp_lock);
1406         }
1407
1408         RETURN(rc);
1409 }
1410
1411 /**
1412  * Helper function to send request \a req over the network for the first time
1413  * Also adjusts request phase.
1414  * Returns 0 on success or error code.
1415  */
1416 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1417 {
1418         struct obd_import     *imp = req->rq_import;
1419         int rc;
1420         ENTRY;
1421
1422         LASSERT(req->rq_phase == RQ_PHASE_NEW);
1423         if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1424             (!req->rq_generation_set ||
1425              req->rq_import_generation == imp->imp_generation))
1426                 RETURN (0);
1427
1428         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1429
1430         spin_lock(&imp->imp_lock);
1431
1432         if (!req->rq_generation_set)
1433                 req->rq_import_generation = imp->imp_generation;
1434
1435         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1436                 spin_lock(&req->rq_lock);
1437                 req->rq_waiting = 1;
1438                 spin_unlock(&req->rq_lock);
1439
1440                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1441                           "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1442                           ptlrpc_import_state_name(req->rq_send_state),
1443                           ptlrpc_import_state_name(imp->imp_state));
1444                 LASSERT(list_empty(&req->rq_list));
1445                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1446                 atomic_inc(&req->rq_import->imp_inflight);
1447                 spin_unlock(&imp->imp_lock);
1448                 RETURN(0);
1449         }
1450
1451         if (rc != 0) {
1452                 spin_unlock(&imp->imp_lock);
1453                 req->rq_status = rc;
1454                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1455                 RETURN(rc);
1456         }
1457
1458         LASSERT(list_empty(&req->rq_list));
1459         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1460         atomic_inc(&req->rq_import->imp_inflight);
1461         spin_unlock(&imp->imp_lock);
1462
1463         lustre_msg_set_status(req->rq_reqmsg, current_pid());
1464
1465         rc = sptlrpc_req_refresh_ctx(req, -1);
1466         if (rc) {
1467                 if (req->rq_err) {
1468                         req->rq_status = rc;
1469                         RETURN(1);
1470                 } else {
1471                         spin_lock(&req->rq_lock);
1472                         req->rq_wait_ctx = 1;
1473                         spin_unlock(&req->rq_lock);
1474                         RETURN(0);
1475                 }
1476         }
1477
1478         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1479                " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1480                imp->imp_obd->obd_uuid.uuid,
1481                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1482                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1483                lustre_msg_get_opc(req->rq_reqmsg));
1484
1485         rc = ptl_send_rpc(req, 0);
1486         if (rc) {
1487                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1488                 spin_lock(&req->rq_lock);
1489                 req->rq_net_err = 1;
1490                 spin_unlock(&req->rq_lock);
1491                 RETURN(rc);
1492         }
1493         RETURN(0);
1494 }
1495
1496 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1497 {
1498         int remaining, rc;
1499         ENTRY;
1500
1501         LASSERT(set->set_producer != NULL);
1502
1503         remaining = atomic_read(&set->set_remaining);
1504
1505         /* populate the ->set_requests list with requests until we
1506          * reach the maximum number of RPCs in flight for this set */
1507         while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1508                 rc = set->set_producer(set, set->set_producer_arg);
1509                 if (rc == -ENOENT) {
1510                         /* no more RPC to produce */
1511                         set->set_producer     = NULL;
1512                         set->set_producer_arg = NULL;
1513                         RETURN(0);
1514                 }
1515         }
1516
1517         RETURN((atomic_read(&set->set_remaining) - remaining));
1518 }
1519
1520 /**
1521  * this sends any unsent RPCs in \a set and returns 1 if all are sent
1522  * and no more replies are expected.
1523  * (it is possible to get less replies than requests sent e.g. due to timed out
1524  * requests or requests that we had trouble to send out)
1525  *
1526  * NOTE: This function contains a potential schedule point (cond_resched()).
1527  */
1528 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1529 {
1530         struct list_head *tmp, *next;
1531         struct list_head  comp_reqs;
1532         int force_timer_recalc = 0;
1533         ENTRY;
1534
1535         if (atomic_read(&set->set_remaining) == 0)
1536                 RETURN(1);
1537
1538         INIT_LIST_HEAD(&comp_reqs);
1539         list_for_each_safe(tmp, next, &set->set_requests) {
1540                 struct ptlrpc_request *req =
1541                         list_entry(tmp, struct ptlrpc_request,
1542                                    rq_set_chain);
1543                 struct obd_import *imp = req->rq_import;
1544                 int unregistered = 0;
1545                 int rc = 0;
1546
1547                 /* This schedule point is mainly for the ptlrpcd caller of this
1548                  * function.  Most ptlrpc sets are not long-lived and unbounded
1549                  * in length, but at the least the set used by the ptlrpcd is.
1550                  * Since the processing time is unbounded, we need to insert an
1551                  * explicit schedule point to make the thread well-behaved.
1552                  */
1553                 cond_resched();
1554
1555                 if (req->rq_phase == RQ_PHASE_NEW &&
1556                     ptlrpc_send_new_req(req)) {
1557                         force_timer_recalc = 1;
1558                 }
1559
1560                 /* delayed send - skip */
1561                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1562                         continue;
1563
1564                 /* delayed resend - skip */
1565                 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1566                     req->rq_sent > cfs_time_current_sec())
1567                         continue;
1568
1569                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1570                       req->rq_phase == RQ_PHASE_BULK ||
1571                       req->rq_phase == RQ_PHASE_INTERPRET ||
1572                       req->rq_phase == RQ_PHASE_UNREGISTERING ||
1573                       req->rq_phase == RQ_PHASE_COMPLETE)) {
1574                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1575                         LBUG();
1576                 }
1577
1578                 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1579                         LASSERT(req->rq_next_phase != req->rq_phase);
1580                         LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1581
1582                         /*
1583                          * Skip processing until reply is unlinked. We
1584                          * can't return to pool before that and we can't
1585                          * call interpret before that. We need to make
1586                          * sure that all rdma transfers finished and will
1587                          * not corrupt any data.
1588                          */
1589                         if (ptlrpc_client_recv_or_unlink(req) ||
1590                             ptlrpc_client_bulk_active(req))
1591                                 continue;
1592
1593                         /*
1594                          * Turn fail_loc off to prevent it from looping
1595                          * forever.
1596                          */
1597                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1598                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1599                                                      OBD_FAIL_ONCE);
1600                         }
1601                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1602                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1603                                                      OBD_FAIL_ONCE);
1604                         }
1605
1606                         /*
1607                          * Move to next phase if reply was successfully
1608                          * unlinked.
1609                          */
1610                         ptlrpc_rqphase_move(req, req->rq_next_phase);
1611                 }
1612
1613                 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1614                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1615                         continue;
1616                 }
1617
1618                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1619                         GOTO(interpret, req->rq_status);
1620
1621                 /*
1622                  * Note that this also will start async reply unlink.
1623                  */
1624                 if (req->rq_net_err && !req->rq_timedout) {
1625                         ptlrpc_expire_one_request(req, 1);
1626
1627                         /*
1628                          * Check if we still need to wait for unlink.
1629                          */
1630                         if (ptlrpc_client_recv_or_unlink(req) ||
1631                             ptlrpc_client_bulk_active(req))
1632                                 continue;
1633                         /* If there is no need to resend, fail it now. */
1634                         if (req->rq_no_resend) {
1635                                 if (req->rq_status == 0)
1636                                         req->rq_status = -EIO;
1637                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1638                                 GOTO(interpret, req->rq_status);
1639                         } else {
1640                                 continue;
1641                         }
1642                 }
1643
1644                 if (req->rq_err) {
1645                         spin_lock(&req->rq_lock);
1646                         req->rq_replied = 0;
1647                         spin_unlock(&req->rq_lock);
1648                         if (req->rq_status == 0)
1649                                 req->rq_status = -EIO;
1650                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1651                         GOTO(interpret, req->rq_status);
1652                 }
1653
1654                 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1655                  * so it sets rq_intr regardless of individual rpc
1656                  * timeouts. The synchronous IO waiting path sets
1657                  * rq_intr irrespective of whether ptlrpcd
1658                  * has seen a timeout.  Our policy is to only interpret
1659                  * interrupted rpcs after they have timed out, so we
1660                  * need to enforce that here.
1661                  */
1662
1663                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1664                                      req->rq_wait_ctx)) {
1665                         req->rq_status = -EINTR;
1666                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1667                         GOTO(interpret, req->rq_status);
1668                 }
1669
1670                 if (req->rq_phase == RQ_PHASE_RPC) {
1671                         if (req->rq_timedout || req->rq_resend ||
1672                             req->rq_waiting || req->rq_wait_ctx) {
1673                                 int status;
1674
1675                                 if (!ptlrpc_unregister_reply(req, 1)) {
1676                                         ptlrpc_unregister_bulk(req, 1);
1677                                         continue;
1678                                 }
1679
1680                                 spin_lock(&imp->imp_lock);
1681                                 if (ptlrpc_import_delay_req(imp, req, &status)){
1682                                         /* put on delay list - only if we wait
1683                                          * recovery finished - before send */
1684                                         list_del_init(&req->rq_list);
1685                                         list_add_tail(&req->rq_list,
1686                                                           &imp->
1687                                                           imp_delayed_list);
1688                                         spin_unlock(&imp->imp_lock);
1689                                         continue;
1690                                 }
1691
1692                                 if (status != 0)  {
1693                                         req->rq_status = status;
1694                                         ptlrpc_rqphase_move(req,
1695                                                 RQ_PHASE_INTERPRET);
1696                                         spin_unlock(&imp->imp_lock);
1697                                         GOTO(interpret, req->rq_status);
1698                                 }
1699                                 if (ptlrpc_no_resend(req) &&
1700                                     !req->rq_wait_ctx) {
1701                                         req->rq_status = -ENOTCONN;
1702                                         ptlrpc_rqphase_move(req,
1703                                                             RQ_PHASE_INTERPRET);
1704                                         spin_unlock(&imp->imp_lock);
1705                                         GOTO(interpret, req->rq_status);
1706                                 }
1707
1708                                 list_del_init(&req->rq_list);
1709                                 list_add_tail(&req->rq_list,
1710                                                   &imp->imp_sending_list);
1711
1712                                 spin_unlock(&imp->imp_lock);
1713
1714                                 spin_lock(&req->rq_lock);
1715                                 req->rq_waiting = 0;
1716                                 spin_unlock(&req->rq_lock);
1717
1718                                 if (req->rq_timedout || req->rq_resend) {
1719                                         /* This is re-sending anyways,
1720                                          * let's mark req as resend. */
1721                                         spin_lock(&req->rq_lock);
1722                                         req->rq_resend = 1;
1723                                         spin_unlock(&req->rq_lock);
1724                                         if (req->rq_bulk) {
1725                                                 __u64 old_xid;
1726
1727                                                 if (!ptlrpc_unregister_bulk(req, 1))
1728                                                         continue;
1729
1730                                                 /* ensure previous bulk fails */
1731                                                 old_xid = req->rq_xid;
1732                                                 req->rq_xid = ptlrpc_next_xid();
1733                                                 CDEBUG(D_HA, "resend bulk "
1734                                                        "old x"LPU64
1735                                                        " new x"LPU64"\n",
1736                                                        old_xid, req->rq_xid);
1737                                         }
1738                                 }
1739                                 /*
1740                                  * rq_wait_ctx is only touched by ptlrpcd,
1741                                  * so no lock is needed here.
1742                                  */
1743                                 status = sptlrpc_req_refresh_ctx(req, -1);
1744                                 if (status) {
1745                                         if (req->rq_err) {
1746                                                 req->rq_status = status;
1747                                                 spin_lock(&req->rq_lock);
1748                                                 req->rq_wait_ctx = 0;
1749                                                 spin_unlock(&req->rq_lock);
1750                                                 force_timer_recalc = 1;
1751                                         } else {
1752                                                 spin_lock(&req->rq_lock);
1753                                                 req->rq_wait_ctx = 1;
1754                                                 spin_unlock(&req->rq_lock);
1755                                         }
1756
1757                                         continue;
1758                                 } else {
1759                                         spin_lock(&req->rq_lock);
1760                                         req->rq_wait_ctx = 0;
1761                                         spin_unlock(&req->rq_lock);
1762                                 }
1763
1764                                 rc = ptl_send_rpc(req, 0);
1765                                 if (rc) {
1766                                         DEBUG_REQ(D_HA, req,
1767                                                   "send failed: rc = %d", rc);
1768                                         force_timer_recalc = 1;
1769                                         spin_lock(&req->rq_lock);
1770                                         req->rq_net_err = 1;
1771                                         spin_unlock(&req->rq_lock);
1772                                         continue;
1773                                 }
1774                                 /* need to reset the timeout */
1775                                 force_timer_recalc = 1;
1776                         }
1777
1778                         spin_lock(&req->rq_lock);
1779
1780                         if (ptlrpc_client_early(req)) {
1781                                 ptlrpc_at_recv_early_reply(req);
1782                                 spin_unlock(&req->rq_lock);
1783                                 continue;
1784                         }
1785
1786                         /* Still waiting for a reply? */
1787                         if (ptlrpc_client_recv(req)) {
1788                                 spin_unlock(&req->rq_lock);
1789                                 continue;
1790                         }
1791
1792                         /* Did we actually receive a reply? */
1793                         if (!ptlrpc_client_replied(req)) {
1794                                 spin_unlock(&req->rq_lock);
1795                                 continue;
1796                         }
1797
1798                         spin_unlock(&req->rq_lock);
1799
1800                         /* unlink from net because we are going to
1801                          * swab in-place of reply buffer */
1802                         unregistered = ptlrpc_unregister_reply(req, 1);
1803                         if (!unregistered)
1804                                 continue;
1805
1806                         req->rq_status = after_reply(req);
1807                         if (req->rq_resend)
1808                                 continue;
1809
1810                         /* If there is no bulk associated with this request,
1811                          * then we're done and should let the interpreter
1812                          * process the reply. Similarly if the RPC returned
1813                          * an error, and therefore the bulk will never arrive.
1814                          */
1815                         if (req->rq_bulk == NULL || req->rq_status < 0) {
1816                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1817                                 GOTO(interpret, req->rq_status);
1818                         }
1819
1820                         ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1821                 }
1822
1823                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1824                 if (ptlrpc_client_bulk_active(req))
1825                         continue;
1826
1827                 if (req->rq_bulk->bd_failure) {
1828                         /* The RPC reply arrived OK, but the bulk screwed
1829                          * up!  Dead weird since the server told us the RPC
1830                          * was good after getting the REPLY for her GET or
1831                          * the ACK for her PUT. */
1832                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1833                         req->rq_status = -EIO;
1834                 }
1835
1836                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1837
1838         interpret:
1839                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1840
1841                 /* This moves to "unregistering" phase we need to wait for
1842                  * reply unlink. */
1843                 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1844                         /* start async bulk unlink too */
1845                         ptlrpc_unregister_bulk(req, 1);
1846                         continue;
1847                 }
1848
1849                 if (!ptlrpc_unregister_bulk(req, 1))
1850                         continue;
1851
1852                 /* When calling interpret receiving already should be
1853                  * finished. */
1854                 LASSERT(!req->rq_receiving_reply);
1855
1856                 ptlrpc_req_interpret(env, req, req->rq_status);
1857
1858                 if (ptlrpcd_check_work(req)) {
1859                         atomic_dec(&set->set_remaining);
1860                         continue;
1861                 }
1862                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1863
1864                 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1865                         "Completed RPC pname:cluuid:pid:xid:nid:"
1866                         "opc %s:%s:%d:"LPU64":%s:%d\n",
1867                         current_comm(), imp->imp_obd->obd_uuid.uuid,
1868                         lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1869                         libcfs_nid2str(imp->imp_connection->c_peer.nid),
1870                         lustre_msg_get_opc(req->rq_reqmsg));
1871
1872                 spin_lock(&imp->imp_lock);
1873                 /* Request already may be not on sending or delaying list. This
1874                  * may happen in the case of marking it erroneous for the case
1875                  * ptlrpc_import_delay_req(req, status) find it impossible to
1876                  * allow sending this rpc and returns *status != 0. */
1877                 if (!list_empty(&req->rq_list)) {
1878                         list_del_init(&req->rq_list);
1879                         atomic_dec(&imp->imp_inflight);
1880                 }
1881                 spin_unlock(&imp->imp_lock);
1882
1883                 atomic_dec(&set->set_remaining);
1884                 wake_up_all(&imp->imp_recovery_waitq);
1885
1886                 if (set->set_producer) {
1887                         /* produce a new request if possible */
1888                         if (ptlrpc_set_producer(set) > 0)
1889                                 force_timer_recalc = 1;
1890
1891                         /* free the request that has just been completed
1892                          * in order not to pollute set->set_requests */
1893                         list_del_init(&req->rq_set_chain);
1894                         spin_lock(&req->rq_lock);
1895                         req->rq_set = NULL;
1896                         req->rq_invalid_rqset = 0;
1897                         spin_unlock(&req->rq_lock);
1898
1899                         /* record rq_status to compute the final status later */
1900                         if (req->rq_status != 0)
1901                                 set->set_rc = req->rq_status;
1902                         ptlrpc_req_finished(req);
1903                 } else {
1904                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1905                 }
1906         }
1907
1908         /* move completed request at the head of list so it's easier for
1909          * caller to find them */
1910         list_splice(&comp_reqs, &set->set_requests);
1911
1912         /* If we hit an error, we want to recover promptly. */
1913         RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1914 }
1915 EXPORT_SYMBOL(ptlrpc_check_set);
1916
1917 /**
1918  * Time out request \a req. is \a async_unlink is set, that means do not wait
1919  * until LNet actually confirms network buffer unlinking.
1920  * Return 1 if we should give up further retrying attempts or 0 otherwise.
1921  */
1922 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1923 {
1924         struct obd_import *imp = req->rq_import;
1925         int rc = 0;
1926         ENTRY;
1927
1928         spin_lock(&req->rq_lock);
1929         req->rq_timedout = 1;
1930         spin_unlock(&req->rq_lock);
1931
1932         DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1933                   "/real "CFS_DURATION_T"]",
1934                   req->rq_net_err ? "failed due to network error" :
1935                      ((req->rq_real_sent == 0 ||
1936                        cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1937                        cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1938                       "timed out for sent delay" : "timed out for slow reply"),
1939                   req->rq_sent, req->rq_real_sent);
1940
1941         if (imp != NULL && obd_debug_peer_on_timeout)
1942                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1943
1944         ptlrpc_unregister_reply(req, async_unlink);
1945         ptlrpc_unregister_bulk(req, async_unlink);
1946
1947         if (obd_dump_on_timeout)
1948                 libcfs_debug_dumplog();
1949
1950         if (imp == NULL) {
1951                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1952                 RETURN(1);
1953         }
1954
1955         atomic_inc(&imp->imp_timeouts);
1956
1957         /* The DLM server doesn't want recovery run on its imports. */
1958         if (imp->imp_dlm_fake)
1959                 RETURN(1);
1960
1961         /* If this request is for recovery or other primordial tasks,
1962          * then error it out here. */
1963         if (req->rq_ctx_init || req->rq_ctx_fini ||
1964             req->rq_send_state != LUSTRE_IMP_FULL ||
1965             imp->imp_obd->obd_no_recov) {
1966                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1967                           ptlrpc_import_state_name(req->rq_send_state),
1968                           ptlrpc_import_state_name(imp->imp_state));
1969                 spin_lock(&req->rq_lock);
1970                 req->rq_status = -ETIMEDOUT;
1971                 req->rq_err = 1;
1972                 spin_unlock(&req->rq_lock);
1973                 RETURN(1);
1974         }
1975
1976         /* if a request can't be resent we can't wait for an answer after
1977            the timeout */
1978         if (ptlrpc_no_resend(req)) {
1979                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1980                 rc = 1;
1981         }
1982
1983         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1984
1985         RETURN(rc);
1986 }
1987
1988 /**
1989  * Time out all uncompleted requests in request set pointed by \a data
1990  * Callback used when waiting on sets with l_wait_event.
1991  * Always returns 1.
1992  */
1993 int ptlrpc_expired_set(void *data)
1994 {
1995         struct ptlrpc_request_set       *set = data;
1996         struct list_head                *tmp;
1997         time_t                          now = cfs_time_current_sec();
1998         ENTRY;
1999
2000         LASSERT(set != NULL);
2001
2002         /*
2003          * A timeout expired. See which reqs it applies to...
2004          */
2005         list_for_each(tmp, &set->set_requests) {
2006                 struct ptlrpc_request *req =
2007                         list_entry(tmp, struct ptlrpc_request,
2008                                    rq_set_chain);
2009
2010                 /* don't expire request waiting for context */
2011                 if (req->rq_wait_ctx)
2012                         continue;
2013
2014                 /* Request in-flight? */
2015                 if (!((req->rq_phase == RQ_PHASE_RPC &&
2016                        !req->rq_waiting && !req->rq_resend) ||
2017                       (req->rq_phase == RQ_PHASE_BULK)))
2018                         continue;
2019
2020                 if (req->rq_timedout ||     /* already dealt with */
2021                     req->rq_deadline > now) /* not expired */
2022                         continue;
2023
2024                 /* Deal with this guy. Do it asynchronously to not block
2025                  * ptlrpcd thread. */
2026                 ptlrpc_expire_one_request(req, 1);
2027         }
2028
2029         /*
2030          * When waiting for a whole set, we always break out of the
2031          * sleep so we can recalculate the timeout, or enable interrupts
2032          * if everyone's timed out.
2033          */
2034         RETURN(1);
2035 }
2036
2037 /**
2038  * Sets rq_intr flag in \a req under spinlock.
2039  */
2040 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2041 {
2042         spin_lock(&req->rq_lock);
2043         req->rq_intr = 1;
2044         spin_unlock(&req->rq_lock);
2045 }
2046 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2047
2048 /**
2049  * Interrupts (sets interrupted flag) all uncompleted requests in
2050  * a set \a data. Callback for l_wait_event for interruptible waits.
2051  */
2052 static void ptlrpc_interrupted_set(void *data)
2053 {
2054         struct ptlrpc_request_set *set = data;
2055         struct list_head *tmp;
2056
2057         LASSERT(set != NULL);
2058         CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2059
2060         list_for_each(tmp, &set->set_requests) {
2061                 struct ptlrpc_request *req =
2062                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2063
2064                 if (req->rq_phase != RQ_PHASE_RPC &&
2065                     req->rq_phase != RQ_PHASE_UNREGISTERING)
2066                         continue;
2067
2068                 ptlrpc_mark_interrupted(req);
2069         }
2070 }
2071
2072 /**
2073  * Get the smallest timeout in the set; this does NOT set a timeout.
2074  */
2075 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2076 {
2077         struct list_head        *tmp;
2078         time_t                   now = cfs_time_current_sec();
2079         int                      timeout = 0;
2080         struct ptlrpc_request   *req;
2081         int                      deadline;
2082         ENTRY;
2083
2084         list_for_each(tmp, &set->set_requests) {
2085                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2086
2087                 /*
2088                  * Request in-flight?
2089                  */
2090                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2091                       (req->rq_phase == RQ_PHASE_BULK) ||
2092                       (req->rq_phase == RQ_PHASE_NEW)))
2093                         continue;
2094
2095                 /*
2096                  * Already timed out.
2097                  */
2098                 if (req->rq_timedout)
2099                         continue;
2100
2101                 /*
2102                  * Waiting for ctx.
2103                  */
2104                 if (req->rq_wait_ctx)
2105                         continue;
2106
2107                 if (req->rq_phase == RQ_PHASE_NEW)
2108                         deadline = req->rq_sent;
2109                 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2110                         deadline = req->rq_sent;
2111                 else
2112                         deadline = req->rq_sent + req->rq_timeout;
2113
2114                 if (deadline <= now)    /* actually expired already */
2115                         timeout = 1;    /* ASAP */
2116                 else if (timeout == 0 || timeout > deadline - now)
2117                         timeout = deadline - now;
2118         }
2119         RETURN(timeout);
2120 }
2121
2122 /**
2123  * Send all unset request from the set and then wait untill all
2124  * requests in the set complete (either get a reply, timeout, get an
2125  * error or otherwise be interrupted).
2126  * Returns 0 on success or error code otherwise.
2127  */
2128 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2129 {
2130         struct list_head            *tmp;
2131         struct ptlrpc_request *req;
2132         struct l_wait_info     lwi;
2133         int                    rc, timeout;
2134         ENTRY;
2135
2136         if (set->set_producer)
2137                 (void)ptlrpc_set_producer(set);
2138         else
2139                 list_for_each(tmp, &set->set_requests) {
2140                         req = list_entry(tmp, struct ptlrpc_request,
2141                                          rq_set_chain);
2142                         if (req->rq_phase == RQ_PHASE_NEW)
2143                                 (void)ptlrpc_send_new_req(req);
2144                 }
2145
2146         if (list_empty(&set->set_requests))
2147                 RETURN(0);
2148
2149         do {
2150                 timeout = ptlrpc_set_next_timeout(set);
2151
2152                 /* wait until all complete, interrupted, or an in-flight
2153                  * req times out */
2154                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2155                        set, timeout);
2156
2157                 if (timeout == 0 && !cfs_signal_pending())
2158                         /*
2159                          * No requests are in-flight (ether timed out
2160                          * or delayed), so we can allow interrupts.
2161                          * We still want to block for a limited time,
2162                          * so we allow interrupts during the timeout.
2163                          */
2164                         lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2165                                                    ptlrpc_expired_set,
2166                                                    ptlrpc_interrupted_set, set);
2167                 else
2168                         /*
2169                          * At least one request is in flight, so no
2170                          * interrupts are allowed. Wait until all
2171                          * complete, or an in-flight req times out.
2172                          */
2173                         lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2174                                           ptlrpc_expired_set, set);
2175
2176                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2177
2178                 /* LU-769 - if we ignored the signal because it was already
2179                  * pending when we started, we need to handle it now or we risk
2180                  * it being ignored forever */
2181                 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2182                     cfs_signal_pending()) {
2183                         sigset_t blocked_sigs =
2184                                            cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2185
2186                         /* In fact we only interrupt for the "fatal" signals
2187                          * like SIGINT or SIGKILL. We still ignore less
2188                          * important signals since ptlrpc set is not easily
2189                          * reentrant from userspace again */
2190                         if (cfs_signal_pending())
2191                                 ptlrpc_interrupted_set(set);
2192                         cfs_restore_sigs(blocked_sigs);
2193                 }
2194
2195                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2196
2197                 /* -EINTR => all requests have been flagged rq_intr so next
2198                  * check completes.
2199                  * -ETIMEDOUT => someone timed out.  When all reqs have
2200                  * timed out, signals are enabled allowing completion with
2201                  * EINTR.
2202                  * I don't really care if we go once more round the loop in
2203                  * the error cases -eeb. */
2204                 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2205                         list_for_each(tmp, &set->set_requests) {
2206                                 req = list_entry(tmp, struct ptlrpc_request,
2207                                                  rq_set_chain);
2208                                 spin_lock(&req->rq_lock);
2209                                 req->rq_invalid_rqset = 1;
2210                                 spin_unlock(&req->rq_lock);
2211                         }
2212                 }
2213         } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2214
2215         LASSERT(atomic_read(&set->set_remaining) == 0);
2216
2217         rc = set->set_rc; /* rq_status of already freed requests if any */
2218         list_for_each(tmp, &set->set_requests) {
2219                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2220
2221                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2222                 if (req->rq_status != 0)
2223                         rc = req->rq_status;
2224         }
2225
2226         if (set->set_interpret != NULL) {
2227                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2228                         set->set_interpret;
2229                 rc = interpreter (set, set->set_arg, rc);
2230         } else {
2231                 struct ptlrpc_set_cbdata *cbdata, *n;
2232                 int err;
2233
2234                 list_for_each_entry_safe(cbdata, n,
2235                                          &set->set_cblist, psc_item) {
2236                         list_del_init(&cbdata->psc_item);
2237                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2238                         if (err && !rc)
2239                                 rc = err;
2240                         OBD_FREE_PTR(cbdata);
2241                 }
2242         }
2243
2244         RETURN(rc);
2245 }
2246 EXPORT_SYMBOL(ptlrpc_set_wait);
2247
2248 /**
2249  * Helper fuction for request freeing.
2250  * Called when request count reached zero and request needs to be freed.
2251  * Removes request from all sorts of sending/replay lists it might be on,
2252  * frees network buffers if any are present.
2253  * If \a locked is set, that means caller is already holding import imp_lock
2254  * and so we no longer need to reobtain it (for certain lists manipulations)
2255  */
2256 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2257 {
2258         ENTRY;
2259
2260         if (request == NULL)
2261                 RETURN_EXIT;
2262
2263         LASSERT(!request->rq_srv_req);
2264         LASSERT(request->rq_export == NULL);
2265         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2266         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2267         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2268         LASSERTF(!request->rq_replay, "req %p\n", request);
2269
2270         req_capsule_fini(&request->rq_pill);
2271
2272         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
2273          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2274         if (request->rq_import != NULL) {
2275                 if (!locked)
2276                         spin_lock(&request->rq_import->imp_lock);
2277                 list_del_init(&request->rq_replay_list);
2278                 if (!locked)
2279                         spin_unlock(&request->rq_import->imp_lock);
2280         }
2281         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2282
2283         if (atomic_read(&request->rq_refcount) != 0) {
2284                 DEBUG_REQ(D_ERROR, request,
2285                           "freeing request with nonzero refcount");
2286                 LBUG();
2287         }
2288
2289         if (request->rq_repbuf != NULL)
2290                 sptlrpc_cli_free_repbuf(request);
2291
2292         if (request->rq_import != NULL) {
2293                 class_import_put(request->rq_import);
2294                 request->rq_import = NULL;
2295         }
2296         if (request->rq_bulk != NULL)
2297                 ptlrpc_free_bulk_pin(request->rq_bulk);
2298
2299         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2300                 sptlrpc_cli_free_reqbuf(request);
2301
2302         if (request->rq_cli_ctx)
2303                 sptlrpc_req_put_ctx(request, !locked);
2304
2305         if (request->rq_pool)
2306                 __ptlrpc_free_req_to_pool(request);
2307         else
2308                 ptlrpc_request_cache_free(request);
2309         EXIT;
2310 }
2311
2312 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2313 /**
2314  * Drop one request reference. Must be called with import imp_lock held.
2315  * When reference count drops to zero, request is freed.
2316  */
2317 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2318 {
2319         assert_spin_locked(&request->rq_import->imp_lock);
2320         (void)__ptlrpc_req_finished(request, 1);
2321 }
2322
2323 /**
2324  * Helper function
2325  * Drops one reference count for request \a request.
2326  * \a locked set indicates that caller holds import imp_lock.
2327  * Frees the request whe reference count reaches zero.
2328  */
2329 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2330 {
2331         ENTRY;
2332         if (request == NULL)
2333                 RETURN(1);
2334
2335         if (request == LP_POISON ||
2336             request->rq_reqmsg == LP_POISON) {
2337                 CERROR("dereferencing freed request (bug 575)\n");
2338                 LBUG();
2339                 RETURN(1);
2340         }
2341
2342         DEBUG_REQ(D_INFO, request, "refcount now %u",
2343                   atomic_read(&request->rq_refcount) - 1);
2344
2345         if (atomic_dec_and_test(&request->rq_refcount)) {
2346                 __ptlrpc_free_req(request, locked);
2347                 RETURN(1);
2348         }
2349
2350         RETURN(0);
2351 }
2352
2353 /**
2354  * Drops one reference count for a request.
2355  */
2356 void ptlrpc_req_finished(struct ptlrpc_request *request)
2357 {
2358         __ptlrpc_req_finished(request, 0);
2359 }
2360 EXPORT_SYMBOL(ptlrpc_req_finished);
2361
2362 /**
2363  * Returns xid of a \a request
2364  */
2365 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2366 {
2367         return request->rq_xid;
2368 }
2369 EXPORT_SYMBOL(ptlrpc_req_xid);
2370
2371 /**
2372  * Disengage the client's reply buffer from the network
2373  * NB does _NOT_ unregister any client-side bulk.
2374  * IDEMPOTENT, but _not_ safe against concurrent callers.
2375  * The request owner (i.e. the thread doing the I/O) must call...
2376  * Returns 0 on success or 1 if unregistering cannot be made.
2377  */
2378 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2379 {
2380         int                rc;
2381         struct l_wait_info lwi;
2382
2383         /*
2384          * Might sleep.
2385          */
2386         LASSERT(!in_interrupt());
2387
2388         /*
2389          * Let's setup deadline for reply unlink.
2390          */
2391         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2392             async && request->rq_reply_deadline == 0)
2393                 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2394
2395         /*
2396          * Nothing left to do.
2397          */
2398         if (!ptlrpc_client_recv_or_unlink(request))
2399                 RETURN(1);
2400
2401         LNetMDUnlink(request->rq_reply_md_h);
2402
2403         /*
2404          * Let's check it once again.
2405          */
2406         if (!ptlrpc_client_recv_or_unlink(request))
2407                 RETURN(1);
2408
2409         /*
2410          * Move to "Unregistering" phase as reply was not unlinked yet.
2411          */
2412         ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2413
2414         /*
2415          * Do not wait for unlink to finish.
2416          */
2417         if (async)
2418                 RETURN(0);
2419
2420         /*
2421          * We have to l_wait_event() whatever the result, to give liblustre
2422          * a chance to run reply_in_callback(), and to make sure we've
2423          * unlinked before returning a req to the pool.
2424          */
2425         for (;;) {
2426                 /* The wq argument is ignored by user-space wait_event macros */
2427                 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2428                                         &request->rq_set->set_waitq :
2429                                         &request->rq_reply_waitq;
2430                 /* Network access will complete in finite time but the HUGE
2431                  * timeout lets us CWARN for visibility of sluggish NALs */
2432                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2433                                            cfs_time_seconds(1), NULL, NULL);
2434                 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2435                                   &lwi);
2436                 if (rc == 0) {
2437                         ptlrpc_rqphase_move(request, request->rq_next_phase);
2438                         RETURN(1);
2439                 }
2440
2441                 LASSERT(rc == -ETIMEDOUT);
2442                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2443                           "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2444                           request->rq_receiving_reply,
2445                           request->rq_req_unlinked,
2446                           request->rq_reply_unlinked);
2447         }
2448         RETURN(0);
2449 }
2450
2451 static void ptlrpc_free_request(struct ptlrpc_request *req)
2452 {
2453         spin_lock(&req->rq_lock);
2454         req->rq_replay = 0;
2455         spin_unlock(&req->rq_lock);
2456
2457         if (req->rq_commit_cb != NULL)
2458                 req->rq_commit_cb(req);
2459         list_del_init(&req->rq_replay_list);
2460
2461         __ptlrpc_req_finished(req, 1);
2462 }
2463
2464 /**
2465  * the request is committed and dropped from the replay list of its import
2466  */
2467 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2468 {
2469         struct obd_import       *imp = req->rq_import;
2470
2471         spin_lock(&imp->imp_lock);
2472         if (list_empty(&req->rq_replay_list)) {
2473                 spin_unlock(&imp->imp_lock);
2474                 return;
2475         }
2476
2477         if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2478                 ptlrpc_free_request(req);
2479
2480         spin_unlock(&imp->imp_lock);
2481 }
2482 EXPORT_SYMBOL(ptlrpc_request_committed);
2483
2484 /**
2485  * Iterates through replay_list on import and prunes
2486  * all requests have transno smaller than last_committed for the
2487  * import and don't have rq_replay set.
2488  * Since requests are sorted in transno order, stops when meetign first
2489  * transno bigger than last_committed.
2490  * caller must hold imp->imp_lock
2491  */
2492 void ptlrpc_free_committed(struct obd_import *imp)
2493 {
2494         struct ptlrpc_request   *req, *saved;
2495         struct ptlrpc_request   *last_req = NULL; /* temporary fire escape */
2496         bool                     skip_committed_list = true;
2497         ENTRY;
2498
2499         LASSERT(imp != NULL);
2500         assert_spin_locked(&imp->imp_lock);
2501
2502         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2503             imp->imp_generation == imp->imp_last_generation_checked) {
2504                 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2505                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2506                 RETURN_EXIT;
2507         }
2508         CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2509                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2510                imp->imp_generation);
2511
2512         if (imp->imp_generation != imp->imp_last_generation_checked ||
2513             imp->imp_last_transno_checked == 0)
2514                 skip_committed_list = false;
2515
2516         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2517         imp->imp_last_generation_checked = imp->imp_generation;
2518
2519         list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2520                                      rq_replay_list) {
2521                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
2522                 LASSERT(req != last_req);
2523                 last_req = req;
2524
2525                 if (req->rq_transno == 0) {
2526                         DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2527                         LBUG();
2528                 }
2529                 if (req->rq_import_generation < imp->imp_generation) {
2530                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2531                         GOTO(free_req, 0);
2532                 }
2533
2534                 /* not yet committed */
2535                 if (req->rq_transno > imp->imp_peer_committed_transno) {
2536                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2537                         break;
2538                 }
2539
2540                 if (req->rq_replay) {
2541                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2542                         list_move_tail(&req->rq_replay_list,
2543                                            &imp->imp_committed_list);
2544                         continue;
2545                 }
2546
2547                 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2548                           imp->imp_peer_committed_transno);
2549 free_req:
2550                 ptlrpc_free_request(req);
2551         }
2552
2553         if (skip_committed_list)
2554                 GOTO(out, 0);
2555
2556         list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2557                                      rq_replay_list) {
2558                 LASSERT(req->rq_transno != 0);
2559                 if (req->rq_import_generation < imp->imp_generation) {
2560                         DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2561                         ptlrpc_free_request(req);
2562                 } else if (!req->rq_replay) {
2563                         DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2564                         ptlrpc_free_request(req);
2565                 }
2566         }
2567 out:
2568         EXIT;
2569 }
2570
2571 void ptlrpc_cleanup_client(struct obd_import *imp)
2572 {
2573         ENTRY;
2574         EXIT;
2575 }
2576
2577 /**
2578  * Schedule previously sent request for resend.
2579  * For bulk requests we assign new xid (to avoid problems with
2580  * lost replies and therefore several transfers landing into same buffer
2581  * from different sending attempts).
2582  */
2583 void ptlrpc_resend_req(struct ptlrpc_request *req)
2584 {
2585         DEBUG_REQ(D_HA, req, "going to resend");
2586         spin_lock(&req->rq_lock);
2587
2588         /* Request got reply but linked to the import list still.
2589            Let ptlrpc_check_set() to process it. */
2590         if (ptlrpc_client_replied(req)) {
2591                 spin_unlock(&req->rq_lock);
2592                 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2593                 return;
2594         }
2595
2596         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2597         req->rq_status = -EAGAIN;
2598
2599         req->rq_resend = 1;
2600         req->rq_net_err = 0;
2601         req->rq_timedout = 0;
2602         if (req->rq_bulk) {
2603                 __u64 old_xid = req->rq_xid;
2604
2605                 /* ensure previous bulk fails */
2606                 req->rq_xid = ptlrpc_next_xid();
2607                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2608                        old_xid, req->rq_xid);
2609         }
2610         ptlrpc_client_wake_req(req);
2611         spin_unlock(&req->rq_lock);
2612 }
2613
2614 /* XXX: this function and rq_status are currently unused */
2615 void ptlrpc_restart_req(struct ptlrpc_request *req)
2616 {
2617         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2618         req->rq_status = -ERESTARTSYS;
2619
2620         spin_lock(&req->rq_lock);
2621         req->rq_restart = 1;
2622         req->rq_timedout = 0;
2623         ptlrpc_client_wake_req(req);
2624         spin_unlock(&req->rq_lock);
2625 }
2626
2627 /**
2628  * Grab additional reference on a request \a req
2629  */
2630 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2631 {
2632         ENTRY;
2633         atomic_inc(&req->rq_refcount);
2634         RETURN(req);
2635 }
2636 EXPORT_SYMBOL(ptlrpc_request_addref);
2637
2638 /**
2639  * Add a request to import replay_list.
2640  * Must be called under imp_lock
2641  */
2642 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2643                                       struct obd_import *imp)
2644 {
2645         struct list_head *tmp;
2646
2647         assert_spin_locked(&imp->imp_lock);
2648
2649         if (req->rq_transno == 0) {
2650                 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2651                 LBUG();
2652         }
2653
2654         /* clear this for new requests that were resent as well
2655            as resent replayed requests. */
2656         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2657
2658         /* don't re-add requests that have been replayed */
2659         if (!list_empty(&req->rq_replay_list))
2660                 return;
2661
2662         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2663
2664         LASSERT(imp->imp_replayable);
2665         /* Balanced in ptlrpc_free_committed, usually. */
2666         ptlrpc_request_addref(req);
2667         list_for_each_prev(tmp, &imp->imp_replay_list) {
2668                 struct ptlrpc_request *iter = list_entry(tmp,
2669                                                          struct ptlrpc_request,
2670                                                          rq_replay_list);
2671
2672                 /* We may have duplicate transnos if we create and then
2673                  * open a file, or for closes retained if to match creating
2674                  * opens, so use req->rq_xid as a secondary key.
2675                  * (See bugs 684, 685, and 428.)
2676                  * XXX no longer needed, but all opens need transnos!
2677                  */
2678                 if (iter->rq_transno > req->rq_transno)
2679                         continue;
2680
2681                 if (iter->rq_transno == req->rq_transno) {
2682                         LASSERT(iter->rq_xid != req->rq_xid);
2683                         if (iter->rq_xid > req->rq_xid)
2684                                 continue;
2685                 }
2686
2687                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2688                 return;
2689         }
2690
2691         list_add(&req->rq_replay_list, &imp->imp_replay_list);
2692 }
2693
2694 /**
2695  * Send request and wait until it completes.
2696  * Returns request processing status.
2697  */
2698 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2699 {
2700         struct ptlrpc_request_set *set;
2701         int rc;
2702         ENTRY;
2703
2704         LASSERT(req->rq_set == NULL);
2705         LASSERT(!req->rq_receiving_reply);
2706
2707         set = ptlrpc_prep_set();
2708         if (set == NULL) {
2709                 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2710                 RETURN(-ENOMEM);
2711         }
2712
2713         /* for distributed debugging */
2714         lustre_msg_set_status(req->rq_reqmsg, current_pid());
2715
2716         /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2717         ptlrpc_request_addref(req);
2718         ptlrpc_set_add_req(set, req);
2719         rc = ptlrpc_set_wait(set);
2720         ptlrpc_set_destroy(set);
2721
2722         RETURN(rc);
2723 }
2724 EXPORT_SYMBOL(ptlrpc_queue_wait);
2725
2726 /**
2727  * Callback used for replayed requests reply processing.
2728  * In case of successful reply calls registered request replay callback.
2729  * In case of error restart replay process.
2730  */
2731 static int ptlrpc_replay_interpret(const struct lu_env *env,
2732                                    struct ptlrpc_request *req,
2733                                    void * data, int rc)
2734 {
2735         struct ptlrpc_replay_async_args *aa = data;
2736         struct obd_import *imp = req->rq_import;
2737
2738         ENTRY;
2739         atomic_dec(&imp->imp_replay_inflight);
2740
2741         if (!ptlrpc_client_replied(req)) {
2742                 CERROR("request replay timed out, restarting recovery\n");
2743                 GOTO(out, rc = -ETIMEDOUT);
2744         }
2745
2746         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2747             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2748              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2749                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2750
2751         /** VBR: check version failure */
2752         if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2753                 /** replay was failed due to version mismatch */
2754                 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2755                 spin_lock(&imp->imp_lock);
2756                 imp->imp_vbr_failed = 1;
2757                 imp->imp_no_lock_replay = 1;
2758                 spin_unlock(&imp->imp_lock);
2759                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2760         } else {
2761                 /** The transno had better not change over replay. */
2762                 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2763                          lustre_msg_get_transno(req->rq_repmsg) ||
2764                          lustre_msg_get_transno(req->rq_repmsg) == 0,
2765                          LPX64"/"LPX64"\n",
2766                          lustre_msg_get_transno(req->rq_reqmsg),
2767                          lustre_msg_get_transno(req->rq_repmsg));
2768         }
2769
2770         spin_lock(&imp->imp_lock);
2771         /** if replays by version then gap occur on server, no trust to locks */
2772         if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2773                 imp->imp_no_lock_replay = 1;
2774         imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2775         spin_unlock(&imp->imp_lock);
2776         LASSERT(imp->imp_last_replay_transno);
2777
2778         /* transaction number shouldn't be bigger than the latest replayed */
2779         if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2780                 DEBUG_REQ(D_ERROR, req,
2781                           "Reported transno "LPU64" is bigger than the "
2782                           "replayed one: "LPU64, req->rq_transno,
2783                           lustre_msg_get_transno(req->rq_reqmsg));
2784                 GOTO(out, rc = -EINVAL);
2785         }
2786
2787         DEBUG_REQ(D_HA, req, "got rep");
2788
2789         /* let the callback do fixups, possibly including in the request */
2790         if (req->rq_replay_cb)
2791                 req->rq_replay_cb(req);
2792
2793         if (ptlrpc_client_replied(req) &&
2794             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2795                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2796                           lustre_msg_get_status(req->rq_repmsg),
2797                           aa->praa_old_status);
2798         } else {
2799                 /* Put it back for re-replay. */
2800                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2801         }
2802
2803         /*
2804          * Errors while replay can set transno to 0, but
2805          * imp_last_replay_transno shouldn't be set to 0 anyway
2806          */
2807         if (req->rq_transno == 0)
2808                 CERROR("Transno is 0 during replay!\n");
2809
2810         /* continue with recovery */
2811         rc = ptlrpc_import_recovery_state_machine(imp);
2812  out:
2813         req->rq_send_state = aa->praa_old_state;
2814
2815         if (rc != 0)
2816                 /* this replay failed, so restart recovery */
2817                 ptlrpc_connect_import(imp);
2818
2819         RETURN(rc);
2820 }
2821
2822 /**
2823  * Prepares and queues request for replay.
2824  * Adds it to ptlrpcd queue for actual sending.
2825  * Returns 0 on success.
2826  */
2827 int ptlrpc_replay_req(struct ptlrpc_request *req)
2828 {
2829         struct ptlrpc_replay_async_args *aa;
2830         ENTRY;
2831
2832         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2833
2834         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2835         aa = ptlrpc_req_async_args(req);
2836         memset(aa, 0, sizeof *aa);
2837
2838         /* Prepare request to be resent with ptlrpcd */
2839         aa->praa_old_state = req->rq_send_state;
2840         req->rq_send_state = LUSTRE_IMP_REPLAY;
2841         req->rq_phase = RQ_PHASE_NEW;
2842         req->rq_next_phase = RQ_PHASE_UNDEFINED;
2843         if (req->rq_repmsg)
2844                 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2845         req->rq_status = 0;
2846         req->rq_interpret_reply = ptlrpc_replay_interpret;
2847         /* Readjust the timeout for current conditions */
2848         ptlrpc_at_set_req_timeout(req);
2849
2850         /* Tell server the net_latency, so the server can calculate how long
2851          * it should wait for next replay */
2852         lustre_msg_set_service_time(req->rq_reqmsg,
2853                                     ptlrpc_at_get_net_latency(req));
2854         DEBUG_REQ(D_HA, req, "REPLAY");
2855
2856         atomic_inc(&req->rq_import->imp_replay_inflight);
2857         ptlrpc_request_addref(req);     /* ptlrpcd needs a ref */
2858
2859         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2860         RETURN(0);
2861 }
2862
2863 /**
2864  * Aborts all in-flight request on import \a imp sending and delayed lists
2865  */
2866 void ptlrpc_abort_inflight(struct obd_import *imp)
2867 {
2868         struct list_head *tmp, *n;
2869         ENTRY;
2870
2871         /* Make sure that no new requests get processed for this import.
2872          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2873          * this flag and then putting requests on sending_list or delayed_list.
2874          */
2875         spin_lock(&imp->imp_lock);
2876
2877         /* XXX locking?  Maybe we should remove each request with the list
2878          * locked?  Also, how do we know if the requests on the list are
2879          * being freed at this time?
2880          */
2881         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2882                 struct ptlrpc_request *req = list_entry(tmp,
2883                                                         struct ptlrpc_request,
2884                                                         rq_list);
2885
2886                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2887
2888                 spin_lock(&req->rq_lock);
2889                 if (req->rq_import_generation < imp->imp_generation) {
2890                         req->rq_err = 1;
2891                         req->rq_status = -EIO;
2892                         ptlrpc_client_wake_req(req);
2893                 }
2894                 spin_unlock(&req->rq_lock);
2895         }
2896
2897         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2898                 struct ptlrpc_request *req =
2899                         list_entry(tmp, struct ptlrpc_request, rq_list);
2900
2901                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2902
2903                 spin_lock(&req->rq_lock);
2904                 if (req->rq_import_generation < imp->imp_generation) {
2905                         req->rq_err = 1;
2906                         req->rq_status = -EIO;
2907                         ptlrpc_client_wake_req(req);
2908                 }
2909                 spin_unlock(&req->rq_lock);
2910         }
2911
2912         /* Last chance to free reqs left on the replay list, but we
2913          * will still leak reqs that haven't committed.  */
2914         if (imp->imp_replayable)
2915                 ptlrpc_free_committed(imp);
2916
2917         spin_unlock(&imp->imp_lock);
2918
2919         EXIT;
2920 }
2921
2922 /**
2923  * Abort all uncompleted requests in request set \a set
2924  */
2925 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2926 {
2927         struct list_head *tmp, *pos;
2928
2929         LASSERT(set != NULL);
2930
2931         list_for_each_safe(pos, tmp, &set->set_requests) {
2932                 struct ptlrpc_request *req =
2933                         list_entry(pos, struct ptlrpc_request,
2934                                    rq_set_chain);
2935
2936                 spin_lock(&req->rq_lock);
2937                 if (req->rq_phase != RQ_PHASE_RPC) {
2938                         spin_unlock(&req->rq_lock);
2939                         continue;
2940                 }
2941
2942                 req->rq_err = 1;
2943                 req->rq_status = -EINTR;
2944                 ptlrpc_client_wake_req(req);
2945                 spin_unlock(&req->rq_lock);
2946         }
2947 }
2948
2949 static __u64 ptlrpc_last_xid;
2950 static spinlock_t ptlrpc_last_xid_lock;
2951
2952 /**
2953  * Initialize the XID for the node.  This is common among all requests on
2954  * this node, and only requires the property that it is monotonically
2955  * increasing.  It does not need to be sequential.  Since this is also used
2956  * as the RDMA match bits, it is important that a single client NOT have
2957  * the same match bits for two different in-flight requests, hence we do
2958  * NOT want to have an XID per target or similar.
2959  *
2960  * To avoid an unlikely collision between match bits after a client reboot
2961  * (which would deliver old data into the wrong RDMA buffer) initialize
2962  * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2963  * If the time is clearly incorrect, we instead use a 62-bit random number.
2964  * In the worst case the random number will overflow 1M RPCs per second in
2965  * 9133 years, or permutations thereof.
2966  */
2967 #define YEAR_2004 (1ULL << 30)
2968 void ptlrpc_init_xid(void)
2969 {
2970         time_t now = cfs_time_current_sec();
2971
2972         spin_lock_init(&ptlrpc_last_xid_lock);
2973         if (now < YEAR_2004) {
2974                 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2975                 ptlrpc_last_xid >>= 2;
2976                 ptlrpc_last_xid |= (1ULL << 61);
2977         } else {
2978                 ptlrpc_last_xid = (__u64)now << 20;
2979         }
2980
2981         /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
2982         CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
2983         ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
2984 }
2985
2986 /**
2987  * Increase xid and returns resulting new value to the caller.
2988  *
2989  * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2990  * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2991  * itself uses the last bulk xid needed, so the server can determine the
2992  * the number of bulk transfers from the RPC XID and a bitmask.  The starting
2993  * xid must align to a power-of-two value.
2994  *
2995  * This is assumed to be true due to the initial ptlrpc_last_xid
2996  * value also being initialized to a power-of-two value. LU-1431
2997  */
2998 __u64 ptlrpc_next_xid(void)
2999 {
3000         __u64 next;
3001
3002         spin_lock(&ptlrpc_last_xid_lock);
3003         next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3004         ptlrpc_last_xid = next;
3005         spin_unlock(&ptlrpc_last_xid_lock);
3006
3007         return next;
3008 }
3009
3010 /**
3011  * Get a glimpse at what next xid value might have been.
3012  * Returns possible next xid.
3013  */
3014 __u64 ptlrpc_sample_next_xid(void)
3015 {
3016 #if BITS_PER_LONG == 32
3017         /* need to avoid possible word tearing on 32-bit systems */
3018         __u64 next;
3019
3020         spin_lock(&ptlrpc_last_xid_lock);
3021         next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3022         spin_unlock(&ptlrpc_last_xid_lock);
3023
3024         return next;
3025 #else
3026         /* No need to lock, since returned value is racy anyways */
3027         return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3028 #endif
3029 }
3030 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3031
3032 /**
3033  * Functions for operating ptlrpc workers.
3034  *
3035  * A ptlrpc work is a function which will be running inside ptlrpc context.
3036  * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3037  *
3038  * 1. after a work is created, it can be used many times, that is:
3039  *         handler = ptlrpcd_alloc_work();
3040  *         ptlrpcd_queue_work();
3041  *
3042  *    queue it again when necessary:
3043  *         ptlrpcd_queue_work();
3044  *         ptlrpcd_destroy_work();
3045  * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3046  *    it will only be queued once in any time. Also as its name implies, it may
3047  *    have delay before it really runs by ptlrpcd thread.
3048  */
3049 struct ptlrpc_work_async_args {
3050         int   (*cb)(const struct lu_env *, void *);
3051         void   *cbdata;
3052 };
3053
3054 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3055 {
3056         /* re-initialize the req */
3057         req->rq_timeout         = obd_timeout;
3058         req->rq_sent            = cfs_time_current_sec();
3059         req->rq_deadline        = req->rq_sent + req->rq_timeout;
3060         req->rq_reply_deadline  = req->rq_deadline;
3061         req->rq_phase           = RQ_PHASE_INTERPRET;
3062         req->rq_next_phase      = RQ_PHASE_COMPLETE;
3063         req->rq_xid             = ptlrpc_next_xid();
3064         req->rq_import_generation = req->rq_import->imp_generation;
3065
3066         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3067 }
3068
3069 static int work_interpreter(const struct lu_env *env,
3070                             struct ptlrpc_request *req, void *data, int rc)
3071 {
3072         struct ptlrpc_work_async_args *arg = data;
3073
3074         LASSERT(ptlrpcd_check_work(req));
3075         LASSERT(arg->cb != NULL);
3076
3077         rc = arg->cb(env, arg->cbdata);
3078
3079         list_del_init(&req->rq_set_chain);
3080         req->rq_set = NULL;
3081
3082         if (atomic_dec_return(&req->rq_refcount) > 1) {
3083                 atomic_set(&req->rq_refcount, 2);
3084                 ptlrpcd_add_work_req(req);
3085         }
3086         return rc;
3087 }
3088
3089 static int worker_format;
3090
3091 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3092 {
3093         return req->rq_pill.rc_fmt == (void *)&worker_format;
3094 }
3095
3096 /**
3097  * Create a work for ptlrpc.
3098  */
3099 void *ptlrpcd_alloc_work(struct obd_import *imp,
3100                          int (*cb)(const struct lu_env *, void *), void *cbdata)
3101 {
3102         struct ptlrpc_request         *req = NULL;
3103         struct ptlrpc_work_async_args *args;
3104         ENTRY;
3105
3106         might_sleep();
3107
3108         if (cb == NULL)
3109                 RETURN(ERR_PTR(-EINVAL));
3110
3111         /* copy some code from deprecated fakereq. */
3112         req = ptlrpc_request_cache_alloc(GFP_NOFS);
3113         if (req == NULL) {
3114                 CERROR("ptlrpc: run out of memory!\n");
3115                 RETURN(ERR_PTR(-ENOMEM));
3116         }
3117
3118         ptlrpc_cli_req_init(req);
3119
3120         req->rq_send_state = LUSTRE_IMP_FULL;
3121         req->rq_type = PTL_RPC_MSG_REQUEST;
3122         req->rq_import = class_import_get(imp);
3123         req->rq_interpret_reply = work_interpreter;
3124         /* don't want reply */
3125         req->rq_no_delay = req->rq_no_resend = 1;
3126         req->rq_pill.rc_fmt = (void *)&worker_format;
3127
3128         CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3129         args = ptlrpc_req_async_args(req);
3130         args->cb     = cb;
3131         args->cbdata = cbdata;
3132
3133         RETURN(req);
3134 }
3135 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3136
3137 void ptlrpcd_destroy_work(void *handler)
3138 {
3139         struct ptlrpc_request *req = handler;
3140
3141         if (req)
3142                 ptlrpc_req_finished(req);
3143 }
3144 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3145
3146 int ptlrpcd_queue_work(void *handler)
3147 {
3148         struct ptlrpc_request *req = handler;
3149
3150         /*
3151          * Check if the req is already being queued.
3152          *
3153          * Here comes a trick: it lacks a way of checking if a req is being
3154          * processed reliably in ptlrpc. Here I have to use refcount of req
3155          * for this purpose. This is okay because the caller should use this
3156          * req as opaque data. - Jinshan
3157          */
3158         LASSERT(atomic_read(&req->rq_refcount) > 0);
3159         if (atomic_inc_return(&req->rq_refcount) == 2)
3160                 ptlrpcd_add_work_req(req);
3161         return 0;
3162 }
3163 EXPORT_SYMBOL(ptlrpcd_queue_work);