Whamcloud - gitweb
LU-6409 nodemap: don't destroy nodemaps in cfs_hash_for_each
[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 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
51         .add_kiov_frag  = ptlrpc_prep_bulk_page_pin,
52         .release_frags  = ptlrpc_release_bulk_page_pin,
53 };
54 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
55
56 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
57         .add_kiov_frag  = ptlrpc_prep_bulk_page_nopin,
58         .release_frags  = ptlrpc_release_bulk_noop,
59 };
60 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
61
62 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
63 static int ptlrpcd_check_work(struct ptlrpc_request *req);
64 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
65
66 /**
67  * Initialize passed in client structure \a cl.
68  */
69 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
70                         struct ptlrpc_client *cl)
71 {
72         cl->cli_request_portal = req_portal;
73         cl->cli_reply_portal   = rep_portal;
74         cl->cli_name           = name;
75 }
76 EXPORT_SYMBOL(ptlrpc_init_client);
77
78 /**
79  * Return PortalRPC connection for remore uud \a uuid
80  */
81 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
82 {
83         struct ptlrpc_connection *c;
84         lnet_nid_t                self;
85         lnet_process_id_t         peer;
86         int                       err;
87
88         /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
89          * before accessing its values. */
90         /* coverity[uninit_use_in_call] */
91         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
92         if (err != 0) {
93                 CNETERR("cannot find peer %s!\n", uuid->uuid);
94                 return NULL;
95         }
96
97         c = ptlrpc_connection_get(peer, self, uuid);
98         if (c) {
99                 memcpy(c->c_remote_uuid.uuid,
100                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
101         }
102
103         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
104
105         return c;
106 }
107
108 /**
109  * Allocate and initialize new bulk descriptor on the sender.
110  * Returns pointer to the descriptor or NULL on error.
111  */
112 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
113                                          enum ptlrpc_bulk_op_type type,
114                                          unsigned portal,
115                                          const struct ptlrpc_bulk_frag_ops *ops)
116 {
117         struct ptlrpc_bulk_desc *desc;
118         int i;
119
120         /* ensure that only one of KIOV or IOVEC is set but not both */
121         LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
122                  ops->add_kiov_frag != NULL) ||
123                 (ptlrpc_is_bulk_desc_kvec(type) &&
124                  ops->add_iov_frag != NULL));
125
126         if (type & PTLRPC_BULK_BUF_KIOV) {
127                 OBD_ALLOC(desc,
128                           offsetof(struct ptlrpc_bulk_desc,
129                                    bd_u.bd_kiov.bd_vec[nfrags]));
130         } else {
131                 OBD_ALLOC(desc,
132                           offsetof(struct ptlrpc_bulk_desc,
133                                    bd_u.bd_kvec.bd_kvec[nfrags]));
134         }
135
136         if (!desc)
137                 return NULL;
138
139         spin_lock_init(&desc->bd_lock);
140         init_waitqueue_head(&desc->bd_waitq);
141         desc->bd_max_iov = nfrags;
142         desc->bd_iov_count = 0;
143         desc->bd_portal = portal;
144         desc->bd_type = type;
145         desc->bd_md_count = 0;
146         desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
147         LASSERT(max_brw > 0);
148         desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
149         /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
150          * node. Negotiated ocd_brw_size will always be <= this number. */
151         for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
152                 LNetInvalidateHandle(&desc->bd_mds[i]);
153
154         return desc;
155 }
156
157 /**
158  * Prepare bulk descriptor for specified outgoing request \a req that
159  * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
160  * the bulk to be sent. Used on client-side.
161  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
162  * error.
163  */
164 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
165                                               unsigned nfrags, unsigned max_brw,
166                                               unsigned int type,
167                                               unsigned portal,
168                                               const struct ptlrpc_bulk_frag_ops
169                                                 *ops)
170 {
171         struct obd_import *imp = req->rq_import;
172         struct ptlrpc_bulk_desc *desc;
173
174         ENTRY;
175         LASSERT(ptlrpc_is_bulk_op_passive(type));
176
177         desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
178         if (desc == NULL)
179                 RETURN(NULL);
180
181         desc->bd_import_generation = req->rq_import_generation;
182         desc->bd_import = class_import_get(imp);
183         desc->bd_req = req;
184
185         desc->bd_cbid.cbid_fn  = client_bulk_callback;
186         desc->bd_cbid.cbid_arg = desc;
187
188         /* This makes req own desc, and free it when she frees herself */
189         req->rq_bulk = desc;
190
191         return desc;
192 }
193 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
194
195 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
196                              struct page *page, int pageoffset, int len,
197                              int pin)
198 {
199         lnet_kiov_t *kiov;
200
201         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
202         LASSERT(page != NULL);
203         LASSERT(pageoffset >= 0);
204         LASSERT(len > 0);
205         LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
206         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
207
208         kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
209
210         desc->bd_nob += len;
211
212         if (pin)
213                 page_cache_get(page);
214
215         kiov->kiov_page = page;
216         kiov->kiov_offset = pageoffset;
217         kiov->kiov_len = len;
218
219         desc->bd_iov_count++;
220 }
221 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
222
223 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
224                           void *frag, int len)
225 {
226         struct kvec *iovec;
227
228         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
229         LASSERT(frag != NULL);
230         LASSERT(len > 0);
231         LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
232
233         iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
234
235         desc->bd_nob += len;
236
237         iovec->iov_base = frag;
238         iovec->iov_len = len;
239
240         desc->bd_iov_count++;
241
242         return desc->bd_nob;
243 }
244 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
245
246 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
247 {
248         ENTRY;
249
250         LASSERT(desc != NULL);
251         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
252         LASSERT(desc->bd_md_count == 0);         /* network hands off */
253         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
254         LASSERT(desc->bd_frag_ops != NULL);
255
256         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
257                 sptlrpc_enc_pool_put_pages(desc);
258
259         if (desc->bd_export)
260                 class_export_put(desc->bd_export);
261         else
262                 class_import_put(desc->bd_import);
263
264         if (desc->bd_frag_ops->release_frags != NULL)
265                 desc->bd_frag_ops->release_frags(desc);
266
267         if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
268                 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
269                                         bd_u.bd_kiov.bd_vec[desc->bd_max_iov]));
270         else
271                 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
272                                         bd_u.bd_kvec.bd_kvec[desc->
273                                                 bd_max_iov]));
274
275         EXIT;
276 }
277 EXPORT_SYMBOL(ptlrpc_free_bulk);
278
279 /**
280  * Set server timelimit for this req, i.e. how long are we willing to wait
281  * for reply before timing out this request.
282  */
283 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
284 {
285         __u32 serv_est;
286         int idx;
287         struct imp_at *at;
288
289         LASSERT(req->rq_import);
290
291         if (AT_OFF) {
292                 /* non-AT settings */
293                 /**
294                  * \a imp_server_timeout means this is reverse import and
295                  * we send (currently only) ASTs to the client and cannot afford
296                  * to wait too long for the reply, otherwise the other client
297                  * (because of which we are sending this request) would
298                  * timeout waiting for us
299                  */
300                 req->rq_timeout = req->rq_import->imp_server_timeout ?
301                                   obd_timeout / 2 : obd_timeout;
302         } else {
303                 at = &req->rq_import->imp_at;
304                 idx = import_at_get_index(req->rq_import,
305                                           req->rq_request_portal);
306                 serv_est = at_get(&at->iat_service_estimate[idx]);
307                 req->rq_timeout = at_est2timeout(serv_est);
308         }
309         /* We could get even fancier here, using history to predict increased
310            loading... */
311
312         /* Let the server know what this RPC timeout is by putting it in the
313            reqmsg*/
314         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
315 }
316 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
317
318 /* Adjust max service estimate based on server value */
319 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
320                                   unsigned int serv_est)
321 {
322         int idx;
323         unsigned int oldse;
324         struct imp_at *at;
325
326         LASSERT(req->rq_import);
327         at = &req->rq_import->imp_at;
328
329         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
330         /* max service estimates are tracked on the server side,
331            so just keep minimal history here */
332         oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
333         if (oldse != 0)
334                 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
335                        "has changed from %d to %d\n",
336                        req->rq_import->imp_obd->obd_name,req->rq_request_portal,
337                        oldse, at_get(&at->iat_service_estimate[idx]));
338 }
339
340 /* Expected network latency per remote node (secs) */
341 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
342 {
343         return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
344 }
345
346 /* Adjust expected network latency */
347 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
348                                unsigned int service_time)
349 {
350         unsigned int nl, oldnl;
351         struct imp_at *at;
352         time_t now = cfs_time_current_sec();
353
354         LASSERT(req->rq_import);
355
356         if (service_time > now - req->rq_sent + 3) {
357                 /* bz16408, however, this can also happen if early reply
358                  * is lost and client RPC is expired and resent, early reply
359                  * or reply of original RPC can still be fit in reply buffer
360                  * of resent RPC, now client is measuring time from the
361                  * resent time, but server sent back service time of original
362                  * RPC.
363                  */
364                 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
365                        D_ADAPTTO : D_WARNING,
366                        "Reported service time %u > total measured time "
367                        CFS_DURATION_T"\n", service_time,
368                        cfs_time_sub(now, req->rq_sent));
369                 return;
370         }
371
372         /* Network latency is total time less server processing time */
373         nl = max_t(int, now - req->rq_sent -
374                         service_time, 0) + 1; /* st rounding */
375         at = &req->rq_import->imp_at;
376
377         oldnl = at_measured(&at->iat_net_latency, nl);
378         if (oldnl != 0)
379                 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
380                        "has changed from %d to %d\n",
381                        req->rq_import->imp_obd->obd_name,
382                        obd_uuid2str(
383                                &req->rq_import->imp_connection->c_remote_uuid),
384                        oldnl, at_get(&at->iat_net_latency));
385 }
386
387 static int unpack_reply(struct ptlrpc_request *req)
388 {
389         int rc;
390
391         if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
392                 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
393                 if (rc) {
394                         DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
395                         return(-EPROTO);
396                 }
397         }
398
399         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
400         if (rc) {
401                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
402                 return(-EPROTO);
403         }
404         return 0;
405 }
406
407 /**
408  * Handle an early reply message, called with the rq_lock held.
409  * If anything goes wrong just ignore it - same as if it never happened
410  */
411 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
412 __must_hold(&req->rq_lock)
413 {
414         struct ptlrpc_request *early_req;
415         time_t                 olddl;
416         int                    rc;
417         ENTRY;
418
419         req->rq_early = 0;
420         spin_unlock(&req->rq_lock);
421
422         rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
423         if (rc) {
424                 spin_lock(&req->rq_lock);
425                 RETURN(rc);
426         }
427
428         rc = unpack_reply(early_req);
429         if (rc != 0) {
430                 sptlrpc_cli_finish_early_reply(early_req);
431                 spin_lock(&req->rq_lock);
432                 RETURN(rc);
433         }
434
435         /* Use new timeout value just to adjust the local value for this
436          * request, don't include it into at_history. It is unclear yet why
437          * service time increased and should it be counted or skipped, e.g.
438          * that can be recovery case or some error or server, the real reply
439          * will add all new data if it is worth to add. */
440         req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
441         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
442
443         /* Network latency can be adjusted, it is pure network delays */
444         ptlrpc_at_adj_net_latency(req,
445                         lustre_msg_get_service_time(early_req->rq_repmsg));
446
447         sptlrpc_cli_finish_early_reply(early_req);
448
449         spin_lock(&req->rq_lock);
450         olddl = req->rq_deadline;
451         /* server assumes it now has rq_timeout from when the request
452          * arrived, so the client should give it at least that long.
453          * since we don't know the arrival time we'll use the original
454          * sent time */
455         req->rq_deadline = req->rq_sent + req->rq_timeout +
456                            ptlrpc_at_get_net_latency(req);
457
458         DEBUG_REQ(D_ADAPTTO, req,
459                   "Early reply #%d, new deadline in "CFS_DURATION_T"s "
460                   "("CFS_DURATION_T"s)", req->rq_early_count,
461                   cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
462                   cfs_time_sub(req->rq_deadline, olddl));
463
464         RETURN(rc);
465 }
466
467 static struct kmem_cache *request_cache;
468
469 int ptlrpc_request_cache_init(void)
470 {
471         request_cache = kmem_cache_create("ptlrpc_cache",
472                                           sizeof(struct ptlrpc_request),
473                                           0, SLAB_HWCACHE_ALIGN, NULL);
474         return request_cache == NULL ? -ENOMEM : 0;
475 }
476
477 void ptlrpc_request_cache_fini(void)
478 {
479         kmem_cache_destroy(request_cache);
480 }
481
482 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
483 {
484         struct ptlrpc_request *req;
485
486         OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
487         return req;
488 }
489
490 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
491 {
492         OBD_SLAB_FREE_PTR(req, request_cache);
493 }
494
495 /**
496  * Wind down request pool \a pool.
497  * Frees all requests from the pool too
498  */
499 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
500 {
501         struct list_head *l, *tmp;
502         struct ptlrpc_request *req;
503
504         LASSERT(pool != NULL);
505
506         spin_lock(&pool->prp_lock);
507         list_for_each_safe(l, tmp, &pool->prp_req_list) {
508                 req = list_entry(l, struct ptlrpc_request, rq_list);
509                 list_del(&req->rq_list);
510                 LASSERT(req->rq_reqbuf);
511                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
512                 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
513                 ptlrpc_request_cache_free(req);
514         }
515         spin_unlock(&pool->prp_lock);
516         OBD_FREE(pool, sizeof(*pool));
517 }
518 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
519
520 /**
521  * Allocates, initializes and adds \a num_rq requests to the pool \a pool
522  */
523 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
524 {
525         int i;
526         int size = 1;
527
528         while (size < pool->prp_rq_size)
529                 size <<= 1;
530
531         LASSERTF(list_empty(&pool->prp_req_list) ||
532                  size == pool->prp_rq_size,
533                  "Trying to change pool size with nonempty pool "
534                  "from %d to %d bytes\n", pool->prp_rq_size, size);
535
536         spin_lock(&pool->prp_lock);
537         pool->prp_rq_size = size;
538         for (i = 0; i < num_rq; i++) {
539                 struct ptlrpc_request *req;
540                 struct lustre_msg *msg;
541
542                 spin_unlock(&pool->prp_lock);
543                 req = ptlrpc_request_cache_alloc(GFP_NOFS);
544                 if (!req)
545                         return;
546                 OBD_ALLOC_LARGE(msg, size);
547                 if (!msg) {
548                         ptlrpc_request_cache_free(req);
549                         return;
550                 }
551                 req->rq_reqbuf = msg;
552                 req->rq_reqbuf_len = size;
553                 req->rq_pool = pool;
554                 spin_lock(&pool->prp_lock);
555                 list_add_tail(&req->rq_list, &pool->prp_req_list);
556         }
557         spin_unlock(&pool->prp_lock);
558         return;
559 }
560 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
561
562 /**
563  * Create and initialize new request pool with given attributes:
564  * \a num_rq - initial number of requests to create for the pool
565  * \a msgsize - maximum message size possible for requests in thid pool
566  * \a populate_pool - function to be called when more requests need to be added
567  *                    to the pool
568  * Returns pointer to newly created pool or NULL on error.
569  */
570 struct ptlrpc_request_pool *
571 ptlrpc_init_rq_pool(int num_rq, int msgsize,
572                     void (*populate_pool)(struct ptlrpc_request_pool *, int))
573 {
574         struct ptlrpc_request_pool *pool;
575
576         OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
577         if (!pool)
578                 return NULL;
579
580         /* Request next power of two for the allocation, because internally
581            kernel would do exactly this */
582
583         spin_lock_init(&pool->prp_lock);
584         INIT_LIST_HEAD(&pool->prp_req_list);
585         pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
586         pool->prp_populate = populate_pool;
587
588         populate_pool(pool, num_rq);
589
590         if (list_empty(&pool->prp_req_list)) {
591                 /* have not allocated a single request for the pool */
592                 OBD_FREE(pool, sizeof(struct ptlrpc_request_pool));
593                 pool = NULL;
594         }
595         return pool;
596 }
597 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
598
599 /**
600  * Fetches one request from pool \a pool
601  */
602 static struct ptlrpc_request *
603 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
604 {
605         struct ptlrpc_request *request;
606         struct lustre_msg *reqbuf;
607
608         if (!pool)
609                 return NULL;
610
611         spin_lock(&pool->prp_lock);
612
613         /* See if we have anything in a pool, and bail out if nothing,
614          * in writeout path, where this matters, this is safe to do, because
615          * nothing is lost in this case, and when some in-flight requests
616          * complete, this code will be called again. */
617         if (unlikely(list_empty(&pool->prp_req_list))) {
618                 spin_unlock(&pool->prp_lock);
619                 return NULL;
620         }
621
622         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
623                              rq_list);
624         list_del_init(&request->rq_list);
625         spin_unlock(&pool->prp_lock);
626
627         LASSERT(request->rq_reqbuf);
628         LASSERT(request->rq_pool);
629
630         reqbuf = request->rq_reqbuf;
631         memset(request, 0, sizeof(*request));
632         request->rq_reqbuf = reqbuf;
633         request->rq_reqbuf_len = pool->prp_rq_size;
634         request->rq_pool = pool;
635
636         return request;
637 }
638
639 /**
640  * Returns freed \a request to pool.
641  */
642 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
643 {
644         struct ptlrpc_request_pool *pool = request->rq_pool;
645
646         spin_lock(&pool->prp_lock);
647         LASSERT(list_empty(&request->rq_list));
648         LASSERT(!request->rq_receiving_reply);
649         list_add_tail(&request->rq_list, &pool->prp_req_list);
650         spin_unlock(&pool->prp_lock);
651 }
652
653 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
654                                       __u32 version, int opcode,
655                                       int count, __u32 *lengths, char **bufs,
656                                       struct ptlrpc_cli_ctx *ctx)
657 {
658         struct obd_import  *imp = request->rq_import;
659         int                 rc;
660         ENTRY;
661
662         if (unlikely(ctx))
663                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
664         else {
665                 rc = sptlrpc_req_get_ctx(request);
666                 if (rc)
667                         GOTO(out_free, rc);
668         }
669
670         sptlrpc_req_set_flavor(request, opcode);
671
672         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
673                                  lengths, bufs);
674         if (rc) {
675                 LASSERT(!request->rq_pool);
676                 GOTO(out_ctx, rc);
677         }
678
679         lustre_msg_add_version(request->rq_reqmsg, version);
680         request->rq_send_state = LUSTRE_IMP_FULL;
681         request->rq_type = PTL_RPC_MSG_REQUEST;
682
683         request->rq_req_cbid.cbid_fn  = request_out_callback;
684         request->rq_req_cbid.cbid_arg = request;
685
686         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
687         request->rq_reply_cbid.cbid_arg = request;
688
689         request->rq_reply_deadline = 0;
690         request->rq_phase = RQ_PHASE_NEW;
691         request->rq_next_phase = RQ_PHASE_UNDEFINED;
692
693         request->rq_request_portal = imp->imp_client->cli_request_portal;
694         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
695
696         ptlrpc_at_set_req_timeout(request);
697
698         lustre_msg_set_opc(request->rq_reqmsg, opcode);
699
700         RETURN(0);
701 out_ctx:
702         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
703 out_free:
704         class_import_put(imp);
705         return rc;
706 }
707
708 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
709                              __u32 version, int opcode, char **bufs,
710                              struct ptlrpc_cli_ctx *ctx)
711 {
712         int count;
713
714         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
715         return __ptlrpc_request_bufs_pack(request, version, opcode, count,
716                                           request->rq_pill.rc_area[RCL_CLIENT],
717                                           bufs, ctx);
718 }
719 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
720
721 /**
722  * Pack request buffers for network transfer, performing necessary encryption
723  * steps if necessary.
724  */
725 int ptlrpc_request_pack(struct ptlrpc_request *request,
726                         __u32 version, int opcode)
727 {
728         int rc;
729         rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
730         if (rc)
731                 return rc;
732
733         /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
734          * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
735          * have to send old ptlrpc_body to keep interoprability with these
736          * clients.
737          *
738          * Only three kinds of server->client RPCs so far:
739          *  - LDLM_BL_CALLBACK
740          *  - LDLM_CP_CALLBACK
741          *  - LDLM_GL_CALLBACK
742          *
743          * XXX This should be removed whenever we drop the interoprability with
744          *     the these old clients.
745          */
746         if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
747             opcode == LDLM_GL_CALLBACK)
748                 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
749                                    sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
750
751         return rc;
752 }
753 EXPORT_SYMBOL(ptlrpc_request_pack);
754
755 /**
756  * Helper function to allocate new request on import \a imp
757  * and possibly using existing request from pool \a pool if provided.
758  * Returns allocated request structure with import field filled or
759  * NULL on error.
760  */
761 static inline
762 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
763                                               struct ptlrpc_request_pool *pool)
764 {
765         struct ptlrpc_request *request = NULL;
766
767         if (pool)
768                 request = ptlrpc_prep_req_from_pool(pool);
769
770         if (!request)
771                 request = ptlrpc_request_cache_alloc(GFP_NOFS);
772
773         if (request) {
774                 ptlrpc_cli_req_init(request);
775
776                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
777                 LASSERT(imp != LP_POISON);
778                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
779                         imp->imp_client);
780                 LASSERT(imp->imp_client != LP_POISON);
781
782                 request->rq_import = class_import_get(imp);
783         } else {
784                 CERROR("request allocation out of memory\n");
785         }
786
787         return request;
788 }
789
790 /**
791  * Helper function for creating a request.
792  * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
793  * buffer structures according to capsule template \a format.
794  * Returns allocated request structure pointer or NULL on error.
795  */
796 static struct ptlrpc_request *
797 ptlrpc_request_alloc_internal(struct obd_import *imp,
798                               struct ptlrpc_request_pool * pool,
799                               const struct req_format *format)
800 {
801         struct ptlrpc_request *request;
802
803         request = __ptlrpc_request_alloc(imp, pool);
804         if (request == NULL)
805                 return NULL;
806
807         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
808         req_capsule_set(&request->rq_pill, format);
809         return request;
810 }
811
812 /**
813  * Allocate new request structure for import \a imp and initialize its
814  * buffer structure according to capsule template \a format.
815  */
816 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
817                                             const struct req_format *format)
818 {
819         return ptlrpc_request_alloc_internal(imp, NULL, format);
820 }
821 EXPORT_SYMBOL(ptlrpc_request_alloc);
822
823 /**
824  * Allocate new request structure for import \a imp from pool \a pool and
825  * initialize its buffer structure according to capsule template \a format.
826  */
827 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
828                                             struct ptlrpc_request_pool * pool,
829                                             const struct req_format *format)
830 {
831         return ptlrpc_request_alloc_internal(imp, pool, format);
832 }
833 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
834
835 /**
836  * For requests not from pool, free memory of the request structure.
837  * For requests obtained from a pool earlier, return request back to pool.
838  */
839 void ptlrpc_request_free(struct ptlrpc_request *request)
840 {
841         if (request->rq_pool)
842                 __ptlrpc_free_req_to_pool(request);
843         else
844                 ptlrpc_request_cache_free(request);
845 }
846 EXPORT_SYMBOL(ptlrpc_request_free);
847
848 /**
849  * Allocate new request for operatione \a opcode and immediatelly pack it for
850  * network transfer.
851  * Only used for simple requests like OBD_PING where the only important
852  * part of the request is operation itself.
853  * Returns allocated request or NULL on error.
854  */
855 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
856                                                 const struct req_format *format,
857                                                 __u32 version, int opcode)
858 {
859         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
860         int                    rc;
861
862         if (req) {
863                 rc = ptlrpc_request_pack(req, version, opcode);
864                 if (rc) {
865                         ptlrpc_request_free(req);
866                         req = NULL;
867                 }
868         }
869         return req;
870 }
871 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
872
873 /**
874  * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
875  * for operation \a opcode. Request would contain \a count buffers.
876  * Sizes of buffers are described in array \a lengths and buffers themselves
877  * are provided by a pointer \a bufs.
878  * Returns prepared request structure pointer or NULL on error.
879  */
880 struct ptlrpc_request *
881 ptlrpc_prep_req_pool(struct obd_import *imp,
882                      __u32 version, int opcode,
883                      int count, __u32 *lengths, char **bufs,
884                      struct ptlrpc_request_pool *pool)
885 {
886         struct ptlrpc_request *request;
887         int                    rc;
888
889         request = __ptlrpc_request_alloc(imp, pool);
890         if (!request)
891                 return NULL;
892
893         rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
894                                         lengths, bufs, NULL);
895         if (rc) {
896                 ptlrpc_request_free(request);
897                 request = NULL;
898         }
899         return request;
900 }
901
902 /**
903  * Same as ptlrpc_prep_req_pool, but without pool
904  */
905 struct ptlrpc_request *
906 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
907                 __u32 *lengths, char **bufs)
908 {
909         return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
910                                     NULL);
911 }
912
913 /**
914  * Allocate and initialize new request set structure.
915  * Returns a pointer to the newly allocated set structure or NULL on error.
916  */
917 struct ptlrpc_request_set *ptlrpc_prep_set(void)
918 {
919         struct ptlrpc_request_set *set;
920
921         ENTRY;
922         OBD_ALLOC(set, sizeof *set);
923         if (!set)
924                 RETURN(NULL);
925         atomic_set(&set->set_refcount, 1);
926         INIT_LIST_HEAD(&set->set_requests);
927         init_waitqueue_head(&set->set_waitq);
928         atomic_set(&set->set_new_count, 0);
929         atomic_set(&set->set_remaining, 0);
930         spin_lock_init(&set->set_new_req_lock);
931         INIT_LIST_HEAD(&set->set_new_requests);
932         INIT_LIST_HEAD(&set->set_cblist);
933         set->set_max_inflight = UINT_MAX;
934         set->set_producer     = NULL;
935         set->set_producer_arg = NULL;
936         set->set_rc           = 0;
937
938         RETURN(set);
939 }
940 EXPORT_SYMBOL(ptlrpc_prep_set);
941
942 /**
943  * Allocate and initialize new request set structure with flow control
944  * extension. This extension allows to control the number of requests in-flight
945  * for the whole set. A callback function to generate requests must be provided
946  * and the request set will keep the number of requests sent over the wire to
947  * @max_inflight.
948  * Returns a pointer to the newly allocated set structure or NULL on error.
949  */
950 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
951                                              void *arg)
952
953 {
954         struct ptlrpc_request_set *set;
955
956         set = ptlrpc_prep_set();
957         if (!set)
958                 RETURN(NULL);
959
960         set->set_max_inflight  = max;
961         set->set_producer      = func;
962         set->set_producer_arg  = arg;
963
964         RETURN(set);
965 }
966
967 /**
968  * Wind down and free request set structure previously allocated with
969  * ptlrpc_prep_set.
970  * Ensures that all requests on the set have completed and removes
971  * all requests from the request list in a set.
972  * If any unsent request happen to be on the list, pretends that they got
973  * an error in flight and calls their completion handler.
974  */
975 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
976 {
977         struct list_head        *tmp;
978         struct list_head        *next;
979         int                      expected_phase;
980         int                      n = 0;
981         ENTRY;
982
983         /* Requests on the set should either all be completed, or all be new */
984         expected_phase = (atomic_read(&set->set_remaining) == 0) ?
985                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
986         list_for_each(tmp, &set->set_requests) {
987                 struct ptlrpc_request *req =
988                         list_entry(tmp, struct ptlrpc_request,
989                                    rq_set_chain);
990
991                 LASSERT(req->rq_phase == expected_phase);
992                 n++;
993         }
994
995         LASSERTF(atomic_read(&set->set_remaining) == 0 ||
996                  atomic_read(&set->set_remaining) == n, "%d / %d\n",
997                  atomic_read(&set->set_remaining), n);
998
999         list_for_each_safe(tmp, next, &set->set_requests) {
1000                 struct ptlrpc_request *req =
1001                         list_entry(tmp, struct ptlrpc_request,
1002                                    rq_set_chain);
1003                 list_del_init(&req->rq_set_chain);
1004
1005                 LASSERT(req->rq_phase == expected_phase);
1006
1007                 if (req->rq_phase == RQ_PHASE_NEW) {
1008                         ptlrpc_req_interpret(NULL, req, -EBADR);
1009                         atomic_dec(&set->set_remaining);
1010                 }
1011
1012                 spin_lock(&req->rq_lock);
1013                 req->rq_set = NULL;
1014                 req->rq_invalid_rqset = 0;
1015                 spin_unlock(&req->rq_lock);
1016
1017                 ptlrpc_req_finished (req);
1018         }
1019
1020         LASSERT(atomic_read(&set->set_remaining) == 0);
1021
1022         ptlrpc_reqset_put(set);
1023         EXIT;
1024 }
1025 EXPORT_SYMBOL(ptlrpc_set_destroy);
1026
1027 /**
1028  * Add a callback function \a fn to the set.
1029  * This function would be called when all requests on this set are completed.
1030  * The function will be passed \a data argument.
1031  */
1032 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1033                       set_interpreter_func fn, void *data)
1034 {
1035         struct ptlrpc_set_cbdata *cbdata;
1036
1037         OBD_ALLOC_PTR(cbdata);
1038         if (cbdata == NULL)
1039                 RETURN(-ENOMEM);
1040
1041         cbdata->psc_interpret = fn;
1042         cbdata->psc_data = data;
1043         list_add_tail(&cbdata->psc_item, &set->set_cblist);
1044
1045         RETURN(0);
1046 }
1047
1048 /**
1049  * Add a new request to the general purpose request set.
1050  * Assumes request reference from the caller.
1051  */
1052 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1053                         struct ptlrpc_request *req)
1054 {
1055         LASSERT(list_empty(&req->rq_set_chain));
1056
1057         /* The set takes over the caller's request reference */
1058         list_add_tail(&req->rq_set_chain, &set->set_requests);
1059         req->rq_set = set;
1060         atomic_inc(&set->set_remaining);
1061         req->rq_queued_time = cfs_time_current();
1062
1063         if (req->rq_reqmsg != NULL)
1064                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1065
1066         if (set->set_producer != NULL)
1067                 /* If the request set has a producer callback, the RPC must be
1068                  * sent straight away */
1069                 ptlrpc_send_new_req(req);
1070 }
1071 EXPORT_SYMBOL(ptlrpc_set_add_req);
1072
1073 /**
1074  * Add a request to a request with dedicated server thread
1075  * and wake the thread to make any necessary processing.
1076  * Currently only used for ptlrpcd.
1077  */
1078 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1079                            struct ptlrpc_request *req)
1080 {
1081         struct ptlrpc_request_set *set = pc->pc_set;
1082         int count, i;
1083
1084         LASSERT(req->rq_set == NULL);
1085         LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1086
1087         spin_lock(&set->set_new_req_lock);
1088         /*
1089          * The set takes over the caller's request reference.
1090          */
1091         req->rq_set = set;
1092         req->rq_queued_time = cfs_time_current();
1093         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1094         count = atomic_inc_return(&set->set_new_count);
1095         spin_unlock(&set->set_new_req_lock);
1096
1097         /* Only need to call wakeup once for the first entry. */
1098         if (count == 1) {
1099                 wake_up(&set->set_waitq);
1100
1101                 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1102                  *      guarantee the async RPC can be processed ASAP, we have
1103                  *      no other better choice. It maybe fixed in future. */
1104                 for (i = 0; i < pc->pc_npartners; i++)
1105                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1106         }
1107 }
1108
1109 /**
1110  * Based on the current state of the import, determine if the request
1111  * can be sent, is an error, or should be delayed.
1112  *
1113  * Returns true if this request should be delayed. If false, and
1114  * *status is set, then the request can not be sent and *status is the
1115  * error code.  If false and status is 0, then request can be sent.
1116  *
1117  * The imp->imp_lock must be held.
1118  */
1119 static int ptlrpc_import_delay_req(struct obd_import *imp,
1120                                    struct ptlrpc_request *req, int *status)
1121 {
1122         int delay = 0;
1123         ENTRY;
1124
1125         LASSERT (status != NULL);
1126         *status = 0;
1127
1128         if (req->rq_ctx_init || req->rq_ctx_fini) {
1129                 /* always allow ctx init/fini rpc go through */
1130         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1131                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1132                 *status = -EIO;
1133         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1134                 /* pings may safely race with umount */
1135                 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1136                           D_HA : D_ERROR, req, "IMP_CLOSED ");
1137                 *status = -EIO;
1138         } else if (ptlrpc_send_limit_expired(req)) {
1139                 /* probably doesn't need to be a D_ERROR after initial testing*/
1140                 DEBUG_REQ(D_HA, req, "send limit expired ");
1141                 *status = -ETIMEDOUT;
1142         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1143                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
1144                 /* allow CONNECT even if import is invalid */ ;
1145                 if (atomic_read(&imp->imp_inval_count) != 0) {
1146                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1147                         *status = -EIO;
1148                 }
1149         } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1150                 if (!imp->imp_deactive)
1151                         DEBUG_REQ(D_NET, req, "IMP_INVALID");
1152                 *status = -ESHUTDOWN; /* bz 12940 */
1153         } else if (req->rq_import_generation != imp->imp_generation) {
1154                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1155                 *status = -EIO;
1156         } else if (req->rq_send_state != imp->imp_state) {
1157                 /* invalidate in progress - any requests should be drop */
1158                 if (atomic_read(&imp->imp_inval_count) != 0) {
1159                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1160                         *status = -EIO;
1161                 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1162                         *status = -EWOULDBLOCK;
1163                 } else if (req->rq_allow_replay &&
1164                           (imp->imp_state == LUSTRE_IMP_REPLAY ||
1165                            imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1166                            imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1167                            imp->imp_state == LUSTRE_IMP_RECOVER)) {
1168                         DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1169                 } else {
1170                         delay = 1;
1171                 }
1172         }
1173
1174         RETURN(delay);
1175 }
1176
1177 /**
1178  * Decide if the error message should be printed to the console or not.
1179  * Makes its decision based on request type, status, and failure frequency.
1180  *
1181  * \param[in] req  request that failed and may need a console message
1182  *
1183  * \retval false if no message should be printed
1184  * \retval true  if console message should be printed
1185  */
1186 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1187 {
1188         __u32 opc;
1189
1190         LASSERT(req->rq_reqmsg != NULL);
1191         opc = lustre_msg_get_opc(req->rq_reqmsg);
1192
1193         /* Suppress particular reconnect errors which are to be expected. */
1194         if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1195                 int err;
1196
1197                 /* Suppress timed out reconnect requests */
1198                 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1199                     req->rq_timedout)
1200                         return false;
1201
1202                 /* Suppress most unavailable/again reconnect requests, but
1203                  * print occasionally so it is clear client is trying to
1204                  * connect to a server where no target is running. */
1205                 err = lustre_msg_get_status(req->rq_repmsg);
1206                 if ((err == -ENODEV || err == -EAGAIN) &&
1207                     req->rq_import->imp_conn_cnt % 30 != 20)
1208                         return false;
1209         }
1210
1211         return true;
1212 }
1213
1214 /**
1215  * Check request processing status.
1216  * Returns the status.
1217  */
1218 static int ptlrpc_check_status(struct ptlrpc_request *req)
1219 {
1220         int err;
1221         ENTRY;
1222
1223         err = lustre_msg_get_status(req->rq_repmsg);
1224         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1225                 struct obd_import *imp = req->rq_import;
1226                 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1227                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1228
1229                 if (ptlrpc_console_allow(req))
1230                         LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1231                                            "failed: rc = %d\n",
1232                                            imp->imp_obd->obd_name,
1233                                            ll_opcode2str(opc),
1234                                            libcfs_nid2str(nid), err);
1235                 RETURN(err < 0 ? err : -EINVAL);
1236         }
1237
1238         if (err < 0) {
1239                 DEBUG_REQ(D_INFO, req, "status is %d", err);
1240         } else if (err > 0) {
1241                 /* XXX: translate this error from net to host */
1242                 DEBUG_REQ(D_INFO, req, "status is %d", err);
1243         }
1244
1245         RETURN(err);
1246 }
1247
1248 /**
1249  * save pre-versions of objects into request for replay.
1250  * Versions are obtained from server reply.
1251  * used for VBR.
1252  */
1253 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1254 {
1255         struct lustre_msg *repmsg = req->rq_repmsg;
1256         struct lustre_msg *reqmsg = req->rq_reqmsg;
1257         __u64 *versions = lustre_msg_get_versions(repmsg);
1258         ENTRY;
1259
1260         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1261                 return;
1262
1263         LASSERT(versions);
1264         lustre_msg_set_versions(reqmsg, versions);
1265         CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1266                versions[0], versions[1]);
1267
1268         EXIT;
1269 }
1270
1271 /**
1272  * Callback function called when client receives RPC reply for \a req.
1273  * Returns 0 on success or error code.
1274  * The return alue would be assigned to req->rq_status by the caller
1275  * as request processing status.
1276  * This function also decides if the request needs to be saved for later replay.
1277  */
1278 static int after_reply(struct ptlrpc_request *req)
1279 {
1280         struct obd_import *imp = req->rq_import;
1281         struct obd_device *obd = req->rq_import->imp_obd;
1282         int rc;
1283         struct timeval work_start;
1284         long timediff;
1285         ENTRY;
1286
1287         LASSERT(obd != NULL);
1288         /* repbuf must be unlinked */
1289         LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1290
1291         if (req->rq_reply_truncated) {
1292                 if (ptlrpc_no_resend(req)) {
1293                         DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1294                                   " expected: %d, actual size: %d",
1295                                   req->rq_nob_received, req->rq_repbuf_len);
1296                         RETURN(-EOVERFLOW);
1297                 }
1298
1299                 sptlrpc_cli_free_repbuf(req);
1300                 /* Pass the required reply buffer size (include
1301                  * space for early reply).
1302                  * NB: no need to roundup because alloc_repbuf
1303                  * will roundup it */
1304                 req->rq_replen       = req->rq_nob_received;
1305                 req->rq_nob_received = 0;
1306                 spin_lock(&req->rq_lock);
1307                 req->rq_resend       = 1;
1308                 spin_unlock(&req->rq_lock);
1309                 RETURN(0);
1310         }
1311
1312         do_gettimeofday(&work_start);
1313         timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1314
1315         /*
1316          * NB Until this point, the whole of the incoming message,
1317          * including buflens, status etc is in the sender's byte order.
1318          */
1319         rc = sptlrpc_cli_unwrap_reply(req);
1320         if (rc) {
1321                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1322                 RETURN(rc);
1323         }
1324
1325         /*
1326          * Security layer unwrap might ask resend this request.
1327          */
1328         if (req->rq_resend)
1329                 RETURN(0);
1330
1331         rc = unpack_reply(req);
1332         if (rc)
1333                 RETURN(rc);
1334
1335         /* retry indefinitely on EINPROGRESS */
1336         if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1337             ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1338                 time_t  now = cfs_time_current_sec();
1339
1340                 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1341                 spin_lock(&req->rq_lock);
1342                 req->rq_resend = 1;
1343                 spin_unlock(&req->rq_lock);
1344                 req->rq_nr_resend++;
1345
1346                 /* allocate new xid to avoid reply reconstruction */
1347                 if (!req->rq_bulk) {
1348                         /* new xid is already allocated for bulk in
1349                          * ptlrpc_check_set() */
1350                         req->rq_xid = ptlrpc_next_xid();
1351                         DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1352                                   "resend on EINPROGRESS");
1353                 }
1354
1355                 /* Readjust the timeout for current conditions */
1356                 ptlrpc_at_set_req_timeout(req);
1357                 /* delay resend to give a chance to the server to get ready.
1358                  * The delay is increased by 1s on every resend and is capped to
1359                  * the current request timeout (i.e. obd_timeout if AT is off,
1360                  * or AT service time x 125% + 5s, see at_est2timeout) */
1361                 if (req->rq_nr_resend > req->rq_timeout)
1362                         req->rq_sent = now + req->rq_timeout;
1363                 else
1364                         req->rq_sent = now + req->rq_nr_resend;
1365
1366                 RETURN(0);
1367         }
1368
1369         if (obd->obd_svc_stats != NULL) {
1370                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1371                                     timediff);
1372                 ptlrpc_lprocfs_rpc_sent(req, timediff);
1373         }
1374
1375         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1376             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1377                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1378                           lustre_msg_get_type(req->rq_repmsg));
1379                 RETURN(-EPROTO);
1380         }
1381
1382         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1383                 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1384         ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1385         ptlrpc_at_adj_net_latency(req,
1386                                   lustre_msg_get_service_time(req->rq_repmsg));
1387
1388         rc = ptlrpc_check_status(req);
1389         imp->imp_connect_error = rc;
1390
1391         if (rc) {
1392                 /*
1393                  * Either we've been evicted, or the server has failed for
1394                  * some reason. Try to reconnect, and if that fails, punt to
1395                  * the upcall.
1396                  */
1397                 if (ptlrpc_recoverable_error(rc)) {
1398                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
1399                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1400                                 RETURN(rc);
1401                         }
1402                         ptlrpc_request_handle_notconn(req);
1403                         RETURN(rc);
1404                 }
1405         } else {
1406                 /*
1407                  * Let's look if server sent slv. Do it only for RPC with
1408                  * rc == 0.
1409                  */
1410                 ldlm_cli_update_pool(req);
1411         }
1412
1413         /*
1414          * Store transno in reqmsg for replay.
1415          */
1416         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1417                 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1418                 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1419         }
1420
1421         if (imp->imp_replayable) {
1422                 spin_lock(&imp->imp_lock);
1423                 /*
1424                  * No point in adding already-committed requests to the replay
1425                  * list, we will just remove them immediately. b=9829
1426                  */
1427                 if (req->rq_transno != 0 &&
1428                     (req->rq_transno >
1429                      lustre_msg_get_last_committed(req->rq_repmsg) ||
1430                      req->rq_replay)) {
1431                         /** version recovery */
1432                         ptlrpc_save_versions(req);
1433                         ptlrpc_retain_replayable_request(req, imp);
1434                 } else if (req->rq_commit_cb != NULL &&
1435                            list_empty(&req->rq_replay_list)) {
1436                         /* NB: don't call rq_commit_cb if it's already on
1437                          * rq_replay_list, ptlrpc_free_committed() will call
1438                          * it later, see LU-3618 for details */
1439                         spin_unlock(&imp->imp_lock);
1440                         req->rq_commit_cb(req);
1441                         spin_lock(&imp->imp_lock);
1442                 }
1443
1444                 /*
1445                  * Replay-enabled imports return commit-status information.
1446                  */
1447                 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1448                         imp->imp_peer_committed_transno =
1449                                 lustre_msg_get_last_committed(req->rq_repmsg);
1450                 }
1451
1452                 ptlrpc_free_committed(imp);
1453
1454                 if (!list_empty(&imp->imp_replay_list)) {
1455                         struct ptlrpc_request *last;
1456
1457                         last = list_entry(imp->imp_replay_list.prev,
1458                                           struct ptlrpc_request,
1459                                           rq_replay_list);
1460                         /*
1461                          * Requests with rq_replay stay on the list even if no
1462                          * commit is expected.
1463                          */
1464                         if (last->rq_transno > imp->imp_peer_committed_transno)
1465                                 ptlrpc_pinger_commit_expected(imp);
1466                 }
1467
1468                 spin_unlock(&imp->imp_lock);
1469         }
1470
1471         RETURN(rc);
1472 }
1473
1474 /**
1475  * Helper function to send request \a req over the network for the first time
1476  * Also adjusts request phase.
1477  * Returns 0 on success or error code.
1478  */
1479 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1480 {
1481         struct obd_import     *imp = req->rq_import;
1482         struct list_head      *tmp;
1483         __u64                  min_xid = ~0ULL;
1484         int rc;
1485         ENTRY;
1486
1487         LASSERT(req->rq_phase == RQ_PHASE_NEW);
1488         if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1489             (!req->rq_generation_set ||
1490              req->rq_import_generation == imp->imp_generation))
1491                 RETURN (0);
1492
1493         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1494
1495         spin_lock(&imp->imp_lock);
1496
1497         /* the very first time we assign XID. it's important to assign XID
1498          * and put it on the list atomically, so that the lowest assigned
1499          * XID is always known. this is vital for multislot last_rcvd */
1500         if (req->rq_send_state == LUSTRE_IMP_REPLAY) {
1501                 LASSERT(req->rq_xid != 0);
1502         } else {
1503                 LASSERT(req->rq_xid == 0);
1504                 req->rq_xid = ptlrpc_next_xid();
1505         }
1506
1507         if (!req->rq_generation_set)
1508                 req->rq_import_generation = imp->imp_generation;
1509
1510         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1511                 spin_lock(&req->rq_lock);
1512                 req->rq_waiting = 1;
1513                 spin_unlock(&req->rq_lock);
1514
1515                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1516                           "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1517                           ptlrpc_import_state_name(req->rq_send_state),
1518                           ptlrpc_import_state_name(imp->imp_state));
1519                 LASSERT(list_empty(&req->rq_list));
1520                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1521                 atomic_inc(&req->rq_import->imp_inflight);
1522                 spin_unlock(&imp->imp_lock);
1523                 RETURN(0);
1524         }
1525
1526         if (rc != 0) {
1527                 spin_unlock(&imp->imp_lock);
1528                 req->rq_status = rc;
1529                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1530                 RETURN(rc);
1531         }
1532
1533         LASSERT(list_empty(&req->rq_list));
1534         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1535         atomic_inc(&req->rq_import->imp_inflight);
1536
1537         /* find the lowest unreplied XID */
1538         list_for_each(tmp, &imp->imp_delayed_list) {
1539                 struct ptlrpc_request *r;
1540                 r = list_entry(tmp, struct ptlrpc_request, rq_list);
1541                 if (r->rq_xid < min_xid)
1542                         min_xid = r->rq_xid;
1543         }
1544         list_for_each(tmp, &imp->imp_sending_list) {
1545                 struct ptlrpc_request *r;
1546                 r = list_entry(tmp, struct ptlrpc_request, rq_list);
1547                 if (r->rq_xid < min_xid)
1548                         min_xid = r->rq_xid;
1549         }
1550         spin_unlock(&imp->imp_lock);
1551
1552         if (likely(min_xid != ~0ULL))
1553                 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid - 1);
1554
1555         lustre_msg_set_status(req->rq_reqmsg, current_pid());
1556
1557         rc = sptlrpc_req_refresh_ctx(req, -1);
1558         if (rc) {
1559                 if (req->rq_err) {
1560                         req->rq_status = rc;
1561                         RETURN(1);
1562                 } else {
1563                         spin_lock(&req->rq_lock);
1564                         req->rq_wait_ctx = 1;
1565                         spin_unlock(&req->rq_lock);
1566                         RETURN(0);
1567                 }
1568         }
1569
1570         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1571                " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1572                imp->imp_obd->obd_uuid.uuid,
1573                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1574                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1575                lustre_msg_get_opc(req->rq_reqmsg));
1576
1577         rc = ptl_send_rpc(req, 0);
1578         if (rc) {
1579                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1580                 spin_lock(&req->rq_lock);
1581                 req->rq_net_err = 1;
1582                 spin_unlock(&req->rq_lock);
1583                 RETURN(rc);
1584         }
1585         RETURN(0);
1586 }
1587
1588 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1589 {
1590         int remaining, rc;
1591         ENTRY;
1592
1593         LASSERT(set->set_producer != NULL);
1594
1595         remaining = atomic_read(&set->set_remaining);
1596
1597         /* populate the ->set_requests list with requests until we
1598          * reach the maximum number of RPCs in flight for this set */
1599         while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1600                 rc = set->set_producer(set, set->set_producer_arg);
1601                 if (rc == -ENOENT) {
1602                         /* no more RPC to produce */
1603                         set->set_producer     = NULL;
1604                         set->set_producer_arg = NULL;
1605                         RETURN(0);
1606                 }
1607         }
1608
1609         RETURN((atomic_read(&set->set_remaining) - remaining));
1610 }
1611
1612 /**
1613  * this sends any unsent RPCs in \a set and returns 1 if all are sent
1614  * and no more replies are expected.
1615  * (it is possible to get less replies than requests sent e.g. due to timed out
1616  * requests or requests that we had trouble to send out)
1617  *
1618  * NOTE: This function contains a potential schedule point (cond_resched()).
1619  */
1620 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1621 {
1622         struct list_head *tmp, *next;
1623         struct list_head  comp_reqs;
1624         int force_timer_recalc = 0;
1625         ENTRY;
1626
1627         if (atomic_read(&set->set_remaining) == 0)
1628                 RETURN(1);
1629
1630         INIT_LIST_HEAD(&comp_reqs);
1631         list_for_each_safe(tmp, next, &set->set_requests) {
1632                 struct ptlrpc_request *req =
1633                         list_entry(tmp, struct ptlrpc_request,
1634                                    rq_set_chain);
1635                 struct obd_import *imp = req->rq_import;
1636                 int unregistered = 0;
1637                 int rc = 0;
1638
1639                 /* This schedule point is mainly for the ptlrpcd caller of this
1640                  * function.  Most ptlrpc sets are not long-lived and unbounded
1641                  * in length, but at the least the set used by the ptlrpcd is.
1642                  * Since the processing time is unbounded, we need to insert an
1643                  * explicit schedule point to make the thread well-behaved.
1644                  */
1645                 cond_resched();
1646
1647                 if (req->rq_phase == RQ_PHASE_NEW &&
1648                     ptlrpc_send_new_req(req)) {
1649                         force_timer_recalc = 1;
1650                 }
1651
1652                 /* delayed send - skip */
1653                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1654                         continue;
1655
1656                 /* delayed resend - skip */
1657                 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1658                     req->rq_sent > cfs_time_current_sec())
1659                         continue;
1660
1661                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1662                       req->rq_phase == RQ_PHASE_BULK ||
1663                       req->rq_phase == RQ_PHASE_INTERPRET ||
1664                       req->rq_phase == RQ_PHASE_UNREGISTERING ||
1665                       req->rq_phase == RQ_PHASE_COMPLETE)) {
1666                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1667                         LBUG();
1668                 }
1669
1670                 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1671                         LASSERT(req->rq_next_phase != req->rq_phase);
1672                         LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1673
1674                         /*
1675                          * Skip processing until reply is unlinked. We
1676                          * can't return to pool before that and we can't
1677                          * call interpret before that. We need to make
1678                          * sure that all rdma transfers finished and will
1679                          * not corrupt any data.
1680                          */
1681                         if (ptlrpc_client_recv_or_unlink(req) ||
1682                             ptlrpc_client_bulk_active(req))
1683                                 continue;
1684
1685                         /*
1686                          * Turn fail_loc off to prevent it from looping
1687                          * forever.
1688                          */
1689                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1690                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1691                                                      OBD_FAIL_ONCE);
1692                         }
1693                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1694                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1695                                                      OBD_FAIL_ONCE);
1696                         }
1697
1698                         /*
1699                          * Move to next phase if reply was successfully
1700                          * unlinked.
1701                          */
1702                         ptlrpc_rqphase_move(req, req->rq_next_phase);
1703                 }
1704
1705                 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1706                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1707                         continue;
1708                 }
1709
1710                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1711                         GOTO(interpret, req->rq_status);
1712
1713                 /*
1714                  * Note that this also will start async reply unlink.
1715                  */
1716                 if (req->rq_net_err && !req->rq_timedout) {
1717                         ptlrpc_expire_one_request(req, 1);
1718
1719                         /*
1720                          * Check if we still need to wait for unlink.
1721                          */
1722                         if (ptlrpc_client_recv_or_unlink(req) ||
1723                             ptlrpc_client_bulk_active(req))
1724                                 continue;
1725                         /* If there is no need to resend, fail it now. */
1726                         if (req->rq_no_resend) {
1727                                 if (req->rq_status == 0)
1728                                         req->rq_status = -EIO;
1729                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1730                                 GOTO(interpret, req->rq_status);
1731                         } else {
1732                                 continue;
1733                         }
1734                 }
1735
1736                 if (req->rq_err) {
1737                         spin_lock(&req->rq_lock);
1738                         req->rq_replied = 0;
1739                         spin_unlock(&req->rq_lock);
1740                         if (req->rq_status == 0)
1741                                 req->rq_status = -EIO;
1742                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1743                         GOTO(interpret, req->rq_status);
1744                 }
1745
1746                 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1747                  * so it sets rq_intr regardless of individual rpc
1748                  * timeouts. The synchronous IO waiting path sets
1749                  * rq_intr irrespective of whether ptlrpcd
1750                  * has seen a timeout.  Our policy is to only interpret
1751                  * interrupted rpcs after they have timed out, so we
1752                  * need to enforce that here.
1753                  */
1754
1755                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1756                                      req->rq_wait_ctx)) {
1757                         req->rq_status = -EINTR;
1758                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1759                         GOTO(interpret, req->rq_status);
1760                 }
1761
1762                 if (req->rq_phase == RQ_PHASE_RPC) {
1763                         if (req->rq_timedout || req->rq_resend ||
1764                             req->rq_waiting || req->rq_wait_ctx) {
1765                                 int status;
1766
1767                                 if (!ptlrpc_unregister_reply(req, 1)) {
1768                                         ptlrpc_unregister_bulk(req, 1);
1769                                         continue;
1770                                 }
1771
1772                                 spin_lock(&imp->imp_lock);
1773                                 if (ptlrpc_import_delay_req(imp, req, &status)){
1774                                         /* put on delay list - only if we wait
1775                                          * recovery finished - before send */
1776                                         list_del_init(&req->rq_list);
1777                                         list_add_tail(&req->rq_list,
1778                                                           &imp->
1779                                                           imp_delayed_list);
1780                                         spin_unlock(&imp->imp_lock);
1781                                         continue;
1782                                 }
1783
1784                                 if (status != 0)  {
1785                                         req->rq_status = status;
1786                                         ptlrpc_rqphase_move(req,
1787                                                 RQ_PHASE_INTERPRET);
1788                                         spin_unlock(&imp->imp_lock);
1789                                         GOTO(interpret, req->rq_status);
1790                                 }
1791                                 if (ptlrpc_no_resend(req) &&
1792                                     !req->rq_wait_ctx) {
1793                                         req->rq_status = -ENOTCONN;
1794                                         ptlrpc_rqphase_move(req,
1795                                                             RQ_PHASE_INTERPRET);
1796                                         spin_unlock(&imp->imp_lock);
1797                                         GOTO(interpret, req->rq_status);
1798                                 }
1799
1800                                 list_del_init(&req->rq_list);
1801                                 list_add_tail(&req->rq_list,
1802                                                   &imp->imp_sending_list);
1803
1804                                 spin_unlock(&imp->imp_lock);
1805
1806                                 spin_lock(&req->rq_lock);
1807                                 req->rq_waiting = 0;
1808                                 spin_unlock(&req->rq_lock);
1809
1810                                 if (req->rq_timedout || req->rq_resend) {
1811                                         /* This is re-sending anyways,
1812                                          * let's mark req as resend. */
1813                                         spin_lock(&req->rq_lock);
1814                                         req->rq_resend = 1;
1815                                         spin_unlock(&req->rq_lock);
1816                                         if (req->rq_bulk) {
1817                                                 __u64 old_xid;
1818
1819                                                 if (!ptlrpc_unregister_bulk(req, 1))
1820                                                         continue;
1821
1822                                                 /* ensure previous bulk fails */
1823                                                 old_xid = req->rq_xid;
1824                                                 req->rq_xid = ptlrpc_next_xid();
1825                                                 CDEBUG(D_HA, "resend bulk "
1826                                                        "old x"LPU64
1827                                                        " new x"LPU64"\n",
1828                                                        old_xid, req->rq_xid);
1829                                         }
1830                                 }
1831                                 /*
1832                                  * rq_wait_ctx is only touched by ptlrpcd,
1833                                  * so no lock is needed here.
1834                                  */
1835                                 status = sptlrpc_req_refresh_ctx(req, -1);
1836                                 if (status) {
1837                                         if (req->rq_err) {
1838                                                 req->rq_status = status;
1839                                                 spin_lock(&req->rq_lock);
1840                                                 req->rq_wait_ctx = 0;
1841                                                 spin_unlock(&req->rq_lock);
1842                                                 force_timer_recalc = 1;
1843                                         } else {
1844                                                 spin_lock(&req->rq_lock);
1845                                                 req->rq_wait_ctx = 1;
1846                                                 spin_unlock(&req->rq_lock);
1847                                         }
1848
1849                                         continue;
1850                                 } else {
1851                                         spin_lock(&req->rq_lock);
1852                                         req->rq_wait_ctx = 0;
1853                                         spin_unlock(&req->rq_lock);
1854                                 }
1855
1856                                 rc = ptl_send_rpc(req, 0);
1857                                 if (rc) {
1858                                         DEBUG_REQ(D_HA, req,
1859                                                   "send failed: rc = %d", rc);
1860                                         force_timer_recalc = 1;
1861                                         spin_lock(&req->rq_lock);
1862                                         req->rq_net_err = 1;
1863                                         spin_unlock(&req->rq_lock);
1864                                         continue;
1865                                 }
1866                                 /* need to reset the timeout */
1867                                 force_timer_recalc = 1;
1868                         }
1869
1870                         spin_lock(&req->rq_lock);
1871
1872                         if (ptlrpc_client_early(req)) {
1873                                 ptlrpc_at_recv_early_reply(req);
1874                                 spin_unlock(&req->rq_lock);
1875                                 continue;
1876                         }
1877
1878                         /* Still waiting for a reply? */
1879                         if (ptlrpc_client_recv(req)) {
1880                                 spin_unlock(&req->rq_lock);
1881                                 continue;
1882                         }
1883
1884                         /* Did we actually receive a reply? */
1885                         if (!ptlrpc_client_replied(req)) {
1886                                 spin_unlock(&req->rq_lock);
1887                                 continue;
1888                         }
1889
1890                         spin_unlock(&req->rq_lock);
1891
1892                         /* unlink from net because we are going to
1893                          * swab in-place of reply buffer */
1894                         unregistered = ptlrpc_unregister_reply(req, 1);
1895                         if (!unregistered)
1896                                 continue;
1897
1898                         req->rq_status = after_reply(req);
1899                         if (req->rq_resend)
1900                                 continue;
1901
1902                         /* If there is no bulk associated with this request,
1903                          * then we're done and should let the interpreter
1904                          * process the reply. Similarly if the RPC returned
1905                          * an error, and therefore the bulk will never arrive.
1906                          */
1907                         if (req->rq_bulk == NULL || req->rq_status < 0) {
1908                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1909                                 GOTO(interpret, req->rq_status);
1910                         }
1911
1912                         ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1913                 }
1914
1915                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1916                 if (ptlrpc_client_bulk_active(req))
1917                         continue;
1918
1919                 if (req->rq_bulk->bd_failure) {
1920                         /* The RPC reply arrived OK, but the bulk screwed
1921                          * up!  Dead weird since the server told us the RPC
1922                          * was good after getting the REPLY for her GET or
1923                          * the ACK for her PUT. */
1924                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1925                         req->rq_status = -EIO;
1926                 }
1927
1928                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1929
1930         interpret:
1931                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1932
1933                 /* This moves to "unregistering" phase we need to wait for
1934                  * reply unlink. */
1935                 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1936                         /* start async bulk unlink too */
1937                         ptlrpc_unregister_bulk(req, 1);
1938                         continue;
1939                 }
1940
1941                 if (!ptlrpc_unregister_bulk(req, 1))
1942                         continue;
1943
1944                 /* When calling interpret receiving already should be
1945                  * finished. */
1946                 LASSERT(!req->rq_receiving_reply);
1947
1948                 ptlrpc_req_interpret(env, req, req->rq_status);
1949
1950                 if (ptlrpcd_check_work(req)) {
1951                         atomic_dec(&set->set_remaining);
1952                         continue;
1953                 }
1954                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1955
1956                 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1957                         "Completed RPC pname:cluuid:pid:xid:nid:"
1958                         "opc %s:%s:%d:"LPU64":%s:%d\n",
1959                         current_comm(), imp->imp_obd->obd_uuid.uuid,
1960                         lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1961                         libcfs_nid2str(imp->imp_connection->c_peer.nid),
1962                         lustre_msg_get_opc(req->rq_reqmsg));
1963
1964                 spin_lock(&imp->imp_lock);
1965                 /* Request already may be not on sending or delaying list. This
1966                  * may happen in the case of marking it erroneous for the case
1967                  * ptlrpc_import_delay_req(req, status) find it impossible to
1968                  * allow sending this rpc and returns *status != 0. */
1969                 if (!list_empty(&req->rq_list)) {
1970                         list_del_init(&req->rq_list);
1971                         atomic_dec(&imp->imp_inflight);
1972                 }
1973                 spin_unlock(&imp->imp_lock);
1974
1975                 atomic_dec(&set->set_remaining);
1976                 wake_up_all(&imp->imp_recovery_waitq);
1977
1978                 if (set->set_producer) {
1979                         /* produce a new request if possible */
1980                         if (ptlrpc_set_producer(set) > 0)
1981                                 force_timer_recalc = 1;
1982
1983                         /* free the request that has just been completed
1984                          * in order not to pollute set->set_requests */
1985                         list_del_init(&req->rq_set_chain);
1986                         spin_lock(&req->rq_lock);
1987                         req->rq_set = NULL;
1988                         req->rq_invalid_rqset = 0;
1989                         spin_unlock(&req->rq_lock);
1990
1991                         /* record rq_status to compute the final status later */
1992                         if (req->rq_status != 0)
1993                                 set->set_rc = req->rq_status;
1994                         ptlrpc_req_finished(req);
1995                 } else {
1996                         list_move_tail(&req->rq_set_chain, &comp_reqs);
1997                 }
1998         }
1999
2000         /* move completed request at the head of list so it's easier for
2001          * caller to find them */
2002         list_splice(&comp_reqs, &set->set_requests);
2003
2004         /* If we hit an error, we want to recover promptly. */
2005         RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2006 }
2007 EXPORT_SYMBOL(ptlrpc_check_set);
2008
2009 /**
2010  * Time out request \a req. is \a async_unlink is set, that means do not wait
2011  * until LNet actually confirms network buffer unlinking.
2012  * Return 1 if we should give up further retrying attempts or 0 otherwise.
2013  */
2014 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2015 {
2016         struct obd_import *imp = req->rq_import;
2017         int rc = 0;
2018         ENTRY;
2019
2020         spin_lock(&req->rq_lock);
2021         req->rq_timedout = 1;
2022         spin_unlock(&req->rq_lock);
2023
2024         DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
2025                   "/real "CFS_DURATION_T"]",
2026                   req->rq_net_err ? "failed due to network error" :
2027                      ((req->rq_real_sent == 0 ||
2028                        cfs_time_before(req->rq_real_sent, req->rq_sent) ||
2029                        cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
2030                       "timed out for sent delay" : "timed out for slow reply"),
2031                   req->rq_sent, req->rq_real_sent);
2032
2033         if (imp != NULL && obd_debug_peer_on_timeout)
2034                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
2035
2036         ptlrpc_unregister_reply(req, async_unlink);
2037         ptlrpc_unregister_bulk(req, async_unlink);
2038
2039         if (obd_dump_on_timeout)
2040                 libcfs_debug_dumplog();
2041
2042         if (imp == NULL) {
2043                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2044                 RETURN(1);
2045         }
2046
2047         atomic_inc(&imp->imp_timeouts);
2048
2049         /* The DLM server doesn't want recovery run on its imports. */
2050         if (imp->imp_dlm_fake)
2051                 RETURN(1);
2052
2053         /* If this request is for recovery or other primordial tasks,
2054          * then error it out here. */
2055         if (req->rq_ctx_init || req->rq_ctx_fini ||
2056             req->rq_send_state != LUSTRE_IMP_FULL ||
2057             imp->imp_obd->obd_no_recov) {
2058                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2059                           ptlrpc_import_state_name(req->rq_send_state),
2060                           ptlrpc_import_state_name(imp->imp_state));
2061                 spin_lock(&req->rq_lock);
2062                 req->rq_status = -ETIMEDOUT;
2063                 req->rq_err = 1;
2064                 spin_unlock(&req->rq_lock);
2065                 RETURN(1);
2066         }
2067
2068         /* if a request can't be resent we can't wait for an answer after
2069            the timeout */
2070         if (ptlrpc_no_resend(req)) {
2071                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2072                 rc = 1;
2073         }
2074
2075         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2076
2077         RETURN(rc);
2078 }
2079
2080 /**
2081  * Time out all uncompleted requests in request set pointed by \a data
2082  * Callback used when waiting on sets with l_wait_event.
2083  * Always returns 1.
2084  */
2085 int ptlrpc_expired_set(void *data)
2086 {
2087         struct ptlrpc_request_set       *set = data;
2088         struct list_head                *tmp;
2089         time_t                          now = cfs_time_current_sec();
2090         ENTRY;
2091
2092         LASSERT(set != NULL);
2093
2094         /*
2095          * A timeout expired. See which reqs it applies to...
2096          */
2097         list_for_each(tmp, &set->set_requests) {
2098                 struct ptlrpc_request *req =
2099                         list_entry(tmp, struct ptlrpc_request,
2100                                    rq_set_chain);
2101
2102                 /* don't expire request waiting for context */
2103                 if (req->rq_wait_ctx)
2104                         continue;
2105
2106                 /* Request in-flight? */
2107                 if (!((req->rq_phase == RQ_PHASE_RPC &&
2108                        !req->rq_waiting && !req->rq_resend) ||
2109                       (req->rq_phase == RQ_PHASE_BULK)))
2110                         continue;
2111
2112                 if (req->rq_timedout ||     /* already dealt with */
2113                     req->rq_deadline > now) /* not expired */
2114                         continue;
2115
2116                 /* Deal with this guy. Do it asynchronously to not block
2117                  * ptlrpcd thread. */
2118                 ptlrpc_expire_one_request(req, 1);
2119         }
2120
2121         /*
2122          * When waiting for a whole set, we always break out of the
2123          * sleep so we can recalculate the timeout, or enable interrupts
2124          * if everyone's timed out.
2125          */
2126         RETURN(1);
2127 }
2128
2129 /**
2130  * Sets rq_intr flag in \a req under spinlock.
2131  */
2132 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2133 {
2134         spin_lock(&req->rq_lock);
2135         req->rq_intr = 1;
2136         spin_unlock(&req->rq_lock);
2137 }
2138 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2139
2140 /**
2141  * Interrupts (sets interrupted flag) all uncompleted requests in
2142  * a set \a data. Callback for l_wait_event for interruptible waits.
2143  */
2144 static void ptlrpc_interrupted_set(void *data)
2145 {
2146         struct ptlrpc_request_set *set = data;
2147         struct list_head *tmp;
2148
2149         LASSERT(set != NULL);
2150         CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2151
2152         list_for_each(tmp, &set->set_requests) {
2153                 struct ptlrpc_request *req =
2154                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2155
2156                 if (req->rq_phase != RQ_PHASE_RPC &&
2157                     req->rq_phase != RQ_PHASE_UNREGISTERING)
2158                         continue;
2159
2160                 ptlrpc_mark_interrupted(req);
2161         }
2162 }
2163
2164 /**
2165  * Get the smallest timeout in the set; this does NOT set a timeout.
2166  */
2167 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2168 {
2169         struct list_head        *tmp;
2170         time_t                   now = cfs_time_current_sec();
2171         int                      timeout = 0;
2172         struct ptlrpc_request   *req;
2173         int                      deadline;
2174         ENTRY;
2175
2176         list_for_each(tmp, &set->set_requests) {
2177                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2178
2179                 /*
2180                  * Request in-flight?
2181                  */
2182                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2183                       (req->rq_phase == RQ_PHASE_BULK) ||
2184                       (req->rq_phase == RQ_PHASE_NEW)))
2185                         continue;
2186
2187                 /*
2188                  * Already timed out.
2189                  */
2190                 if (req->rq_timedout)
2191                         continue;
2192
2193                 /*
2194                  * Waiting for ctx.
2195                  */
2196                 if (req->rq_wait_ctx)
2197                         continue;
2198
2199                 if (req->rq_phase == RQ_PHASE_NEW)
2200                         deadline = req->rq_sent;
2201                 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2202                         deadline = req->rq_sent;
2203                 else
2204                         deadline = req->rq_sent + req->rq_timeout;
2205
2206                 if (deadline <= now)    /* actually expired already */
2207                         timeout = 1;    /* ASAP */
2208                 else if (timeout == 0 || timeout > deadline - now)
2209                         timeout = deadline - now;
2210         }
2211         RETURN(timeout);
2212 }
2213
2214 /**
2215  * Send all unset request from the set and then wait untill all
2216  * requests in the set complete (either get a reply, timeout, get an
2217  * error or otherwise be interrupted).
2218  * Returns 0 on success or error code otherwise.
2219  */
2220 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2221 {
2222         struct list_head            *tmp;
2223         struct ptlrpc_request *req;
2224         struct l_wait_info     lwi;
2225         int                    rc, timeout;
2226         ENTRY;
2227
2228         if (set->set_producer)
2229                 (void)ptlrpc_set_producer(set);
2230         else
2231                 list_for_each(tmp, &set->set_requests) {
2232                         req = list_entry(tmp, struct ptlrpc_request,
2233                                          rq_set_chain);
2234                         if (req->rq_phase == RQ_PHASE_NEW)
2235                                 (void)ptlrpc_send_new_req(req);
2236                 }
2237
2238         if (list_empty(&set->set_requests))
2239                 RETURN(0);
2240
2241         do {
2242                 timeout = ptlrpc_set_next_timeout(set);
2243
2244                 /* wait until all complete, interrupted, or an in-flight
2245                  * req times out */
2246                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2247                        set, timeout);
2248
2249                 if (timeout == 0 && !signal_pending(current))
2250                         /*
2251                          * No requests are in-flight (ether timed out
2252                          * or delayed), so we can allow interrupts.
2253                          * We still want to block for a limited time,
2254                          * so we allow interrupts during the timeout.
2255                          */
2256                         lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2257                                                    ptlrpc_expired_set,
2258                                                    ptlrpc_interrupted_set, set);
2259                 else
2260                         /*
2261                          * At least one request is in flight, so no
2262                          * interrupts are allowed. Wait until all
2263                          * complete, or an in-flight req times out.
2264                          */
2265                         lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2266                                           ptlrpc_expired_set, set);
2267
2268                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2269
2270                 /* LU-769 - if we ignored the signal because it was already
2271                  * pending when we started, we need to handle it now or we risk
2272                  * it being ignored forever */
2273                 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2274                     signal_pending(current)) {
2275                         sigset_t blocked_sigs =
2276                                            cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2277
2278                         /* In fact we only interrupt for the "fatal" signals
2279                          * like SIGINT or SIGKILL. We still ignore less
2280                          * important signals since ptlrpc set is not easily
2281                          * reentrant from userspace again */
2282                         if (signal_pending(current))
2283                                 ptlrpc_interrupted_set(set);
2284                         cfs_restore_sigs(blocked_sigs);
2285                 }
2286
2287                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2288
2289                 /* -EINTR => all requests have been flagged rq_intr so next
2290                  * check completes.
2291                  * -ETIMEDOUT => someone timed out.  When all reqs have
2292                  * timed out, signals are enabled allowing completion with
2293                  * EINTR.
2294                  * I don't really care if we go once more round the loop in
2295                  * the error cases -eeb. */
2296                 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2297                         list_for_each(tmp, &set->set_requests) {
2298                                 req = list_entry(tmp, struct ptlrpc_request,
2299                                                  rq_set_chain);
2300                                 spin_lock(&req->rq_lock);
2301                                 req->rq_invalid_rqset = 1;
2302                                 spin_unlock(&req->rq_lock);
2303                         }
2304                 }
2305         } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2306
2307         LASSERT(atomic_read(&set->set_remaining) == 0);
2308
2309         rc = set->set_rc; /* rq_status of already freed requests if any */
2310         list_for_each(tmp, &set->set_requests) {
2311                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2312
2313                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2314                 if (req->rq_status != 0)
2315                         rc = req->rq_status;
2316         }
2317
2318         if (set->set_interpret != NULL) {
2319                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2320                         set->set_interpret;
2321                 rc = interpreter (set, set->set_arg, rc);
2322         } else {
2323                 struct ptlrpc_set_cbdata *cbdata, *n;
2324                 int err;
2325
2326                 list_for_each_entry_safe(cbdata, n,
2327                                          &set->set_cblist, psc_item) {
2328                         list_del_init(&cbdata->psc_item);
2329                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2330                         if (err && !rc)
2331                                 rc = err;
2332                         OBD_FREE_PTR(cbdata);
2333                 }
2334         }
2335
2336         RETURN(rc);
2337 }
2338 EXPORT_SYMBOL(ptlrpc_set_wait);
2339
2340 /**
2341  * Helper fuction for request freeing.
2342  * Called when request count reached zero and request needs to be freed.
2343  * Removes request from all sorts of sending/replay lists it might be on,
2344  * frees network buffers if any are present.
2345  * If \a locked is set, that means caller is already holding import imp_lock
2346  * and so we no longer need to reobtain it (for certain lists manipulations)
2347  */
2348 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2349 {
2350         ENTRY;
2351
2352         if (request == NULL)
2353                 RETURN_EXIT;
2354
2355         LASSERT(!request->rq_srv_req);
2356         LASSERT(request->rq_export == NULL);
2357         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2358         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2359         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2360         LASSERTF(!request->rq_replay, "req %p\n", request);
2361
2362         req_capsule_fini(&request->rq_pill);
2363
2364         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
2365          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2366         if (request->rq_import != NULL) {
2367                 if (!locked)
2368                         spin_lock(&request->rq_import->imp_lock);
2369                 list_del_init(&request->rq_replay_list);
2370                 if (!locked)
2371                         spin_unlock(&request->rq_import->imp_lock);
2372         }
2373         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2374
2375         if (atomic_read(&request->rq_refcount) != 0) {
2376                 DEBUG_REQ(D_ERROR, request,
2377                           "freeing request with nonzero refcount");
2378                 LBUG();
2379         }
2380
2381         if (request->rq_repbuf != NULL)
2382                 sptlrpc_cli_free_repbuf(request);
2383
2384         if (request->rq_import != NULL) {
2385                 class_import_put(request->rq_import);
2386                 request->rq_import = NULL;
2387         }
2388         if (request->rq_bulk != NULL)
2389                 ptlrpc_free_bulk(request->rq_bulk);
2390
2391         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2392                 sptlrpc_cli_free_reqbuf(request);
2393
2394         if (request->rq_cli_ctx)
2395                 sptlrpc_req_put_ctx(request, !locked);
2396
2397         if (request->rq_pool)
2398                 __ptlrpc_free_req_to_pool(request);
2399         else
2400                 ptlrpc_request_cache_free(request);
2401         EXIT;
2402 }
2403
2404 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2405 /**
2406  * Drop one request reference. Must be called with import imp_lock held.
2407  * When reference count drops to zero, request is freed.
2408  */
2409 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2410 {
2411         assert_spin_locked(&request->rq_import->imp_lock);
2412         (void)__ptlrpc_req_finished(request, 1);
2413 }
2414
2415 /**
2416  * Helper function
2417  * Drops one reference count for request \a request.
2418  * \a locked set indicates that caller holds import imp_lock.
2419  * Frees the request whe reference count reaches zero.
2420  */
2421 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2422 {
2423         ENTRY;
2424         if (request == NULL)
2425                 RETURN(1);
2426
2427         if (request == LP_POISON ||
2428             request->rq_reqmsg == LP_POISON) {
2429                 CERROR("dereferencing freed request (bug 575)\n");
2430                 LBUG();
2431                 RETURN(1);
2432         }
2433
2434         DEBUG_REQ(D_INFO, request, "refcount now %u",
2435                   atomic_read(&request->rq_refcount) - 1);
2436
2437         if (atomic_dec_and_test(&request->rq_refcount)) {
2438                 __ptlrpc_free_req(request, locked);
2439                 RETURN(1);
2440         }
2441
2442         RETURN(0);
2443 }
2444
2445 /**
2446  * Drops one reference count for a request.
2447  */
2448 void ptlrpc_req_finished(struct ptlrpc_request *request)
2449 {
2450         __ptlrpc_req_finished(request, 0);
2451 }
2452 EXPORT_SYMBOL(ptlrpc_req_finished);
2453
2454 /**
2455  * Returns xid of a \a request
2456  */
2457 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2458 {
2459         return request->rq_xid;
2460 }
2461 EXPORT_SYMBOL(ptlrpc_req_xid);
2462
2463 /**
2464  * Disengage the client's reply buffer from the network
2465  * NB does _NOT_ unregister any client-side bulk.
2466  * IDEMPOTENT, but _not_ safe against concurrent callers.
2467  * The request owner (i.e. the thread doing the I/O) must call...
2468  * Returns 0 on success or 1 if unregistering cannot be made.
2469  */
2470 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2471 {
2472         int                rc;
2473         struct l_wait_info lwi;
2474
2475         /*
2476          * Might sleep.
2477          */
2478         LASSERT(!in_interrupt());
2479
2480         /*
2481          * Let's setup deadline for reply unlink.
2482          */
2483         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2484             async && request->rq_reply_deadline == 0)
2485                 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2486
2487         /*
2488          * Nothing left to do.
2489          */
2490         if (!ptlrpc_client_recv_or_unlink(request))
2491                 RETURN(1);
2492
2493         LNetMDUnlink(request->rq_reply_md_h);
2494
2495         /*
2496          * Let's check it once again.
2497          */
2498         if (!ptlrpc_client_recv_or_unlink(request))
2499                 RETURN(1);
2500
2501         /*
2502          * Move to "Unregistering" phase as reply was not unlinked yet.
2503          */
2504         ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2505
2506         /*
2507          * Do not wait for unlink to finish.
2508          */
2509         if (async)
2510                 RETURN(0);
2511
2512         /*
2513          * We have to l_wait_event() whatever the result, to give liblustre
2514          * a chance to run reply_in_callback(), and to make sure we've
2515          * unlinked before returning a req to the pool.
2516          */
2517         for (;;) {
2518                 /* The wq argument is ignored by user-space wait_event macros */
2519                 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2520                                         &request->rq_set->set_waitq :
2521                                         &request->rq_reply_waitq;
2522                 /* Network access will complete in finite time but the HUGE
2523                  * timeout lets us CWARN for visibility of sluggish NALs */
2524                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2525                                            cfs_time_seconds(1), NULL, NULL);
2526                 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2527                                   &lwi);
2528                 if (rc == 0) {
2529                         ptlrpc_rqphase_move(request, request->rq_next_phase);
2530                         RETURN(1);
2531                 }
2532
2533                 LASSERT(rc == -ETIMEDOUT);
2534                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2535                           "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2536                           request->rq_receiving_reply,
2537                           request->rq_req_unlinked,
2538                           request->rq_reply_unlinked);
2539         }
2540         RETURN(0);
2541 }
2542
2543 static void ptlrpc_free_request(struct ptlrpc_request *req)
2544 {
2545         spin_lock(&req->rq_lock);
2546         req->rq_replay = 0;
2547         spin_unlock(&req->rq_lock);
2548
2549         if (req->rq_commit_cb != NULL)
2550                 req->rq_commit_cb(req);
2551         list_del_init(&req->rq_replay_list);
2552
2553         __ptlrpc_req_finished(req, 1);
2554 }
2555
2556 /**
2557  * the request is committed and dropped from the replay list of its import
2558  */
2559 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2560 {
2561         struct obd_import       *imp = req->rq_import;
2562
2563         spin_lock(&imp->imp_lock);
2564         if (list_empty(&req->rq_replay_list)) {
2565                 spin_unlock(&imp->imp_lock);
2566                 return;
2567         }
2568
2569         if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2570                 ptlrpc_free_request(req);
2571
2572         spin_unlock(&imp->imp_lock);
2573 }
2574 EXPORT_SYMBOL(ptlrpc_request_committed);
2575
2576 /**
2577  * Iterates through replay_list on import and prunes
2578  * all requests have transno smaller than last_committed for the
2579  * import and don't have rq_replay set.
2580  * Since requests are sorted in transno order, stops when meetign first
2581  * transno bigger than last_committed.
2582  * caller must hold imp->imp_lock
2583  */
2584 void ptlrpc_free_committed(struct obd_import *imp)
2585 {
2586         struct ptlrpc_request   *req, *saved;
2587         struct ptlrpc_request   *last_req = NULL; /* temporary fire escape */
2588         bool                     skip_committed_list = true;
2589         ENTRY;
2590
2591         LASSERT(imp != NULL);
2592         assert_spin_locked(&imp->imp_lock);
2593
2594         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2595             imp->imp_generation == imp->imp_last_generation_checked) {
2596                 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2597                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2598                 RETURN_EXIT;
2599         }
2600         CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2601                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2602                imp->imp_generation);
2603
2604         if (imp->imp_generation != imp->imp_last_generation_checked ||
2605             imp->imp_last_transno_checked == 0)
2606                 skip_committed_list = false;
2607
2608         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2609         imp->imp_last_generation_checked = imp->imp_generation;
2610
2611         list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2612                                      rq_replay_list) {
2613                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
2614                 LASSERT(req != last_req);
2615                 last_req = req;
2616
2617                 if (req->rq_transno == 0) {
2618                         DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2619                         LBUG();
2620                 }
2621                 if (req->rq_import_generation < imp->imp_generation) {
2622                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2623                         GOTO(free_req, 0);
2624                 }
2625
2626                 /* not yet committed */
2627                 if (req->rq_transno > imp->imp_peer_committed_transno) {
2628                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2629                         break;
2630                 }
2631
2632                 if (req->rq_replay) {
2633                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2634                         list_move_tail(&req->rq_replay_list,
2635                                            &imp->imp_committed_list);
2636                         continue;
2637                 }
2638
2639                 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2640                           imp->imp_peer_committed_transno);
2641 free_req:
2642                 ptlrpc_free_request(req);
2643         }
2644
2645         if (skip_committed_list)
2646                 GOTO(out, 0);
2647
2648         list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2649                                      rq_replay_list) {
2650                 LASSERT(req->rq_transno != 0);
2651                 if (req->rq_import_generation < imp->imp_generation) {
2652                         DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2653                         ptlrpc_free_request(req);
2654                 } else if (!req->rq_replay) {
2655                         DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2656                         ptlrpc_free_request(req);
2657                 }
2658         }
2659 out:
2660         EXIT;
2661 }
2662
2663 void ptlrpc_cleanup_client(struct obd_import *imp)
2664 {
2665         ENTRY;
2666         EXIT;
2667 }
2668
2669 /**
2670  * Schedule previously sent request for resend.
2671  * For bulk requests we assign new xid (to avoid problems with
2672  * lost replies and therefore several transfers landing into same buffer
2673  * from different sending attempts).
2674  */
2675 void ptlrpc_resend_req(struct ptlrpc_request *req)
2676 {
2677         DEBUG_REQ(D_HA, req, "going to resend");
2678         spin_lock(&req->rq_lock);
2679
2680         /* Request got reply but linked to the import list still.
2681            Let ptlrpc_check_set() to process it. */
2682         if (ptlrpc_client_replied(req)) {
2683                 spin_unlock(&req->rq_lock);
2684                 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2685                 return;
2686         }
2687
2688         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2689         req->rq_status = -EAGAIN;
2690
2691         req->rq_resend = 1;
2692         req->rq_net_err = 0;
2693         req->rq_timedout = 0;
2694         if (req->rq_bulk) {
2695                 __u64 old_xid = req->rq_xid;
2696
2697                 /* ensure previous bulk fails */
2698                 req->rq_xid = ptlrpc_next_xid();
2699                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2700                        old_xid, req->rq_xid);
2701         }
2702         ptlrpc_client_wake_req(req);
2703         spin_unlock(&req->rq_lock);
2704 }
2705
2706 /* XXX: this function and rq_status are currently unused */
2707 void ptlrpc_restart_req(struct ptlrpc_request *req)
2708 {
2709         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2710         req->rq_status = -ERESTARTSYS;
2711
2712         spin_lock(&req->rq_lock);
2713         req->rq_restart = 1;
2714         req->rq_timedout = 0;
2715         ptlrpc_client_wake_req(req);
2716         spin_unlock(&req->rq_lock);
2717 }
2718
2719 /**
2720  * Grab additional reference on a request \a req
2721  */
2722 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2723 {
2724         ENTRY;
2725         atomic_inc(&req->rq_refcount);
2726         RETURN(req);
2727 }
2728 EXPORT_SYMBOL(ptlrpc_request_addref);
2729
2730 /**
2731  * Add a request to import replay_list.
2732  * Must be called under imp_lock
2733  */
2734 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2735                                       struct obd_import *imp)
2736 {
2737         struct list_head *tmp;
2738
2739         assert_spin_locked(&imp->imp_lock);
2740
2741         if (req->rq_transno == 0) {
2742                 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2743                 LBUG();
2744         }
2745
2746         /* clear this for new requests that were resent as well
2747            as resent replayed requests. */
2748         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2749
2750         /* don't re-add requests that have been replayed */
2751         if (!list_empty(&req->rq_replay_list))
2752                 return;
2753
2754         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2755
2756         LASSERT(imp->imp_replayable);
2757         /* Balanced in ptlrpc_free_committed, usually. */
2758         ptlrpc_request_addref(req);
2759         list_for_each_prev(tmp, &imp->imp_replay_list) {
2760                 struct ptlrpc_request *iter = list_entry(tmp,
2761                                                          struct ptlrpc_request,
2762                                                          rq_replay_list);
2763
2764                 /* We may have duplicate transnos if we create and then
2765                  * open a file, or for closes retained if to match creating
2766                  * opens, so use req->rq_xid as a secondary key.
2767                  * (See bugs 684, 685, and 428.)
2768                  * XXX no longer needed, but all opens need transnos!
2769                  */
2770                 if (iter->rq_transno > req->rq_transno)
2771                         continue;
2772
2773                 if (iter->rq_transno == req->rq_transno) {
2774                         LASSERT(iter->rq_xid != req->rq_xid);
2775                         if (iter->rq_xid > req->rq_xid)
2776                                 continue;
2777                 }
2778
2779                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2780                 return;
2781         }
2782
2783         list_add(&req->rq_replay_list, &imp->imp_replay_list);
2784 }
2785
2786 /**
2787  * Send request and wait until it completes.
2788  * Returns request processing status.
2789  */
2790 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2791 {
2792         struct ptlrpc_request_set *set;
2793         int rc;
2794         ENTRY;
2795
2796         LASSERT(req->rq_set == NULL);
2797         LASSERT(!req->rq_receiving_reply);
2798
2799         set = ptlrpc_prep_set();
2800         if (set == NULL) {
2801                 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2802                 RETURN(-ENOMEM);
2803         }
2804
2805         /* for distributed debugging */
2806         lustre_msg_set_status(req->rq_reqmsg, current_pid());
2807
2808         /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2809         ptlrpc_request_addref(req);
2810         ptlrpc_set_add_req(set, req);
2811         rc = ptlrpc_set_wait(set);
2812         ptlrpc_set_destroy(set);
2813
2814         RETURN(rc);
2815 }
2816 EXPORT_SYMBOL(ptlrpc_queue_wait);
2817
2818 /**
2819  * Callback used for replayed requests reply processing.
2820  * In case of successful reply calls registered request replay callback.
2821  * In case of error restart replay process.
2822  */
2823 static int ptlrpc_replay_interpret(const struct lu_env *env,
2824                                    struct ptlrpc_request *req,
2825                                    void * data, int rc)
2826 {
2827         struct ptlrpc_replay_async_args *aa = data;
2828         struct obd_import *imp = req->rq_import;
2829
2830         ENTRY;
2831         atomic_dec(&imp->imp_replay_inflight);
2832
2833         if (!ptlrpc_client_replied(req)) {
2834                 CERROR("request replay timed out, restarting recovery\n");
2835                 GOTO(out, rc = -ETIMEDOUT);
2836         }
2837
2838         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2839             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2840              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2841                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2842
2843         /** VBR: check version failure */
2844         if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2845                 /** replay was failed due to version mismatch */
2846                 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2847                 spin_lock(&imp->imp_lock);
2848                 imp->imp_vbr_failed = 1;
2849                 imp->imp_no_lock_replay = 1;
2850                 spin_unlock(&imp->imp_lock);
2851                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2852         } else {
2853                 /** The transno had better not change over replay. */
2854                 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2855                          lustre_msg_get_transno(req->rq_repmsg) ||
2856                          lustre_msg_get_transno(req->rq_repmsg) == 0,
2857                          LPX64"/"LPX64"\n",
2858                          lustre_msg_get_transno(req->rq_reqmsg),
2859                          lustre_msg_get_transno(req->rq_repmsg));
2860         }
2861
2862         spin_lock(&imp->imp_lock);
2863         /** if replays by version then gap occur on server, no trust to locks */
2864         if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2865                 imp->imp_no_lock_replay = 1;
2866         imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2867         spin_unlock(&imp->imp_lock);
2868         LASSERT(imp->imp_last_replay_transno);
2869
2870         /* transaction number shouldn't be bigger than the latest replayed */
2871         if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2872                 DEBUG_REQ(D_ERROR, req,
2873                           "Reported transno "LPU64" is bigger than the "
2874                           "replayed one: "LPU64, req->rq_transno,
2875                           lustre_msg_get_transno(req->rq_reqmsg));
2876                 GOTO(out, rc = -EINVAL);
2877         }
2878
2879         DEBUG_REQ(D_HA, req, "got rep");
2880
2881         /* let the callback do fixups, possibly including in the request */
2882         if (req->rq_replay_cb)
2883                 req->rq_replay_cb(req);
2884
2885         if (ptlrpc_client_replied(req) &&
2886             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2887                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2888                           lustre_msg_get_status(req->rq_repmsg),
2889                           aa->praa_old_status);
2890         } else {
2891                 /* Put it back for re-replay. */
2892                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2893         }
2894
2895         /*
2896          * Errors while replay can set transno to 0, but
2897          * imp_last_replay_transno shouldn't be set to 0 anyway
2898          */
2899         if (req->rq_transno == 0)
2900                 CERROR("Transno is 0 during replay!\n");
2901
2902         /* continue with recovery */
2903         rc = ptlrpc_import_recovery_state_machine(imp);
2904  out:
2905         req->rq_send_state = aa->praa_old_state;
2906
2907         if (rc != 0)
2908                 /* this replay failed, so restart recovery */
2909                 ptlrpc_connect_import(imp);
2910
2911         RETURN(rc);
2912 }
2913
2914 /**
2915  * Prepares and queues request for replay.
2916  * Adds it to ptlrpcd queue for actual sending.
2917  * Returns 0 on success.
2918  */
2919 int ptlrpc_replay_req(struct ptlrpc_request *req)
2920 {
2921         struct ptlrpc_replay_async_args *aa;
2922         ENTRY;
2923
2924         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2925
2926         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2927         aa = ptlrpc_req_async_args(req);
2928         memset(aa, 0, sizeof *aa);
2929
2930         /* Prepare request to be resent with ptlrpcd */
2931         aa->praa_old_state = req->rq_send_state;
2932         req->rq_send_state = LUSTRE_IMP_REPLAY;
2933         req->rq_phase = RQ_PHASE_NEW;
2934         req->rq_next_phase = RQ_PHASE_UNDEFINED;
2935         if (req->rq_repmsg)
2936                 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2937         req->rq_status = 0;
2938         req->rq_interpret_reply = ptlrpc_replay_interpret;
2939         /* Readjust the timeout for current conditions */
2940         ptlrpc_at_set_req_timeout(req);
2941
2942         /* Tell server the net_latency, so the server can calculate how long
2943          * it should wait for next replay */
2944         lustre_msg_set_service_time(req->rq_reqmsg,
2945                                     ptlrpc_at_get_net_latency(req));
2946         DEBUG_REQ(D_HA, req, "REPLAY");
2947
2948         atomic_inc(&req->rq_import->imp_replay_inflight);
2949         ptlrpc_request_addref(req);     /* ptlrpcd needs a ref */
2950
2951         ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2952         RETURN(0);
2953 }
2954
2955 /**
2956  * Aborts all in-flight request on import \a imp sending and delayed lists
2957  */
2958 void ptlrpc_abort_inflight(struct obd_import *imp)
2959 {
2960         struct list_head *tmp, *n;
2961         ENTRY;
2962
2963         /* Make sure that no new requests get processed for this import.
2964          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2965          * this flag and then putting requests on sending_list or delayed_list.
2966          */
2967         spin_lock(&imp->imp_lock);
2968
2969         /* XXX locking?  Maybe we should remove each request with the list
2970          * locked?  Also, how do we know if the requests on the list are
2971          * being freed at this time?
2972          */
2973         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2974                 struct ptlrpc_request *req = list_entry(tmp,
2975                                                         struct ptlrpc_request,
2976                                                         rq_list);
2977
2978                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2979
2980                 spin_lock(&req->rq_lock);
2981                 if (req->rq_import_generation < imp->imp_generation) {
2982                         req->rq_err = 1;
2983                         req->rq_status = -EIO;
2984                         ptlrpc_client_wake_req(req);
2985                 }
2986                 spin_unlock(&req->rq_lock);
2987         }
2988
2989         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2990                 struct ptlrpc_request *req =
2991                         list_entry(tmp, struct ptlrpc_request, rq_list);
2992
2993                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2994
2995                 spin_lock(&req->rq_lock);
2996                 if (req->rq_import_generation < imp->imp_generation) {
2997                         req->rq_err = 1;
2998                         req->rq_status = -EIO;
2999                         ptlrpc_client_wake_req(req);
3000                 }
3001                 spin_unlock(&req->rq_lock);
3002         }
3003
3004         /* Last chance to free reqs left on the replay list, but we
3005          * will still leak reqs that haven't committed.  */
3006         if (imp->imp_replayable)
3007                 ptlrpc_free_committed(imp);
3008
3009         spin_unlock(&imp->imp_lock);
3010
3011         EXIT;
3012 }
3013
3014 /**
3015  * Abort all uncompleted requests in request set \a set
3016  */
3017 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3018 {
3019         struct list_head *tmp, *pos;
3020
3021         LASSERT(set != NULL);
3022
3023         list_for_each_safe(pos, tmp, &set->set_requests) {
3024                 struct ptlrpc_request *req =
3025                         list_entry(pos, struct ptlrpc_request,
3026                                    rq_set_chain);
3027
3028                 spin_lock(&req->rq_lock);
3029                 if (req->rq_phase != RQ_PHASE_RPC) {
3030                         spin_unlock(&req->rq_lock);
3031                         continue;
3032                 }
3033
3034                 req->rq_err = 1;
3035                 req->rq_status = -EINTR;
3036                 ptlrpc_client_wake_req(req);
3037                 spin_unlock(&req->rq_lock);
3038         }
3039 }
3040
3041 static __u64 ptlrpc_last_xid;
3042 static spinlock_t ptlrpc_last_xid_lock;
3043
3044 /**
3045  * Initialize the XID for the node.  This is common among all requests on
3046  * this node, and only requires the property that it is monotonically
3047  * increasing.  It does not need to be sequential.  Since this is also used
3048  * as the RDMA match bits, it is important that a single client NOT have
3049  * the same match bits for two different in-flight requests, hence we do
3050  * NOT want to have an XID per target or similar.
3051  *
3052  * To avoid an unlikely collision between match bits after a client reboot
3053  * (which would deliver old data into the wrong RDMA buffer) initialize
3054  * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3055  * If the time is clearly incorrect, we instead use a 62-bit random number.
3056  * In the worst case the random number will overflow 1M RPCs per second in
3057  * 9133 years, or permutations thereof.
3058  */
3059 #define YEAR_2004 (1ULL << 30)
3060 void ptlrpc_init_xid(void)
3061 {
3062         time_t now = cfs_time_current_sec();
3063
3064         spin_lock_init(&ptlrpc_last_xid_lock);
3065         if (now < YEAR_2004) {
3066                 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3067                 ptlrpc_last_xid >>= 2;
3068                 ptlrpc_last_xid |= (1ULL << 61);
3069         } else {
3070                 ptlrpc_last_xid = (__u64)now << 20;
3071         }
3072
3073         /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3074         CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3075         ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3076 }
3077
3078 /**
3079  * Increase xid and returns resulting new value to the caller.
3080  *
3081  * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3082  * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3083  * itself uses the last bulk xid needed, so the server can determine the
3084  * the number of bulk transfers from the RPC XID and a bitmask.  The starting
3085  * xid must align to a power-of-two value.
3086  *
3087  * This is assumed to be true due to the initial ptlrpc_last_xid
3088  * value also being initialized to a power-of-two value. LU-1431
3089  */
3090 __u64 ptlrpc_next_xid(void)
3091 {
3092         __u64 next;
3093
3094         spin_lock(&ptlrpc_last_xid_lock);
3095         next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3096         ptlrpc_last_xid = next;
3097         spin_unlock(&ptlrpc_last_xid_lock);
3098
3099         return next;
3100 }
3101
3102 /**
3103  * Get a glimpse at what next xid value might have been.
3104  * Returns possible next xid.
3105  */
3106 __u64 ptlrpc_sample_next_xid(void)
3107 {
3108 #if BITS_PER_LONG == 32
3109         /* need to avoid possible word tearing on 32-bit systems */
3110         __u64 next;
3111
3112         spin_lock(&ptlrpc_last_xid_lock);
3113         next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3114         spin_unlock(&ptlrpc_last_xid_lock);
3115
3116         return next;
3117 #else
3118         /* No need to lock, since returned value is racy anyways */
3119         return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3120 #endif
3121 }
3122 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3123
3124 /**
3125  * Functions for operating ptlrpc workers.
3126  *
3127  * A ptlrpc work is a function which will be running inside ptlrpc context.
3128  * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3129  *
3130  * 1. after a work is created, it can be used many times, that is:
3131  *         handler = ptlrpcd_alloc_work();
3132  *         ptlrpcd_queue_work();
3133  *
3134  *    queue it again when necessary:
3135  *         ptlrpcd_queue_work();
3136  *         ptlrpcd_destroy_work();
3137  * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3138  *    it will only be queued once in any time. Also as its name implies, it may
3139  *    have delay before it really runs by ptlrpcd thread.
3140  */
3141 struct ptlrpc_work_async_args {
3142         int   (*cb)(const struct lu_env *, void *);
3143         void   *cbdata;
3144 };
3145
3146 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3147 {
3148         /* re-initialize the req */
3149         req->rq_timeout         = obd_timeout;
3150         req->rq_sent            = cfs_time_current_sec();
3151         req->rq_deadline        = req->rq_sent + req->rq_timeout;
3152         req->rq_reply_deadline  = req->rq_deadline;
3153         req->rq_phase           = RQ_PHASE_INTERPRET;
3154         req->rq_next_phase      = RQ_PHASE_COMPLETE;
3155         req->rq_xid             = ptlrpc_next_xid();
3156         req->rq_import_generation = req->rq_import->imp_generation;
3157
3158         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3159 }
3160
3161 static int work_interpreter(const struct lu_env *env,
3162                             struct ptlrpc_request *req, void *data, int rc)
3163 {
3164         struct ptlrpc_work_async_args *arg = data;
3165
3166         LASSERT(ptlrpcd_check_work(req));
3167         LASSERT(arg->cb != NULL);
3168
3169         rc = arg->cb(env, arg->cbdata);
3170
3171         list_del_init(&req->rq_set_chain);
3172         req->rq_set = NULL;
3173
3174         if (atomic_dec_return(&req->rq_refcount) > 1) {
3175                 atomic_set(&req->rq_refcount, 2);
3176                 ptlrpcd_add_work_req(req);
3177         }
3178         return rc;
3179 }
3180
3181 static int worker_format;
3182
3183 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3184 {
3185         return req->rq_pill.rc_fmt == (void *)&worker_format;
3186 }
3187
3188 /**
3189  * Create a work for ptlrpc.
3190  */
3191 void *ptlrpcd_alloc_work(struct obd_import *imp,
3192                          int (*cb)(const struct lu_env *, void *), void *cbdata)
3193 {
3194         struct ptlrpc_request         *req = NULL;
3195         struct ptlrpc_work_async_args *args;
3196         ENTRY;
3197
3198         might_sleep();
3199
3200         if (cb == NULL)
3201                 RETURN(ERR_PTR(-EINVAL));
3202
3203         /* copy some code from deprecated fakereq. */
3204         req = ptlrpc_request_cache_alloc(GFP_NOFS);
3205         if (req == NULL) {
3206                 CERROR("ptlrpc: run out of memory!\n");
3207                 RETURN(ERR_PTR(-ENOMEM));
3208         }
3209
3210         ptlrpc_cli_req_init(req);
3211
3212         req->rq_send_state = LUSTRE_IMP_FULL;
3213         req->rq_type = PTL_RPC_MSG_REQUEST;
3214         req->rq_import = class_import_get(imp);
3215         req->rq_interpret_reply = work_interpreter;
3216         /* don't want reply */
3217         req->rq_no_delay = req->rq_no_resend = 1;
3218         req->rq_pill.rc_fmt = (void *)&worker_format;
3219
3220         CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3221         args = ptlrpc_req_async_args(req);
3222         args->cb     = cb;
3223         args->cbdata = cbdata;
3224
3225         RETURN(req);
3226 }
3227 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3228
3229 void ptlrpcd_destroy_work(void *handler)
3230 {
3231         struct ptlrpc_request *req = handler;
3232
3233         if (req)
3234                 ptlrpc_req_finished(req);
3235 }
3236 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3237
3238 int ptlrpcd_queue_work(void *handler)
3239 {
3240         struct ptlrpc_request *req = handler;
3241
3242         /*
3243          * Check if the req is already being queued.
3244          *
3245          * Here comes a trick: it lacks a way of checking if a req is being