Whamcloud - gitweb
4cb6a056f2b24ddc4232071d771614be0fb677ac
[fs/lustre-release.git] / lustre / osc / osc_request.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, 2013, 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 #define DEBUG_SUBSYSTEM S_OSC
38
39 #include <libcfs/libcfs.h>
40
41 #ifndef __KERNEL__
42 # include <liblustre.h>
43 #endif
44
45 #include <lustre_dlm.h>
46 #include <lustre_net.h>
47 #include <lustre/lustre_user.h>
48 #include <obd_cksum.h>
49 #include <obd_ost.h>
50 #include <obd_lov.h>
51
52 #ifdef  __CYGWIN__
53 # include <ctype.h>
54 #endif
55
56 #include <lustre_ha.h>
57 #include <lprocfs_status.h>
58 #include <lustre_log.h>
59 #include <lustre_debug.h>
60 #include <lustre_param.h>
61 #include <lustre_fid.h>
62 #include "osc_internal.h"
63 #include "osc_cl_internal.h"
64
65 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
66 static int brw_interpret(const struct lu_env *env,
67                          struct ptlrpc_request *req, void *data, int rc);
68 int osc_cleanup(struct obd_device *obd);
69
70 /* Pack OSC object metadata for disk storage (LE byte order). */
71 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
72                       struct lov_stripe_md *lsm)
73 {
74         int lmm_size;
75         ENTRY;
76
77         lmm_size = sizeof(**lmmp);
78         if (lmmp == NULL)
79                 RETURN(lmm_size);
80
81         if (*lmmp != NULL && lsm == NULL) {
82                 OBD_FREE(*lmmp, lmm_size);
83                 *lmmp = NULL;
84                 RETURN(0);
85         } else if (unlikely(lsm != NULL && ostid_id(&lsm->lsm_oi) == 0)) {
86                 RETURN(-EBADF);
87         }
88
89         if (*lmmp == NULL) {
90                 OBD_ALLOC(*lmmp, lmm_size);
91                 if (*lmmp == NULL)
92                         RETURN(-ENOMEM);
93         }
94
95         if (lsm)
96                 ostid_cpu_to_le(&lsm->lsm_oi, &(*lmmp)->lmm_oi);
97
98         RETURN(lmm_size);
99 }
100
101 /* Unpack OSC object metadata from disk storage (LE byte order). */
102 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
103                         struct lov_mds_md *lmm, int lmm_bytes)
104 {
105         int lsm_size;
106         struct obd_import *imp = class_exp2cliimp(exp);
107         ENTRY;
108
109         if (lmm != NULL) {
110                 if (lmm_bytes < sizeof(*lmm)) {
111                         CERROR("%s: lov_mds_md too small: %d, need %d\n",
112                                exp->exp_obd->obd_name, lmm_bytes,
113                                (int)sizeof(*lmm));
114                         RETURN(-EINVAL);
115                 }
116                 /* XXX LOV_MAGIC etc check? */
117
118                 if (unlikely(ostid_id(&lmm->lmm_oi) == 0)) {
119                         CERROR("%s: zero lmm_object_id: rc = %d\n",
120                                exp->exp_obd->obd_name, -EINVAL);
121                         RETURN(-EINVAL);
122                 }
123         }
124
125         lsm_size = lov_stripe_md_size(1);
126         if (lsmp == NULL)
127                 RETURN(lsm_size);
128
129         if (*lsmp != NULL && lmm == NULL) {
130                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
131                 OBD_FREE(*lsmp, lsm_size);
132                 *lsmp = NULL;
133                 RETURN(0);
134         }
135
136         if (*lsmp == NULL) {
137                 OBD_ALLOC(*lsmp, lsm_size);
138                 if (unlikely(*lsmp == NULL))
139                         RETURN(-ENOMEM);
140                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
141                 if (unlikely((*lsmp)->lsm_oinfo[0] == NULL)) {
142                         OBD_FREE(*lsmp, lsm_size);
143                         RETURN(-ENOMEM);
144                 }
145                 loi_init((*lsmp)->lsm_oinfo[0]);
146         } else if (unlikely(ostid_id(&(*lsmp)->lsm_oi) == 0)) {
147                 RETURN(-EBADF);
148         }
149
150         if (lmm != NULL)
151                 /* XXX zero *lsmp? */
152                 ostid_le_to_cpu(&lmm->lmm_oi, &(*lsmp)->lsm_oi);
153
154         if (imp != NULL &&
155             (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
156                 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
157         else
158                 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
159
160         RETURN(lsm_size);
161 }
162
163 static inline void osc_pack_capa(struct ptlrpc_request *req,
164                                  struct ost_body *body, void *capa)
165 {
166         struct obd_capa *oc = (struct obd_capa *)capa;
167         struct lustre_capa *c;
168
169         if (!capa)
170                 return;
171
172         c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
173         LASSERT(c);
174         capa_cpy(c, oc);
175         body->oa.o_valid |= OBD_MD_FLOSSCAPA;
176         DEBUG_CAPA(D_SEC, c, "pack");
177 }
178
179 static inline void osc_pack_req_body(struct ptlrpc_request *req,
180                                      struct obd_info *oinfo)
181 {
182         struct ost_body *body;
183
184         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
185         LASSERT(body);
186
187         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
188                              oinfo->oi_oa);
189         osc_pack_capa(req, body, oinfo->oi_capa);
190 }
191
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193                                      const struct req_msg_field *field,
194                                      struct obd_capa *oc)
195 {
196         if (oc == NULL)
197                 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
198         else
199                 /* it is already calculated as sizeof struct obd_capa */
200                 ;
201 }
202
203 static int osc_getattr_interpret(const struct lu_env *env,
204                                  struct ptlrpc_request *req,
205                                  struct osc_async_args *aa, int rc)
206 {
207         struct ost_body *body;
208         ENTRY;
209
210         if (rc != 0)
211                 GOTO(out, rc);
212
213         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
214         if (body) {
215                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
216                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
217                                      aa->aa_oi->oi_oa, &body->oa);
218
219                 /* This should really be sent by the OST */
220                 aa->aa_oi->oi_oa->o_blksize = DT_MAX_BRW_SIZE;
221                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
222         } else {
223                 CDEBUG(D_INFO, "can't unpack ost_body\n");
224                 rc = -EPROTO;
225                 aa->aa_oi->oi_oa->o_valid = 0;
226         }
227 out:
228         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
229         RETURN(rc);
230 }
231
232 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
233                              struct ptlrpc_request_set *set)
234 {
235         struct ptlrpc_request *req;
236         struct osc_async_args *aa;
237         int                    rc;
238         ENTRY;
239
240         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
241         if (req == NULL)
242                 RETURN(-ENOMEM);
243
244         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
245         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
246         if (rc) {
247                 ptlrpc_request_free(req);
248                 RETURN(rc);
249         }
250
251         osc_pack_req_body(req, oinfo);
252
253         ptlrpc_request_set_replen(req);
254         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
255
256         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
257         aa = ptlrpc_req_async_args(req);
258         aa->aa_oi = oinfo;
259
260         ptlrpc_set_add_req(set, req);
261         RETURN(0);
262 }
263
264 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
265                        struct obd_info *oinfo)
266 {
267         struct ptlrpc_request *req;
268         struct ost_body       *body;
269         int                    rc;
270         ENTRY;
271
272         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
273         if (req == NULL)
274                 RETURN(-ENOMEM);
275
276         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
277         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
278         if (rc) {
279                 ptlrpc_request_free(req);
280                 RETURN(rc);
281         }
282
283         osc_pack_req_body(req, oinfo);
284
285         ptlrpc_request_set_replen(req);
286
287         rc = ptlrpc_queue_wait(req);
288         if (rc)
289                 GOTO(out, rc);
290
291         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
292         if (body == NULL)
293                 GOTO(out, rc = -EPROTO);
294
295         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
296         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
297                              &body->oa);
298
299         oinfo->oi_oa->o_blksize = cli_brw_size(exp->exp_obd);
300         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
301
302         EXIT;
303  out:
304         ptlrpc_req_finished(req);
305         return rc;
306 }
307
308 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
309                        struct obd_info *oinfo, struct obd_trans_info *oti)
310 {
311         struct ptlrpc_request *req;
312         struct ost_body       *body;
313         int                    rc;
314         ENTRY;
315
316         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
317
318         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
319         if (req == NULL)
320                 RETURN(-ENOMEM);
321
322         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
323         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
324         if (rc) {
325                 ptlrpc_request_free(req);
326                 RETURN(rc);
327         }
328
329         osc_pack_req_body(req, oinfo);
330
331         ptlrpc_request_set_replen(req);
332
333         rc = ptlrpc_queue_wait(req);
334         if (rc)
335                 GOTO(out, rc);
336
337         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
338         if (body == NULL)
339                 GOTO(out, rc = -EPROTO);
340
341         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
342                              &body->oa);
343
344         EXIT;
345 out:
346         ptlrpc_req_finished(req);
347         RETURN(rc);
348 }
349
350 static int osc_setattr_interpret(const struct lu_env *env,
351                                  struct ptlrpc_request *req,
352                                  struct osc_setattr_args *sa, int rc)
353 {
354         struct ost_body *body;
355         ENTRY;
356
357         if (rc != 0)
358                 GOTO(out, rc);
359
360         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
361         if (body == NULL)
362                 GOTO(out, rc = -EPROTO);
363
364         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
365                              &body->oa);
366 out:
367         rc = sa->sa_upcall(sa->sa_cookie, rc);
368         RETURN(rc);
369 }
370
371 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
372                            struct obd_trans_info *oti,
373                            obd_enqueue_update_f upcall, void *cookie,
374                            struct ptlrpc_request_set *rqset)
375 {
376         struct ptlrpc_request   *req;
377         struct osc_setattr_args *sa;
378         int                      rc;
379         ENTRY;
380
381         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
382         if (req == NULL)
383                 RETURN(-ENOMEM);
384
385         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
386         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
387         if (rc) {
388                 ptlrpc_request_free(req);
389                 RETURN(rc);
390         }
391
392         if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
393                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
394
395         osc_pack_req_body(req, oinfo);
396
397         ptlrpc_request_set_replen(req);
398
399         /* do mds to ost setattr asynchronously */
400         if (!rqset) {
401                 /* Do not wait for response. */
402                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
403         } else {
404                 req->rq_interpret_reply =
405                         (ptlrpc_interpterer_t)osc_setattr_interpret;
406
407                 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
408                 sa = ptlrpc_req_async_args(req);
409                 sa->sa_oa = oinfo->oi_oa;
410                 sa->sa_upcall = upcall;
411                 sa->sa_cookie = cookie;
412
413                 if (rqset == PTLRPCD_SET)
414                         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
415                 else
416                         ptlrpc_set_add_req(rqset, req);
417         }
418
419         RETURN(0);
420 }
421
422 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
423                              struct obd_trans_info *oti,
424                              struct ptlrpc_request_set *rqset)
425 {
426         return osc_setattr_async_base(exp, oinfo, oti,
427                                       oinfo->oi_cb_up, oinfo, rqset);
428 }
429
430 int osc_real_create(struct obd_export *exp, struct obdo *oa,
431                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
432 {
433         struct ptlrpc_request *req;
434         struct ost_body       *body;
435         struct lov_stripe_md  *lsm;
436         int                    rc;
437         ENTRY;
438
439         LASSERT(oa);
440         LASSERT(ea);
441
442         lsm = *ea;
443         if (!lsm) {
444                 rc = obd_alloc_memmd(exp, &lsm);
445                 if (rc < 0)
446                         RETURN(rc);
447         }
448
449         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
450         if (req == NULL)
451                 GOTO(out, rc = -ENOMEM);
452
453         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
454         if (rc) {
455                 ptlrpc_request_free(req);
456                 GOTO(out, rc);
457         }
458
459         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
460         LASSERT(body);
461
462         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
463
464         ptlrpc_request_set_replen(req);
465
466         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
467             oa->o_flags == OBD_FL_DELORPHAN) {
468                 DEBUG_REQ(D_HA, req,
469                           "delorphan from OST integration");
470                 /* Don't resend the delorphan req */
471                 req->rq_no_resend = req->rq_no_delay = 1;
472         }
473
474         rc = ptlrpc_queue_wait(req);
475         if (rc)
476                 GOTO(out_req, rc);
477
478         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
479         if (body == NULL)
480                 GOTO(out_req, rc = -EPROTO);
481
482         CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
483         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
484
485         oa->o_blksize = cli_brw_size(exp->exp_obd);
486         oa->o_valid |= OBD_MD_FLBLKSZ;
487
488         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
489          * have valid lsm_oinfo data structs, so don't go touching that.
490          * This needs to be fixed in a big way.
491          */
492         lsm->lsm_oi = oa->o_oi;
493         *ea = lsm;
494
495         if (oti != NULL) {
496                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
497
498                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
499                         if (!oti->oti_logcookies)
500                                 oti_alloc_cookies(oti, 1);
501                         *oti->oti_logcookies = oa->o_lcookie;
502                 }
503         }
504
505         CDEBUG(D_HA, "transno: "LPD64"\n",
506                lustre_msg_get_transno(req->rq_repmsg));
507 out_req:
508         ptlrpc_req_finished(req);
509 out:
510         if (rc && !*ea)
511                 obd_free_memmd(exp, &lsm);
512         RETURN(rc);
513 }
514
515 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
516                    obd_enqueue_update_f upcall, void *cookie,
517                    struct ptlrpc_request_set *rqset)
518 {
519         struct ptlrpc_request   *req;
520         struct osc_setattr_args *sa;
521         struct ost_body         *body;
522         int                      rc;
523         ENTRY;
524
525         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
526         if (req == NULL)
527                 RETURN(-ENOMEM);
528
529         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
530         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
531         if (rc) {
532                 ptlrpc_request_free(req);
533                 RETURN(rc);
534         }
535         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
536         ptlrpc_at_set_req_timeout(req);
537
538         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
539         LASSERT(body);
540         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
541                              oinfo->oi_oa);
542         osc_pack_capa(req, body, oinfo->oi_capa);
543
544         ptlrpc_request_set_replen(req);
545
546         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
547         CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
548         sa = ptlrpc_req_async_args(req);
549         sa->sa_oa     = oinfo->oi_oa;
550         sa->sa_upcall = upcall;
551         sa->sa_cookie = cookie;
552         if (rqset == PTLRPCD_SET)
553                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
554         else
555                 ptlrpc_set_add_req(rqset, req);
556
557         RETURN(0);
558 }
559
560 static int osc_punch(const struct lu_env *env, struct obd_export *exp,
561                      struct obd_info *oinfo, struct obd_trans_info *oti,
562                      struct ptlrpc_request_set *rqset)
563 {
564         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
565         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
566         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
567         return osc_punch_base(exp, oinfo,
568                               oinfo->oi_cb_up, oinfo, rqset);
569 }
570
571 static int osc_sync_interpret(const struct lu_env *env,
572                               struct ptlrpc_request *req,
573                               void *arg, int rc)
574 {
575         struct osc_fsync_args *fa = arg;
576         struct ost_body *body;
577         ENTRY;
578
579         if (rc)
580                 GOTO(out, rc);
581
582         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
583         if (body == NULL) {
584                 CERROR ("can't unpack ost_body\n");
585                 GOTO(out, rc = -EPROTO);
586         }
587
588         *fa->fa_oi->oi_oa = body->oa;
589 out:
590         rc = fa->fa_upcall(fa->fa_cookie, rc);
591         RETURN(rc);
592 }
593
594 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
595                   obd_enqueue_update_f upcall, void *cookie,
596                   struct ptlrpc_request_set *rqset)
597 {
598         struct ptlrpc_request *req;
599         struct ost_body       *body;
600         struct osc_fsync_args *fa;
601         int                    rc;
602         ENTRY;
603
604         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
605         if (req == NULL)
606                 RETURN(-ENOMEM);
607
608         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
609         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
610         if (rc) {
611                 ptlrpc_request_free(req);
612                 RETURN(rc);
613         }
614
615         /* overload the size and blocks fields in the oa with start/end */
616         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
617         LASSERT(body);
618         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
619                              oinfo->oi_oa);
620         osc_pack_capa(req, body, oinfo->oi_capa);
621
622         ptlrpc_request_set_replen(req);
623         req->rq_interpret_reply = osc_sync_interpret;
624
625         CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
626         fa = ptlrpc_req_async_args(req);
627         fa->fa_oi = oinfo;
628         fa->fa_upcall = upcall;
629         fa->fa_cookie = cookie;
630
631         if (rqset == PTLRPCD_SET)
632                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
633         else
634                 ptlrpc_set_add_req(rqset, req);
635
636         RETURN (0);
637 }
638
639 static int osc_sync(const struct lu_env *env, struct obd_export *exp,
640                     struct obd_info *oinfo, obd_size start, obd_size end,
641                     struct ptlrpc_request_set *set)
642 {
643         ENTRY;
644
645         if (!oinfo->oi_oa) {
646                 CDEBUG(D_INFO, "oa NULL\n");
647                 RETURN(-EINVAL);
648         }
649
650         oinfo->oi_oa->o_size = start;
651         oinfo->oi_oa->o_blocks = end;
652         oinfo->oi_oa->o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
653
654         RETURN(osc_sync_base(exp, oinfo, oinfo->oi_cb_up, oinfo, set));
655 }
656
657 /* Find and cancel locally locks matched by @mode in the resource found by
658  * @objid. Found locks are added into @cancel list. Returns the amount of
659  * locks added to @cancels list. */
660 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
661                                    cfs_list_t *cancels,
662                                    ldlm_mode_t mode, __u64 lock_flags)
663 {
664         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
665         struct ldlm_res_id res_id;
666         struct ldlm_resource *res;
667         int count;
668         ENTRY;
669
670         /* Return, i.e. cancel nothing, only if ELC is supported (flag in
671          * export) but disabled through procfs (flag in NS).
672          *
673          * This distinguishes from a case when ELC is not supported originally,
674          * when we still want to cancel locks in advance and just cancel them
675          * locally, without sending any RPC. */
676         if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
677                 RETURN(0);
678
679         ostid_build_res_name(&oa->o_oi, &res_id);
680         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
681         if (res == NULL)
682                 RETURN(0);
683
684         LDLM_RESOURCE_ADDREF(res);
685         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
686                                            lock_flags, 0, NULL);
687         LDLM_RESOURCE_DELREF(res);
688         ldlm_resource_putref(res);
689         RETURN(count);
690 }
691
692 static int osc_destroy_interpret(const struct lu_env *env,
693                                  struct ptlrpc_request *req, void *data,
694                                  int rc)
695 {
696         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
697
698         cfs_atomic_dec(&cli->cl_destroy_in_flight);
699         wake_up(&cli->cl_destroy_waitq);
700         return 0;
701 }
702
703 static int osc_can_send_destroy(struct client_obd *cli)
704 {
705         if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
706             cli->cl_max_rpcs_in_flight) {
707                 /* The destroy request can be sent */
708                 return 1;
709         }
710         if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
711             cli->cl_max_rpcs_in_flight) {
712                 /*
713                  * The counter has been modified between the two atomic
714                  * operations.
715                  */
716                 wake_up(&cli->cl_destroy_waitq);
717         }
718         return 0;
719 }
720
721 int osc_create(const struct lu_env *env, struct obd_export *exp,
722                struct obdo *oa, struct lov_stripe_md **ea,
723                struct obd_trans_info *oti)
724 {
725         int rc = 0;
726         ENTRY;
727
728         LASSERT(oa);
729         LASSERT(ea);
730         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
731
732         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
733             oa->o_flags == OBD_FL_RECREATE_OBJS) {
734                 RETURN(osc_real_create(exp, oa, ea, oti));
735         }
736
737         if (!fid_seq_is_mdt(ostid_seq(&oa->o_oi)))
738                 RETURN(osc_real_create(exp, oa, ea, oti));
739
740         /* we should not get here anymore */
741         LBUG();
742
743         RETURN(rc);
744 }
745
746 /* Destroy requests can be async always on the client, and we don't even really
747  * care about the return code since the client cannot do anything at all about
748  * a destroy failure.
749  * When the MDS is unlinking a filename, it saves the file objects into a
750  * recovery llog, and these object records are cancelled when the OST reports
751  * they were destroyed and sync'd to disk (i.e. transaction committed).
752  * If the client dies, or the OST is down when the object should be destroyed,
753  * the records are not cancelled, and when the OST reconnects to the MDS next,
754  * it will retrieve the llog unlink logs and then sends the log cancellation
755  * cookies to the MDS after committing destroy transactions. */
756 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
757                        struct obdo *oa, struct lov_stripe_md *ea,
758                        struct obd_trans_info *oti, struct obd_export *md_export,
759                        void *capa)
760 {
761         struct client_obd     *cli = &exp->exp_obd->u.cli;
762         struct ptlrpc_request *req;
763         struct ost_body       *body;
764         CFS_LIST_HEAD(cancels);
765         int rc, count;
766         ENTRY;
767
768         if (!oa) {
769                 CDEBUG(D_INFO, "oa NULL\n");
770                 RETURN(-EINVAL);
771         }
772
773         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
774                                         LDLM_FL_DISCARD_DATA);
775
776         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
777         if (req == NULL) {
778                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
779                 RETURN(-ENOMEM);
780         }
781
782         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
783         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
784                                0, &cancels, count);
785         if (rc) {
786                 ptlrpc_request_free(req);
787                 RETURN(rc);
788         }
789
790         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
791         ptlrpc_at_set_req_timeout(req);
792
793         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
794                 oa->o_lcookie = *oti->oti_logcookies;
795         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
796         LASSERT(body);
797         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
798
799         osc_pack_capa(req, body, (struct obd_capa *)capa);
800         ptlrpc_request_set_replen(req);
801
802         /* If osc_destory is for destroying the unlink orphan,
803          * sent from MDT to OST, which should not be blocked here,
804          * because the process might be triggered by ptlrpcd, and
805          * it is not good to block ptlrpcd thread (b=16006)*/
806         if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
807                 req->rq_interpret_reply = osc_destroy_interpret;
808                 if (!osc_can_send_destroy(cli)) {
809                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
810                                                           NULL);
811
812                         /*
813                          * Wait until the number of on-going destroy RPCs drops
814                          * under max_rpc_in_flight
815                          */
816                         l_wait_event_exclusive(cli->cl_destroy_waitq,
817                                                osc_can_send_destroy(cli), &lwi);
818                 }
819         }
820
821         /* Do not wait for response */
822         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
823         RETURN(0);
824 }
825
826 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
827                                 long writing_bytes)
828 {
829         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
830
831         LASSERT(!(oa->o_valid & bits));
832
833         oa->o_valid |= bits;
834         client_obd_list_lock(&cli->cl_loi_list_lock);
835         oa->o_dirty = cli->cl_dirty;
836         if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
837                      cli->cl_dirty_max)) {
838                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
839                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
840                 oa->o_undirty = 0;
841         } else if (unlikely(cfs_atomic_read(&obd_unstable_pages) +
842                             cfs_atomic_read(&obd_dirty_pages) -
843                             cfs_atomic_read(&obd_dirty_transit_pages) >
844                             (long)(obd_max_dirty_pages + 1))) {
845                 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
846                  * not covered by a lock thus they may safely race and trip
847                  * this CERROR() unless we add in a small fudge factor (+1). */
848                 CERROR("%s: dirty %d + %d - %d > system dirty_max %d\n",
849                        cli->cl_import->imp_obd->obd_name,
850                        cfs_atomic_read(&obd_unstable_pages),
851                        cfs_atomic_read(&obd_dirty_pages),
852                        cfs_atomic_read(&obd_dirty_transit_pages),
853                        obd_max_dirty_pages);
854                 oa->o_undirty = 0;
855         } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
856                 CERROR("dirty %lu - dirty_max %lu too big???\n",
857                        cli->cl_dirty, cli->cl_dirty_max);
858                 oa->o_undirty = 0;
859         } else {
860                 long max_in_flight = (cli->cl_max_pages_per_rpc <<
861                                       PAGE_CACHE_SHIFT) *
862                                      (cli->cl_max_rpcs_in_flight + 1);
863                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
864         }
865         oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
866         oa->o_dropped = cli->cl_lost_grant;
867         cli->cl_lost_grant = 0;
868         client_obd_list_unlock(&cli->cl_loi_list_lock);
869         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
870                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
871
872 }
873
874 void osc_update_next_shrink(struct client_obd *cli)
875 {
876         cli->cl_next_shrink_grant =
877                 cfs_time_shift(cli->cl_grant_shrink_interval);
878         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
879                cli->cl_next_shrink_grant);
880 }
881
882 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
883 {
884         client_obd_list_lock(&cli->cl_loi_list_lock);
885         cli->cl_avail_grant += grant;
886         client_obd_list_unlock(&cli->cl_loi_list_lock);
887 }
888
889 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
890 {
891         if (body->oa.o_valid & OBD_MD_FLGRANT) {
892                 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
893                 __osc_update_grant(cli, body->oa.o_grant);
894         }
895 }
896
897 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
898                               obd_count keylen, void *key, obd_count vallen,
899                               void *val, struct ptlrpc_request_set *set);
900
901 static int osc_shrink_grant_interpret(const struct lu_env *env,
902                                       struct ptlrpc_request *req,
903                                       void *aa, int rc)
904 {
905         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
906         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
907         struct ost_body *body;
908
909         if (rc != 0) {
910                 __osc_update_grant(cli, oa->o_grant);
911                 GOTO(out, rc);
912         }
913
914         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
915         LASSERT(body);
916         osc_update_grant(cli, body);
917 out:
918         OBDO_FREE(oa);
919         return rc;
920 }
921
922 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
923 {
924         client_obd_list_lock(&cli->cl_loi_list_lock);
925         oa->o_grant = cli->cl_avail_grant / 4;
926         cli->cl_avail_grant -= oa->o_grant;
927         client_obd_list_unlock(&cli->cl_loi_list_lock);
928         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
929                 oa->o_valid |= OBD_MD_FLFLAGS;
930                 oa->o_flags = 0;
931         }
932         oa->o_flags |= OBD_FL_SHRINK_GRANT;
933         osc_update_next_shrink(cli);
934 }
935
936 /* Shrink the current grant, either from some large amount to enough for a
937  * full set of in-flight RPCs, or if we have already shrunk to that limit
938  * then to enough for a single RPC.  This avoids keeping more grant than
939  * needed, and avoids shrinking the grant piecemeal. */
940 static int osc_shrink_grant(struct client_obd *cli)
941 {
942         __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
943                              (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
944
945         client_obd_list_lock(&cli->cl_loi_list_lock);
946         if (cli->cl_avail_grant <= target_bytes)
947                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
948         client_obd_list_unlock(&cli->cl_loi_list_lock);
949
950         return osc_shrink_grant_to_target(cli, target_bytes);
951 }
952
953 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
954 {
955         int                     rc = 0;
956         struct ost_body        *body;
957         ENTRY;
958
959         client_obd_list_lock(&cli->cl_loi_list_lock);
960         /* Don't shrink if we are already above or below the desired limit
961          * We don't want to shrink below a single RPC, as that will negatively
962          * impact block allocation and long-term performance. */
963         if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
964                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
965
966         if (target_bytes >= cli->cl_avail_grant) {
967                 client_obd_list_unlock(&cli->cl_loi_list_lock);
968                 RETURN(0);
969         }
970         client_obd_list_unlock(&cli->cl_loi_list_lock);
971
972         OBD_ALLOC_PTR(body);
973         if (!body)
974                 RETURN(-ENOMEM);
975
976         osc_announce_cached(cli, &body->oa, 0);
977
978         client_obd_list_lock(&cli->cl_loi_list_lock);
979         body->oa.o_grant = cli->cl_avail_grant - target_bytes;
980         cli->cl_avail_grant = target_bytes;
981         client_obd_list_unlock(&cli->cl_loi_list_lock);
982         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
983                 body->oa.o_valid |= OBD_MD_FLFLAGS;
984                 body->oa.o_flags = 0;
985         }
986         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
987         osc_update_next_shrink(cli);
988
989         rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
990                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
991                                 sizeof(*body), body, NULL);
992         if (rc != 0)
993                 __osc_update_grant(cli, body->oa.o_grant);
994         OBD_FREE_PTR(body);
995         RETURN(rc);
996 }
997
998 static int osc_should_shrink_grant(struct client_obd *client)
999 {
1000         cfs_time_t time = cfs_time_current();
1001         cfs_time_t next_shrink = client->cl_next_shrink_grant;
1002
1003         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1004              OBD_CONNECT_GRANT_SHRINK) == 0)
1005                 return 0;
1006
1007         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1008                 /* Get the current RPC size directly, instead of going via:
1009                  * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
1010                  * Keep comment here so that it can be found by searching. */
1011                 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
1012
1013                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1014                     client->cl_avail_grant > brw_size)
1015                         return 1;
1016                 else
1017                         osc_update_next_shrink(client);
1018         }
1019         return 0;
1020 }
1021
1022 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1023 {
1024         struct client_obd *client;
1025
1026         cfs_list_for_each_entry(client, &item->ti_obd_list,
1027                                 cl_grant_shrink_list) {
1028                 if (osc_should_shrink_grant(client))
1029                         osc_shrink_grant(client);
1030         }
1031         return 0;
1032 }
1033
1034 static int osc_add_shrink_grant(struct client_obd *client)
1035 {
1036         int rc;
1037
1038         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1039                                        TIMEOUT_GRANT,
1040                                        osc_grant_shrink_grant_cb, NULL,
1041                                        &client->cl_grant_shrink_list);
1042         if (rc) {
1043                 CERROR("add grant client %s error %d\n",
1044                         client->cl_import->imp_obd->obd_name, rc);
1045                 return rc;
1046         }
1047         CDEBUG(D_CACHE, "add grant client %s \n",
1048                client->cl_import->imp_obd->obd_name);
1049         osc_update_next_shrink(client);
1050         return 0;
1051 }
1052
1053 static int osc_del_shrink_grant(struct client_obd *client)
1054 {
1055         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1056                                          TIMEOUT_GRANT);
1057 }
1058
1059 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1060 {
1061         /*
1062          * ocd_grant is the total grant amount we're expect to hold: if we've
1063          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1064          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1065          *
1066          * race is tolerable here: if we're evicted, but imp_state already
1067          * left EVICTED state, then cl_dirty must be 0 already.
1068          */
1069         client_obd_list_lock(&cli->cl_loi_list_lock);
1070         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1071                 cli->cl_avail_grant = ocd->ocd_grant;
1072         else
1073                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1074
1075         if (cli->cl_avail_grant < 0) {
1076                 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1077                       cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1078                       ocd->ocd_grant, cli->cl_dirty);
1079                 /* workaround for servers which do not have the patch from
1080                  * LU-2679 */
1081                 cli->cl_avail_grant = ocd->ocd_grant;
1082         }
1083
1084         /* determine the appropriate chunk size used by osc_extent. */
1085         cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
1086         client_obd_list_unlock(&cli->cl_loi_list_lock);
1087
1088         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
1089                 "chunk bits: %d.\n", cli->cl_import->imp_obd->obd_name,
1090                 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1091
1092         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1093             cfs_list_empty(&cli->cl_grant_shrink_list))
1094                 osc_add_shrink_grant(cli);
1095 }
1096
1097 /* We assume that the reason this OSC got a short read is because it read
1098  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1099  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1100  * this stripe never got written at or beyond this stripe offset yet. */
1101 static void handle_short_read(int nob_read, obd_count page_count,
1102                               struct brw_page **pga)
1103 {
1104         char *ptr;
1105         int i = 0;
1106
1107         /* skip bytes read OK */
1108         while (nob_read > 0) {
1109                 LASSERT (page_count > 0);
1110
1111                 if (pga[i]->count > nob_read) {
1112                         /* EOF inside this page */
1113                         ptr = kmap(pga[i]->pg) +
1114                                 (pga[i]->off & ~CFS_PAGE_MASK);
1115                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1116                         kunmap(pga[i]->pg);
1117                         page_count--;
1118                         i++;
1119                         break;
1120                 }
1121
1122                 nob_read -= pga[i]->count;
1123                 page_count--;
1124                 i++;
1125         }
1126
1127         /* zero remaining pages */
1128         while (page_count-- > 0) {
1129                 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1130                 memset(ptr, 0, pga[i]->count);
1131                 kunmap(pga[i]->pg);
1132                 i++;
1133         }
1134 }
1135
1136 static int check_write_rcs(struct ptlrpc_request *req,
1137                            int requested_nob, int niocount,
1138                            obd_count page_count, struct brw_page **pga)
1139 {
1140         int     i;
1141         __u32   *remote_rcs;
1142
1143         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1144                                                   sizeof(*remote_rcs) *
1145                                                   niocount);
1146         if (remote_rcs == NULL) {
1147                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1148                 return(-EPROTO);
1149         }
1150
1151         /* return error if any niobuf was in error */
1152         for (i = 0; i < niocount; i++) {
1153                 if ((int)remote_rcs[i] < 0)
1154                         return(remote_rcs[i]);
1155
1156                 if (remote_rcs[i] != 0) {
1157                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1158                                 i, remote_rcs[i], req);
1159                         return(-EPROTO);
1160                 }
1161         }
1162
1163         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1164                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1165                        req->rq_bulk->bd_nob_transferred, requested_nob);
1166                 return(-EPROTO);
1167         }
1168
1169         return (0);
1170 }
1171
1172 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1173 {
1174         if (p1->flag != p2->flag) {
1175                 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1176                                   OBD_BRW_SYNC       | OBD_BRW_ASYNC   |
1177                                   OBD_BRW_NOQUOTA    | OBD_BRW_SOFT_SYNC);
1178
1179                 /* warn if we try to combine flags that we don't know to be
1180                  * safe to combine */
1181                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1182                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1183                               "report this at http://bugs.whamcloud.com/\n",
1184                               p1->flag, p2->flag);
1185                 }
1186                 return 0;
1187         }
1188
1189         return (p1->off + p1->count == p2->off);
1190 }
1191
1192 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1193                                    struct brw_page **pga, int opc,
1194                                    cksum_type_t cksum_type)
1195 {
1196         __u32                           cksum;
1197         int                             i = 0;
1198         struct cfs_crypto_hash_desc     *hdesc;
1199         unsigned int                    bufsize;
1200         int                             err;
1201         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1202
1203         LASSERT(pg_count > 0);
1204
1205         hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1206         if (IS_ERR(hdesc)) {
1207                 CERROR("Unable to initialize checksum hash %s\n",
1208                        cfs_crypto_hash_name(cfs_alg));
1209                 return PTR_ERR(hdesc);
1210         }
1211
1212         while (nob > 0 && pg_count > 0) {
1213                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1214
1215                 /* corrupt the data before we compute the checksum, to
1216                  * simulate an OST->client data error */
1217                 if (i == 0 && opc == OST_READ &&
1218                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1219                         unsigned char *ptr = kmap(pga[i]->pg);
1220                         int off = pga[i]->off & ~CFS_PAGE_MASK;
1221                         memcpy(ptr + off, "bad1", min(4, nob));
1222                         kunmap(pga[i]->pg);
1223                 }
1224                 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1225                                   pga[i]->off & ~CFS_PAGE_MASK,
1226                                   count);
1227                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1228                                (int)(pga[i]->off & ~CFS_PAGE_MASK));
1229
1230                 nob -= pga[i]->count;
1231                 pg_count--;
1232                 i++;
1233         }
1234
1235         bufsize = 4;
1236         err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1237
1238         if (err)
1239                 cfs_crypto_hash_final(hdesc, NULL, NULL);
1240
1241         /* For sending we only compute the wrong checksum instead
1242          * of corrupting the data so it is still correct on a redo */
1243         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1244                 cksum++;
1245
1246         return cksum;
1247 }
1248
1249 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1250                                 struct lov_stripe_md *lsm, obd_count page_count,
1251                                 struct brw_page **pga,
1252                                 struct ptlrpc_request **reqp,
1253                                 struct obd_capa *ocapa, int reserve,
1254                                 int resend)
1255 {
1256         struct ptlrpc_request   *req;
1257         struct ptlrpc_bulk_desc *desc;
1258         struct ost_body         *body;
1259         struct obd_ioobj        *ioobj;
1260         struct niobuf_remote    *niobuf;
1261         int niocount, i, requested_nob, opc, rc;
1262         struct osc_brw_async_args *aa;
1263         struct req_capsule      *pill;
1264         struct brw_page *pg_prev;
1265
1266         ENTRY;
1267         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1268                 RETURN(-ENOMEM); /* Recoverable */
1269         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1270                 RETURN(-EINVAL); /* Fatal */
1271
1272         if ((cmd & OBD_BRW_WRITE) != 0) {
1273                 opc = OST_WRITE;
1274                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1275                                                 cli->cl_import->imp_rq_pool,
1276                                                 &RQF_OST_BRW_WRITE);
1277         } else {
1278                 opc = OST_READ;
1279                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1280         }
1281         if (req == NULL)
1282                 RETURN(-ENOMEM);
1283
1284         for (niocount = i = 1; i < page_count; i++) {
1285                 if (!can_merge_pages(pga[i - 1], pga[i]))
1286                         niocount++;
1287         }
1288
1289         pill = &req->rq_pill;
1290         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1291                              sizeof(*ioobj));
1292         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1293                              niocount * sizeof(*niobuf));
1294         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1295
1296         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1297         if (rc) {
1298                 ptlrpc_request_free(req);
1299                 RETURN(rc);
1300         }
1301         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1302         ptlrpc_at_set_req_timeout(req);
1303         /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1304          * retry logic */
1305         req->rq_no_retry_einprogress = 1;
1306
1307         desc = ptlrpc_prep_bulk_imp(req, page_count,
1308                 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1309                 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1310                 OST_BULK_PORTAL);
1311
1312         if (desc == NULL)
1313                 GOTO(out, rc = -ENOMEM);
1314         /* NB request now owns desc and will free it when it gets freed */
1315
1316         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1317         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1318         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1319         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1320
1321         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1322
1323         obdo_to_ioobj(oa, ioobj);
1324         ioobj->ioo_bufcnt = niocount;
1325         /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1326          * that might be send for this request.  The actual number is decided
1327          * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1328          * "max - 1" for old client compatibility sending "0", and also so the
1329          * the actual maximum is a power-of-two number, not one less. LU-1431 */
1330         ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1331         osc_pack_capa(req, body, ocapa);
1332         LASSERT(page_count > 0);
1333         pg_prev = pga[0];
1334         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1335                 struct brw_page *pg = pga[i];
1336                 int poff = pg->off & ~CFS_PAGE_MASK;
1337
1338                 LASSERT(pg->count > 0);
1339                 /* make sure there is no gap in the middle of page array */
1340                 LASSERTF(page_count == 1 ||
1341                          (ergo(i == 0, poff + pg->count == PAGE_CACHE_SIZE) &&
1342                           ergo(i > 0 && i < page_count - 1,
1343                                poff == 0 && pg->count == PAGE_CACHE_SIZE)   &&
1344                           ergo(i == page_count - 1, poff == 0)),
1345                          "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1346                          i, page_count, pg, pg->off, pg->count);
1347 #ifdef __linux__
1348                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1349                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1350                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1351                          i, page_count,
1352                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1353                          pg_prev->pg, page_private(pg_prev->pg),
1354                          pg_prev->pg->index, pg_prev->off);
1355 #else
1356                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1357                          "i %d p_c %u\n", i, page_count);
1358 #endif
1359                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1360                         (pg->flag & OBD_BRW_SRVLOCK));
1361
1362                 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1363                 requested_nob += pg->count;
1364
1365                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1366                         niobuf--;
1367                         niobuf->len += pg->count;
1368                 } else {
1369                         niobuf->offset = pg->off;
1370                         niobuf->len    = pg->count;
1371                         niobuf->flags  = pg->flag;
1372                 }
1373                 pg_prev = pg;
1374         }
1375
1376         LASSERTF((void *)(niobuf - niocount) ==
1377                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1378                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1379                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1380
1381         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1382         if (resend) {
1383                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1384                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1385                         body->oa.o_flags = 0;
1386                 }
1387                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1388         }
1389
1390         if (osc_should_shrink_grant(cli))
1391                 osc_shrink_grant_local(cli, &body->oa);
1392
1393         /* size[REQ_REC_OFF] still sizeof (*body) */
1394         if (opc == OST_WRITE) {
1395                 if (cli->cl_checksum &&
1396                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1397                         /* store cl_cksum_type in a local variable since
1398                          * it can be changed via lprocfs */
1399                         cksum_type_t cksum_type = cli->cl_cksum_type;
1400
1401                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1402                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1403                                 body->oa.o_flags = 0;
1404                         }
1405                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1406                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1407                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1408                                                              page_count, pga,
1409                                                              OST_WRITE,
1410                                                              cksum_type);
1411                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1412                                body->oa.o_cksum);
1413                         /* save this in 'oa', too, for later checking */
1414                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1415                         oa->o_flags |= cksum_type_pack(cksum_type);
1416                 } else {
1417                         /* clear out the checksum flag, in case this is a
1418                          * resend but cl_checksum is no longer set. b=11238 */
1419                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1420                 }
1421                 oa->o_cksum = body->oa.o_cksum;
1422                 /* 1 RC per niobuf */
1423                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1424                                      sizeof(__u32) * niocount);
1425         } else {
1426                 if (cli->cl_checksum &&
1427                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1428                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1429                                 body->oa.o_flags = 0;
1430                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1431                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1432                 }
1433         }
1434         ptlrpc_request_set_replen(req);
1435
1436         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1437         aa = ptlrpc_req_async_args(req);
1438         aa->aa_oa = oa;
1439         aa->aa_requested_nob = requested_nob;
1440         aa->aa_nio_count = niocount;
1441         aa->aa_page_count = page_count;
1442         aa->aa_resends = 0;
1443         aa->aa_ppga = pga;
1444         aa->aa_cli = cli;
1445         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1446         if (ocapa && reserve)
1447                 aa->aa_ocapa = capa_get(ocapa);
1448
1449         *reqp = req;
1450         RETURN(0);
1451
1452  out:
1453         ptlrpc_req_finished(req);
1454         RETURN(rc);
1455 }
1456
1457 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1458                                 __u32 client_cksum, __u32 server_cksum, int nob,
1459                                 obd_count page_count, struct brw_page **pga,
1460                                 cksum_type_t client_cksum_type)
1461 {
1462         __u32 new_cksum;
1463         char *msg;
1464         cksum_type_t cksum_type;
1465
1466         if (server_cksum == client_cksum) {
1467                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1468                 return 0;
1469         }
1470
1471         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1472                                        oa->o_flags : 0);
1473         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1474                                       cksum_type);
1475
1476         if (cksum_type != client_cksum_type)
1477                 msg = "the server did not use the checksum type specified in "
1478                       "the original request - likely a protocol problem";
1479         else if (new_cksum == server_cksum)
1480                 msg = "changed on the client after we checksummed it - "
1481                       "likely false positive due to mmap IO (bug 11742)";
1482         else if (new_cksum == client_cksum)
1483                 msg = "changed in transit before arrival at OST";
1484         else
1485                 msg = "changed in transit AND doesn't match the original - "
1486                       "likely false positive due to mmap IO (bug 11742)";
1487
1488         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1489                            " object "DOSTID" extent ["LPU64"-"LPU64"]\n",
1490                            msg, libcfs_nid2str(peer->nid),
1491                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1492                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1493                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1494                            POSTID(&oa->o_oi), pga[0]->off,
1495                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1496         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1497                "client csum now %x\n", client_cksum, client_cksum_type,
1498                server_cksum, cksum_type, new_cksum);
1499         return 1;
1500 }
1501
1502 /* Note rc enters this function as number of bytes transferred */
1503 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1504 {
1505         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1506         const lnet_process_id_t *peer =
1507                         &req->rq_import->imp_connection->c_peer;
1508         struct client_obd *cli = aa->aa_cli;
1509         struct ost_body *body;
1510         __u32 client_cksum = 0;
1511         ENTRY;
1512
1513         if (rc < 0 && rc != -EDQUOT) {
1514                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1515                 RETURN(rc);
1516         }
1517
1518         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1519         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1520         if (body == NULL) {
1521                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1522                 RETURN(-EPROTO);
1523         }
1524
1525         /* set/clear over quota flag for a uid/gid */
1526         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1527             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1528                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1529
1530                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1531                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1532                        body->oa.o_flags);
1533                 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1534         }
1535
1536         osc_update_grant(cli, body);
1537
1538         if (rc < 0)
1539                 RETURN(rc);
1540
1541         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1542                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1543
1544         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1545                 if (rc > 0) {
1546                         CERROR("Unexpected +ve rc %d\n", rc);
1547                         RETURN(-EPROTO);
1548                 }
1549                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1550
1551                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1552                         RETURN(-EAGAIN);
1553
1554                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1555                     check_write_checksum(&body->oa, peer, client_cksum,
1556                                          body->oa.o_cksum, aa->aa_requested_nob,
1557                                          aa->aa_page_count, aa->aa_ppga,
1558                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1559                         RETURN(-EAGAIN);
1560
1561                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1562                                      aa->aa_page_count, aa->aa_ppga);
1563                 GOTO(out, rc);
1564         }
1565
1566         /* The rest of this function executes only for OST_READs */
1567
1568         /* if unwrap_bulk failed, return -EAGAIN to retry */
1569         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1570         if (rc < 0)
1571                 GOTO(out, rc = -EAGAIN);
1572
1573         if (rc > aa->aa_requested_nob) {
1574                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1575                        aa->aa_requested_nob);
1576                 RETURN(-EPROTO);
1577         }
1578
1579         if (rc != req->rq_bulk->bd_nob_transferred) {
1580                 CERROR ("Unexpected rc %d (%d transferred)\n",
1581                         rc, req->rq_bulk->bd_nob_transferred);
1582                 return (-EPROTO);
1583         }
1584
1585         if (rc < aa->aa_requested_nob)
1586                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1587
1588         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1589                 static int cksum_counter;
1590                 __u32      server_cksum = body->oa.o_cksum;
1591                 char      *via;
1592                 char      *router;
1593                 cksum_type_t cksum_type;
1594
1595                 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1596                                                body->oa.o_flags : 0);
1597                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1598                                                  aa->aa_ppga, OST_READ,
1599                                                  cksum_type);
1600
1601                 if (peer->nid == req->rq_bulk->bd_sender) {
1602                         via = router = "";
1603                 } else {
1604                         via = " via ";
1605                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1606                 }
1607
1608                 if (server_cksum == ~0 && rc > 0) {
1609                         CERROR("Protocol error: server %s set the 'checksum' "
1610                                "bit, but didn't send a checksum.  Not fatal, "
1611                                "but please notify on http://bugs.whamcloud.com/\n",
1612                                libcfs_nid2str(peer->nid));
1613                 } else if (server_cksum != client_cksum) {
1614                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1615                                            "%s%s%s inode "DFID" object "DOSTID
1616                                            " extent ["LPU64"-"LPU64"]\n",
1617                                            req->rq_import->imp_obd->obd_name,
1618                                            libcfs_nid2str(peer->nid),
1619                                            via, router,
1620                                            body->oa.o_valid & OBD_MD_FLFID ?
1621                                                 body->oa.o_parent_seq : (__u64)0,
1622                                            body->oa.o_valid & OBD_MD_FLFID ?
1623                                                 body->oa.o_parent_oid : 0,
1624                                            body->oa.o_valid & OBD_MD_FLFID ?
1625                                                 body->oa.o_parent_ver : 0,
1626                                            POSTID(&body->oa.o_oi),
1627                                            aa->aa_ppga[0]->off,
1628                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1629                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1630                                                                         1);
1631                         CERROR("client %x, server %x, cksum_type %x\n",
1632                                client_cksum, server_cksum, cksum_type);
1633                         cksum_counter = 0;
1634                         aa->aa_oa->o_cksum = client_cksum;
1635                         rc = -EAGAIN;
1636                 } else {
1637                         cksum_counter++;
1638                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1639                         rc = 0;
1640                 }
1641         } else if (unlikely(client_cksum)) {
1642                 static int cksum_missed;
1643
1644                 cksum_missed++;
1645                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1646                         CERROR("Checksum %u requested from %s but not sent\n",
1647                                cksum_missed, libcfs_nid2str(peer->nid));
1648         } else {
1649                 rc = 0;
1650         }
1651 out:
1652         if (rc >= 0)
1653                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1654                                      aa->aa_oa, &body->oa);
1655
1656         RETURN(rc);
1657 }
1658
1659 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1660                             struct lov_stripe_md *lsm,
1661                             obd_count page_count, struct brw_page **pga,
1662                             struct obd_capa *ocapa)
1663 {
1664         struct ptlrpc_request *req;
1665         int                    rc;
1666         wait_queue_head_t            waitq;
1667         int                    generation, resends = 0;
1668         struct l_wait_info     lwi;
1669
1670         ENTRY;
1671
1672         init_waitqueue_head(&waitq);
1673         generation = exp->exp_obd->u.cli.cl_import->imp_generation;
1674
1675 restart_bulk:
1676         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1677                                   page_count, pga, &req, ocapa, 0, resends);
1678         if (rc != 0)
1679                 return (rc);
1680
1681         if (resends) {
1682                 req->rq_generation_set = 1;
1683                 req->rq_import_generation = generation;
1684                 req->rq_sent = cfs_time_current_sec() + resends;
1685         }
1686
1687         rc = ptlrpc_queue_wait(req);
1688
1689         if (rc == -ETIMEDOUT && req->rq_resend) {
1690                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1691                 ptlrpc_req_finished(req);
1692                 goto restart_bulk;
1693         }
1694
1695         rc = osc_brw_fini_request(req, rc);
1696
1697         ptlrpc_req_finished(req);
1698         /* When server return -EINPROGRESS, client should always retry
1699          * regardless of the number of times the bulk was resent already.*/
1700         if (osc_recoverable_error(rc)) {
1701                 resends++;
1702                 if (rc != -EINPROGRESS &&
1703                     !client_should_resend(resends, &exp->exp_obd->u.cli)) {
1704                         CERROR("%s: too many resend retries for object: "
1705                                ""DOSTID", rc = %d.\n", exp->exp_obd->obd_name,
1706                                POSTID(&oa->o_oi), rc);
1707                         goto out;
1708                 }
1709                 if (generation !=
1710                     exp->exp_obd->u.cli.cl_import->imp_generation) {
1711                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
1712                                ""DOSTID", rc = %d.\n", exp->exp_obd->obd_name,
1713                                POSTID(&oa->o_oi), rc);
1714                         goto out;
1715                 }
1716
1717                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1718                                        NULL);
1719                 l_wait_event(waitq, 0, &lwi);
1720
1721                 goto restart_bulk;
1722         }
1723 out:
1724         if (rc == -EAGAIN || rc == -EINPROGRESS)
1725                 rc = -EIO;
1726         RETURN (rc);
1727 }
1728
1729 static int osc_brw_redo_request(struct ptlrpc_request *request,
1730                                 struct osc_brw_async_args *aa, int rc)
1731 {
1732         struct ptlrpc_request *new_req;
1733         struct osc_brw_async_args *new_aa;
1734         struct osc_async_page *oap;
1735         ENTRY;
1736
1737         DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1738                   "redo for recoverable error %d", rc);
1739
1740         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1741                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1742                                   aa->aa_cli, aa->aa_oa,
1743                                   NULL /* lsm unused by osc currently */,
1744                                   aa->aa_page_count, aa->aa_ppga,
1745                                   &new_req, aa->aa_ocapa, 0, 1);
1746         if (rc)
1747                 RETURN(rc);
1748
1749         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1750                 if (oap->oap_request != NULL) {
1751                         LASSERTF(request == oap->oap_request,
1752                                  "request %p != oap_request %p\n",
1753                                  request, oap->oap_request);
1754                         if (oap->oap_interrupted) {
1755                                 ptlrpc_req_finished(new_req);
1756                                 RETURN(-EINTR);
1757                         }
1758                 }
1759         }
1760         /* New request takes over pga and oaps from old request.
1761          * Note that copying a list_head doesn't work, need to move it... */
1762         aa->aa_resends++;
1763         new_req->rq_interpret_reply = request->rq_interpret_reply;
1764         new_req->rq_async_args = request->rq_async_args;
1765         new_req->rq_commit_cb = request->rq_commit_cb;
1766         /* cap resend delay to the current request timeout, this is similar to
1767          * what ptlrpc does (see after_reply()) */
1768         if (aa->aa_resends > new_req->rq_timeout)
1769                 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1770         else
1771                 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1772         new_req->rq_generation_set = 1;
1773         new_req->rq_import_generation = request->rq_import_generation;
1774
1775         new_aa = ptlrpc_req_async_args(new_req);
1776
1777         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1778         cfs_list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1779         CFS_INIT_LIST_HEAD(&new_aa->aa_exts);
1780         cfs_list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1781         new_aa->aa_resends = aa->aa_resends;
1782
1783         cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1784                 if (oap->oap_request) {
1785                         ptlrpc_req_finished(oap->oap_request);
1786                         oap->oap_request = ptlrpc_request_addref(new_req);
1787                 }
1788         }
1789
1790         new_aa->aa_ocapa = aa->aa_ocapa;
1791         aa->aa_ocapa = NULL;
1792
1793         /* XXX: This code will run into problem if we're going to support
1794          * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1795          * and wait for all of them to be finished. We should inherit request
1796          * set from old request. */
1797         ptlrpcd_add_req(new_req, PDL_POLICY_SAME, -1);
1798
1799         DEBUG_REQ(D_INFO, new_req, "new request");
1800         RETURN(0);
1801 }
1802
1803 /*
1804  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1805  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1806  * fine for our small page arrays and doesn't require allocation.  its an
1807  * insertion sort that swaps elements that are strides apart, shrinking the
1808  * stride down until its '1' and the array is sorted.
1809  */
1810 static void sort_brw_pages(struct brw_page **array, int num)
1811 {
1812         int stride, i, j;
1813         struct brw_page *tmp;
1814
1815         if (num == 1)
1816                 return;
1817         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1818                 ;
1819
1820         do {
1821                 stride /= 3;
1822                 for (i = stride ; i < num ; i++) {
1823                         tmp = array[i];
1824                         j = i;
1825                         while (j >= stride && array[j - stride]->off > tmp->off) {
1826                                 array[j] = array[j - stride];
1827                                 j -= stride;
1828                         }
1829                         array[j] = tmp;
1830                 }
1831         } while (stride > 1);
1832 }
1833
1834 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1835 {
1836         int count = 1;
1837         int offset;
1838         int i = 0;
1839
1840         LASSERT (pages > 0);
1841         offset = pg[i]->off & ~CFS_PAGE_MASK;
1842
1843         for (;;) {
1844                 pages--;
1845                 if (pages == 0)         /* that's all */
1846                         return count;
1847
1848                 if (offset + pg[i]->count < PAGE_CACHE_SIZE)
1849                         return count;   /* doesn't end on page boundary */
1850
1851                 i++;
1852                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1853                 if (offset != 0)        /* doesn't start on page boundary */
1854                         return count;
1855
1856                 count++;
1857         }
1858 }
1859
1860 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1861 {
1862         struct brw_page **ppga;
1863         int i;
1864
1865         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1866         if (ppga == NULL)
1867                 return NULL;
1868
1869         for (i = 0; i < count; i++)
1870                 ppga[i] = pga + i;
1871         return ppga;
1872 }
1873
1874 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1875 {
1876         LASSERT(ppga != NULL);
1877         OBD_FREE(ppga, sizeof(*ppga) * count);
1878 }
1879
1880 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1881                    obd_count page_count, struct brw_page *pga,
1882                    struct obd_trans_info *oti)
1883 {
1884         struct obdo *saved_oa = NULL;
1885         struct brw_page **ppga, **orig;
1886         struct obd_import *imp = class_exp2cliimp(exp);
1887         struct client_obd *cli;
1888         int rc, page_count_orig;
1889         ENTRY;
1890
1891         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1892         cli = &imp->imp_obd->u.cli;
1893
1894         if (cmd & OBD_BRW_CHECK) {
1895                 /* The caller just wants to know if there's a chance that this
1896                  * I/O can succeed */
1897
1898                 if (imp->imp_invalid)
1899                         RETURN(-EIO);
1900                 RETURN(0);
1901         }
1902
1903         /* test_brw with a failed create can trip this, maybe others. */
1904         LASSERT(cli->cl_max_pages_per_rpc);
1905
1906         rc = 0;
1907
1908         orig = ppga = osc_build_ppga(pga, page_count);
1909         if (ppga == NULL)
1910                 RETURN(-ENOMEM);
1911         page_count_orig = page_count;
1912
1913         sort_brw_pages(ppga, page_count);
1914         while (page_count) {
1915                 obd_count pages_per_brw;
1916
1917                 if (page_count > cli->cl_max_pages_per_rpc)
1918                         pages_per_brw = cli->cl_max_pages_per_rpc;
1919                 else
1920                         pages_per_brw = page_count;
1921
1922                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1923
1924                 if (saved_oa != NULL) {
1925                         /* restore previously saved oa */
1926                         *oinfo->oi_oa = *saved_oa;
1927                 } else if (page_count > pages_per_brw) {
1928                         /* save a copy of oa (brw will clobber it) */
1929                         OBDO_ALLOC(saved_oa);
1930                         if (saved_oa == NULL)
1931                                 GOTO(out, rc = -ENOMEM);
1932                         *saved_oa = *oinfo->oi_oa;
1933                 }
1934
1935                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1936                                       pages_per_brw, ppga, oinfo->oi_capa);
1937
1938                 if (rc != 0)
1939                         break;
1940
1941                 page_count -= pages_per_brw;
1942                 ppga += pages_per_brw;
1943         }
1944
1945 out:
1946         osc_release_ppga(orig, page_count_orig);
1947
1948         if (saved_oa != NULL)
1949                 OBDO_FREE(saved_oa);
1950
1951         RETURN(rc);
1952 }
1953
1954 static int brw_interpret(const struct lu_env *env,
1955                          struct ptlrpc_request *req, void *data, int rc)
1956 {
1957         struct osc_brw_async_args *aa = data;
1958         struct osc_extent *ext;
1959         struct osc_extent *tmp;
1960         struct client_obd *cli = aa->aa_cli;
1961         ENTRY;
1962
1963         rc = osc_brw_fini_request(req, rc);
1964         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1965         /* When server return -EINPROGRESS, client should always retry
1966          * regardless of the number of times the bulk was resent already. */
1967         if (osc_recoverable_error(rc)) {
1968                 if (req->rq_import_generation !=
1969                     req->rq_import->imp_generation) {
1970                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
1971                                ""DOSTID", rc = %d.\n",
1972                                req->rq_import->imp_obd->obd_name,
1973                                POSTID(&aa->aa_oa->o_oi), rc);
1974                 } else if (rc == -EINPROGRESS ||
1975                     client_should_resend(aa->aa_resends, aa->aa_cli)) {
1976                         rc = osc_brw_redo_request(req, aa, rc);
1977                 } else {
1978                         CERROR("%s: too many resent retries for object: "
1979                                ""LPU64":"LPU64", rc = %d.\n",
1980                                req->rq_import->imp_obd->obd_name,
1981                                POSTID(&aa->aa_oa->o_oi), rc);
1982                 }
1983
1984                 if (rc == 0)
1985                         RETURN(0);
1986                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1987                         rc = -EIO;
1988         }
1989
1990         if (aa->aa_ocapa) {
1991                 capa_put(aa->aa_ocapa);
1992                 aa->aa_ocapa = NULL;
1993         }
1994
1995         if (rc == 0) {
1996                 struct obdo *oa = aa->aa_oa;
1997                 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1998                 unsigned long valid = 0;
1999                 struct cl_object *obj;
2000                 struct osc_async_page *last;
2001
2002                 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2003                 obj = osc2cl(last->oap_obj);
2004
2005                 cl_object_attr_lock(obj);
2006                 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2007                         attr->cat_blocks = oa->o_blocks;
2008                         valid |= CAT_BLOCKS;
2009                 }
2010                 if (oa->o_valid & OBD_MD_FLMTIME) {
2011                         attr->cat_mtime = oa->o_mtime;
2012                         valid |= CAT_MTIME;
2013                 }
2014                 if (oa->o_valid & OBD_MD_FLATIME) {
2015                         attr->cat_atime = oa->o_atime;
2016                         valid |= CAT_ATIME;
2017                 }
2018                 if (oa->o_valid & OBD_MD_FLCTIME) {
2019                         attr->cat_ctime = oa->o_ctime;
2020                         valid |= CAT_CTIME;
2021                 }
2022
2023                 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2024                         struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2025                         loff_t last_off = last->oap_count + last->oap_obj_off;
2026
2027                         /* Change file size if this is an out of quota or
2028                          * direct IO write and it extends the file size */
2029                         if (loi->loi_lvb.lvb_size < last_off) {
2030                                 attr->cat_size = last_off;
2031                                 valid |= CAT_SIZE;
2032                         }
2033                         /* Extend KMS if it's not a lockless write */
2034                         if (loi->loi_kms < last_off &&
2035                             oap2osc_page(last)->ops_srvlock == 0) {
2036                                 attr->cat_kms = last_off;
2037                                 valid |= CAT_KMS;
2038                         }
2039                 }
2040
2041                 if (valid != 0)
2042                         cl_object_attr_set(env, obj, attr, valid);
2043                 cl_object_attr_unlock(obj);
2044         }
2045         OBDO_FREE(aa->aa_oa);
2046
2047         cfs_list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2048                 cfs_list_del_init(&ext->oe_link);
2049                 osc_extent_finish(env, ext, 1, rc);
2050         }
2051         LASSERT(cfs_list_empty(&aa->aa_exts));
2052         LASSERT(cfs_list_empty(&aa->aa_oaps));
2053
2054         cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2055                           req->rq_bulk->bd_nob_transferred);
2056         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2057         ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
2058
2059         client_obd_list_lock(&cli->cl_loi_list_lock);
2060         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2061          * is called so we know whether to go to sync BRWs or wait for more
2062          * RPCs to complete */
2063         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2064                 cli->cl_w_in_flight--;
2065         else
2066                 cli->cl_r_in_flight--;
2067         osc_wake_cache_waiters(cli);
2068         client_obd_list_unlock(&cli->cl_loi_list_lock);
2069
2070         osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
2071         RETURN(rc);
2072 }
2073
2074 static void brw_commit(struct ptlrpc_request *req)
2075 {
2076         spin_lock(&req->rq_lock);
2077         /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2078          * this called via the rq_commit_cb, I need to ensure
2079          * osc_dec_unstable_pages is still called. Otherwise unstable
2080          * pages may be leaked. */
2081         if (req->rq_unstable) {
2082                 spin_unlock(&req->rq_lock);
2083                 osc_dec_unstable_pages(req);
2084                 spin_lock(&req->rq_lock);
2085         } else {
2086                 req->rq_committed = 1;
2087         }
2088         spin_unlock(&req->rq_lock);
2089 }
2090
2091 /**
2092  * Build an RPC by the list of extent @ext_list. The caller must ensure
2093  * that the total pages in this list are NOT over max pages per RPC.
2094  * Extents in the list must be in OES_RPC state.
2095  */
2096 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2097                   cfs_list_t *ext_list, int cmd, pdl_policy_t pol)
2098 {
2099         struct ptlrpc_request           *req = NULL;
2100         struct osc_extent               *ext;
2101         struct brw_page                 **pga = NULL;
2102         struct osc_brw_async_args       *aa = NULL;
2103         struct obdo                     *oa = NULL;
2104         struct osc_async_page           *oap;
2105         struct osc_async_page           *tmp;
2106         struct cl_req                   *clerq = NULL;
2107         enum cl_req_type                crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE :
2108                                                                       CRT_READ;
2109         struct ldlm_lock                *lock = NULL;
2110         struct cl_req_attr              *crattr = NULL;
2111         obd_off                         starting_offset = OBD_OBJECT_EOF;
2112         obd_off                         ending_offset = 0;
2113         int                             mpflag = 0;
2114         int                             mem_tight = 0;
2115         int                             page_count = 0;
2116         int                             i;
2117         int                             rc;
2118         CFS_LIST_HEAD(rpc_list);
2119
2120         ENTRY;
2121         LASSERT(!cfs_list_empty(ext_list));
2122
2123         /* add pages into rpc_list to build BRW rpc */
2124         cfs_list_for_each_entry(ext, ext_list, oe_link) {
2125                 LASSERT(ext->oe_state == OES_RPC);
2126                 mem_tight |= ext->oe_memalloc;
2127                 cfs_list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2128                         ++page_count;
2129                         cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2130                         if (starting_offset > oap->oap_obj_off)
2131                                 starting_offset = oap->oap_obj_off;
2132                         else
2133                                 LASSERT(oap->oap_page_off == 0);
2134                         if (ending_offset < oap->oap_obj_off + oap->oap_count)
2135                                 ending_offset = oap->oap_obj_off +
2136                                                 oap->oap_count;
2137                         else
2138                                 LASSERT(oap->oap_page_off + oap->oap_count ==
2139                                         PAGE_CACHE_SIZE);
2140                 }
2141         }
2142
2143         if (mem_tight)
2144                 mpflag = cfs_memory_pressure_get_and_set();
2145
2146         OBD_ALLOC(crattr, sizeof(*crattr));
2147         if (crattr == NULL)
2148                 GOTO(out, rc = -ENOMEM);
2149
2150         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2151         if (pga == NULL)
2152                 GOTO(out, rc = -ENOMEM);
2153
2154         OBDO_ALLOC(oa);
2155         if (oa == NULL)
2156                 GOTO(out, rc = -ENOMEM);
2157
2158         i = 0;
2159         cfs_list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
2160                 struct cl_page *page = oap2cl_page(oap);
2161                 if (clerq == NULL) {
2162                         clerq = cl_req_alloc(env, page, crt,
2163                                              1 /* only 1-object rpcs for now */);
2164                         if (IS_ERR(clerq))
2165                                 GOTO(out, rc = PTR_ERR(clerq));
2166                         lock = oap->oap_ldlm_lock;
2167                 }
2168                 if (mem_tight)
2169                         oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2170                 pga[i] = &oap->oap_brw_page;
2171                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2172                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2173                        pga[i]->pg, page_index(oap->oap_page), oap,
2174                        pga[i]->flag);
2175                 i++;
2176                 cl_req_page_add(env, clerq, page);
2177         }
2178
2179         /* always get the data for the obdo for the rpc */
2180         LASSERT(clerq != NULL);
2181         crattr->cra_oa = oa;
2182         cl_req_attr_set(env, clerq, crattr, ~0ULL);
2183         if (lock) {
2184                 oa->o_handle = lock->l_remote_handle;
2185                 oa->o_valid |= OBD_MD_FLHANDLE;
2186         }
2187
2188         rc = cl_req_prep(env, clerq);
2189         if (rc != 0) {
2190                 CERROR("cl_req_prep failed: %d\n", rc);
2191                 GOTO(out, rc);
2192         }
2193
2194         sort_brw_pages(pga, page_count);
2195         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2196                         pga, &req, crattr->cra_capa, 1, 0);
2197         if (rc != 0) {
2198                 CERROR("prep_req failed: %d\n", rc);
2199                 GOTO(out, rc);
2200         }
2201
2202         req->rq_commit_cb = brw_commit;
2203         req->rq_interpret_reply = brw_interpret;
2204
2205         if (mem_tight != 0)
2206                 req->rq_memalloc = 1;
2207
2208         /* Need to update the timestamps after the request is built in case
2209          * we race with setattr (locally or in queue at OST).  If OST gets
2210          * later setattr before earlier BRW (as determined by the request xid),
2211          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2212          * way to do this in a single call.  bug 10150 */
2213         cl_req_attr_set(env, clerq, crattr,
2214                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2215
2216         lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2217
2218         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2219         aa = ptlrpc_req_async_args(req);
2220         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2221         cfs_list_splice_init(&rpc_list, &aa->aa_oaps);
2222         CFS_INIT_LIST_HEAD(&aa->aa_exts);
2223         cfs_list_splice_init(ext_list, &aa->aa_exts);
2224         aa->aa_clerq = clerq;
2225
2226         /* queued sync pages can be torn down while the pages
2227          * were between the pending list and the rpc */
2228         tmp = NULL;
2229         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2230                 /* only one oap gets a request reference */
2231                 if (tmp == NULL)
2232                         tmp = oap;
2233                 if (oap->oap_interrupted && !req->rq_intr) {
2234                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2235                                         oap, req);
2236                         ptlrpc_mark_interrupted(req);
2237                 }
2238         }
2239         if (tmp != NULL)
2240                 tmp->oap_request = ptlrpc_request_addref(req);
2241
2242         client_obd_list_lock(&cli->cl_loi_list_lock);
2243         starting_offset >>= PAGE_CACHE_SHIFT;
2244         if (cmd == OBD_BRW_READ) {
2245                 cli->cl_r_in_flight++;
2246                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2247                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2248                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2249                                       starting_offset + 1);
2250         } else {
2251                 cli->cl_w_in_flight++;
2252                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2253                 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2254                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2255                                       starting_offset + 1);
2256         }
2257         client_obd_list_unlock(&cli->cl_loi_list_lock);
2258
2259         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2260                   page_count, aa, cli->cl_r_in_flight,
2261                   cli->cl_w_in_flight);
2262
2263         /* XXX: Maybe the caller can check the RPC bulk descriptor to
2264          * see which CPU/NUMA node the majority of pages were allocated
2265          * on, and try to assign the async RPC to the CPU core
2266          * (PDL_POLICY_PREFERRED) to reduce cross-CPU memory traffic.
2267          *
2268          * But on the other hand, we expect that multiple ptlrpcd
2269          * threads and the initial write sponsor can run in parallel,
2270          * especially when data checksum is enabled, which is CPU-bound
2271          * operation and single ptlrpcd thread cannot process in time.
2272          * So more ptlrpcd threads sharing BRW load
2273          * (with PDL_POLICY_ROUND) seems better.
2274          */
2275         ptlrpcd_add_req(req, pol, -1);
2276         rc = 0;
2277         EXIT;
2278
2279 out:
2280         if (mem_tight != 0)
2281                 cfs_memory_pressure_restore(mpflag);
2282
2283         if (crattr != NULL) {
2284                 capa_put(crattr->cra_capa);
2285                 OBD_FREE(crattr, sizeof(*crattr));
2286         }
2287
2288         if (rc != 0) {
2289                 LASSERT(req == NULL);
2290
2291                 if (oa)
2292                         OBDO_FREE(oa);
2293                 if (pga)
2294                         OBD_FREE(pga, sizeof(*pga) * page_count);
2295                 /* this should happen rarely and is pretty bad, it makes the
2296                  * pending list not follow the dirty order */
2297                 while (!cfs_list_empty(ext_list)) {
2298                         ext = cfs_list_entry(ext_list->next, struct osc_extent,
2299                                              oe_link);
2300                         cfs_list_del_init(&ext->oe_link);
2301                         osc_extent_finish(env, ext, 0, rc);
2302                 }
2303                 if (clerq && !IS_ERR(clerq))
2304                         cl_req_completion(env, clerq, rc);
2305         }
2306         RETURN(rc);
2307 }
2308
2309 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2310                                         struct ldlm_enqueue_info *einfo)
2311 {
2312         void *data = einfo->ei_cbdata;
2313         int set = 0;
2314
2315         LASSERT(lock != NULL);
2316         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2317         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2318         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2319         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2320
2321         lock_res_and_lock(lock);
2322         spin_lock(&osc_ast_guard);
2323
2324         if (lock->l_ast_data == NULL)
2325                 lock->l_ast_data = data;
2326         if (lock->l_ast_data == data)
2327                 set = 1;
2328
2329         spin_unlock(&osc_ast_guard);
2330         unlock_res_and_lock(lock);
2331
2332         return set;
2333 }
2334
2335 static int osc_set_data_with_check(struct lustre_handle *lockh,
2336                                    struct ldlm_enqueue_info *einfo)
2337 {
2338         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2339         int set = 0;
2340
2341         if (lock != NULL) {
2342                 set = osc_set_lock_data_with_check(lock, einfo);
2343                 LDLM_LOCK_PUT(lock);
2344         } else
2345                 CERROR("lockh %p, data %p - client evicted?\n",
2346                        lockh, einfo->ei_cbdata);
2347         return set;
2348 }
2349
2350 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2351                              ldlm_iterator_t replace, void *data)
2352 {
2353         struct ldlm_res_id res_id;
2354         struct obd_device *obd = class_exp2obd(exp);
2355
2356         ostid_build_res_name(&lsm->lsm_oi, &res_id);
2357         ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2358         return 0;
2359 }
2360
2361 /* find any ldlm lock of the inode in osc
2362  * return 0    not find
2363  *        1    find one
2364  *      < 0    error */
2365 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2366                            ldlm_iterator_t replace, void *data)
2367 {
2368         struct ldlm_res_id res_id;
2369         struct obd_device *obd = class_exp2obd(exp);
2370         int rc = 0;
2371
2372         ostid_build_res_name(&lsm->lsm_oi, &res_id);
2373         rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2374         if (rc == LDLM_ITER_STOP)
2375                 return(1);
2376         if (rc == LDLM_ITER_CONTINUE)
2377                 return(0);
2378         return(rc);
2379 }
2380
2381 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2382                             obd_enqueue_update_f upcall, void *cookie,
2383                             __u64 *flags, int agl, int rc)
2384 {
2385         int intent = *flags & LDLM_FL_HAS_INTENT;
2386         ENTRY;
2387
2388         if (intent) {
2389                 /* The request was created before ldlm_cli_enqueue call. */
2390                 if (rc == ELDLM_LOCK_ABORTED) {
2391                         struct ldlm_reply *rep;
2392                         rep = req_capsule_server_get(&req->rq_pill,
2393                                                      &RMF_DLM_REP);
2394
2395                         LASSERT(rep != NULL);
2396                         rep->lock_policy_res1 =
2397                                 ptlrpc_status_ntoh(rep->lock_policy_res1);
2398                         if (rep->lock_policy_res1)
2399                                 rc = rep->lock_policy_res1;
2400                 }
2401         }
2402
2403         if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2404             (rc == 0)) {
2405                 *flags |= LDLM_FL_LVB_READY;
2406                 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
2407                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2408         }
2409
2410         /* Call the update callback. */
2411         rc = (*upcall)(cookie, rc);
2412         RETURN(rc);
2413 }
2414
2415 static int osc_enqueue_interpret(const struct lu_env *env,
2416                                  struct ptlrpc_request *req,
2417                                  struct osc_enqueue_args *aa, int rc)
2418 {
2419         struct ldlm_lock *lock;
2420         struct lustre_handle handle;
2421         __u32 mode;
2422         struct ost_lvb *lvb;
2423         __u32 lvb_len;
2424         __u64 *flags = aa->oa_flags;
2425
2426         /* Make a local copy of a lock handle and a mode, because aa->oa_*
2427          * might be freed anytime after lock upcall has been called. */
2428         lustre_handle_copy(&handle, aa->oa_lockh);
2429         mode = aa->oa_ei->ei_mode;
2430
2431         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2432          * be valid. */
2433         lock = ldlm_handle2lock(&handle);
2434
2435         /* Take an additional reference so that a blocking AST that
2436          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2437          * to arrive after an upcall has been executed by
2438          * osc_enqueue_fini(). */
2439         ldlm_lock_addref(&handle, mode);
2440
2441         /* Let CP AST to grant the lock first. */
2442         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2443
2444         if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2445                 lvb = NULL;
2446                 lvb_len = 0;
2447         } else {
2448                 lvb = aa->oa_lvb;
2449                 lvb_len = sizeof(*aa->oa_lvb);
2450         }
2451
2452         /* Complete obtaining the lock procedure. */
2453         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2454                                    mode, flags, lvb, lvb_len, &handle, rc);
2455         /* Complete osc stuff. */
2456         rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2457                               flags, aa->oa_agl, rc);
2458
2459         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2460
2461         /* Release the lock for async request. */
2462         if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2463                 /*
2464                  * Releases a reference taken by ldlm_cli_enqueue(), if it is
2465                  * not already released by
2466                  * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2467                  */
2468                 ldlm_lock_decref(&handle, mode);
2469
2470         LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
2471                  aa->oa_lockh, req, aa);
2472         ldlm_lock_decref(&handle, mode);
2473         LDLM_LOCK_PUT(lock);
2474         return rc;
2475 }
2476
2477 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
2478                         struct lov_oinfo *loi, __u64 flags,
2479                         struct ost_lvb *lvb, __u32 mode, int rc)
2480 {
2481         struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2482
2483         if (rc == ELDLM_OK) {
2484                 __u64 tmp;
2485
2486                 LASSERT(lock != NULL);
2487                 loi->loi_lvb = *lvb;
2488                 tmp = loi->loi_lvb.lvb_size;
2489                 /* Extend KMS up to the end of this lock and no further
2490                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2491                 if (tmp > lock->l_policy_data.l_extent.end)
2492                         tmp = lock->l_policy_data.l_extent.end + 1;
2493                 if (tmp >= loi->loi_kms) {
2494                         LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
2495                                    ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
2496                         loi_kms_set(loi, tmp);
2497                 } else {
2498                         LDLM_DEBUG(lock, "lock acquired, setting rss="
2499                                    LPU64"; leaving kms="LPU64", end="LPU64,
2500                                    loi->loi_lvb.lvb_size, loi->loi_kms,
2501                                    lock->l_policy_data.l_extent.end);
2502                 }
2503                 ldlm_lock_allow_match(lock);
2504         } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
2505                 LASSERT(lock != NULL);
2506                 loi->loi_lvb = *lvb;
2507                 ldlm_lock_allow_match(lock);
2508                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2509                        " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
2510                 rc = ELDLM_OK;
2511         }
2512
2513         if (lock != NULL) {
2514                 if (rc != ELDLM_OK)
2515                         ldlm_lock_fail_match(lock);
2516
2517                 LDLM_LOCK_PUT(lock);
2518         }
2519 }
2520 EXPORT_SYMBOL(osc_update_enqueue);
2521
2522 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2523
2524 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2525  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2526  * other synchronous requests, however keeping some locks and trying to obtain
2527  * others may take a considerable amount of time in a case of ost failure; and
2528  * when other sync requests do not get released lock from a client, the client
2529  * is excluded from the cluster -- such scenarious make the life difficult, so
2530  * release locks just after they are obtained. */
2531 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2532                      __u64 *flags, ldlm_policy_data_t *policy,
2533                      struct ost_lvb *lvb, int kms_valid,
2534                      obd_enqueue_update_f upcall, void *cookie,
2535                      struct ldlm_enqueue_info *einfo,
2536                      struct lustre_handle *lockh,
2537                      struct ptlrpc_request_set *rqset, int async, int agl)
2538 {
2539         struct obd_device *obd = exp->exp_obd;
2540         struct ptlrpc_request *req = NULL;
2541         int intent = *flags & LDLM_FL_HAS_INTENT;
2542         __u64 match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2543         ldlm_mode_t mode;
2544         int rc;
2545         ENTRY;
2546
2547         /* Filesystem lock extents are extended to page boundaries so that
2548          * dealing with the page cache is a little smoother.  */
2549         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2550         policy->l_extent.end |= ~CFS_PAGE_MASK;
2551
2552         /*
2553          * kms is not valid when either object is completely fresh (so that no
2554          * locks are cached), or object was evicted. In the latter case cached
2555          * lock cannot be used, because it would prime inode state with
2556          * potentially stale LVB.
2557          */
2558         if (!kms_valid)
2559                 goto no_match;
2560
2561         /* Next, search for already existing extent locks that will cover us */
2562         /* If we're trying to read, we also search for an existing PW lock.  The
2563          * VFS and page cache already protect us locally, so lots of readers/
2564          * writers can share a single PW lock.
2565          *
2566          * There are problems with conversion deadlocks, so instead of
2567          * converting a read lock to a write lock, we'll just enqueue a new
2568          * one.
2569          *
2570          * At some point we should cancel the read lock instead of making them
2571          * send us a blocking callback, but there are problems with canceling
2572          * locks out from other users right now, too. */
2573         mode = einfo->ei_mode;
2574         if (einfo->ei_mode == LCK_PR)
2575                 mode |= LCK_PW;
2576         mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2577                                einfo->ei_type, policy, mode, lockh, 0);
2578         if (mode) {
2579                 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2580
2581                 if ((agl != 0) && !ldlm_is_lvb_ready(matched)) {
2582                         /* For AGL, if enqueue RPC is sent but the lock is not
2583                          * granted, then skip to process this strpe.
2584                          * Return -ECANCELED to tell the caller. */
2585                         ldlm_lock_decref(lockh, mode);
2586                         LDLM_LOCK_PUT(matched);
2587                         RETURN(-ECANCELED);
2588                 } else if (osc_set_lock_data_with_check(matched, einfo)) {
2589                         *flags |= LDLM_FL_LVB_READY;
2590                         /* addref the lock only if not async requests and PW
2591                          * lock is matched whereas we asked for PR. */
2592                         if (!rqset && einfo->ei_mode != mode)
2593                                 ldlm_lock_addref(lockh, LCK_PR);
2594                         if (intent) {
2595                                 /* I would like to be able to ASSERT here that
2596                                  * rss <= kms, but I can't, for reasons which
2597                                  * are explained in lov_enqueue() */
2598                         }
2599
2600                         /* We already have a lock, and it's referenced.
2601                          *
2602                          * At this point, the cl_lock::cll_state is CLS_QUEUING,
2603                          * AGL upcall may change it to CLS_HELD directly. */
2604                         (*upcall)(cookie, ELDLM_OK);
2605
2606                         if (einfo->ei_mode != mode)
2607                                 ldlm_lock_decref(lockh, LCK_PW);
2608                         else if (rqset)
2609                                 /* For async requests, decref the lock. */
2610                                 ldlm_lock_decref(lockh, einfo->ei_mode);
2611                         LDLM_LOCK_PUT(matched);
2612                         RETURN(ELDLM_OK);
2613                 } else {
2614                         ldlm_lock_decref(lockh, mode);
2615                         LDLM_LOCK_PUT(matched);
2616                 }
2617         }
2618
2619  no_match:
2620         if (intent) {
2621                 CFS_LIST_HEAD(cancels);
2622                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2623                                            &RQF_LDLM_ENQUEUE_LVB);
2624                 if (req == NULL)
2625                         RETURN(-ENOMEM);
2626
2627                 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
2628                 if (rc) {
2629                         ptlrpc_request_free(req);
2630                         RETURN(rc);
2631                 }
2632
2633                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2634                                      sizeof *lvb);
2635                 ptlrpc_request_set_replen(req);
2636         }
2637
2638         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2639         *flags &= ~LDLM_FL_BLOCK_GRANTED;
2640
2641         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2642                               sizeof(*lvb), LVB_T_OST, lockh, async);
2643         if (rqset) {
2644                 if (!rc) {
2645                         struct osc_enqueue_args *aa;
2646                         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2647                         aa = ptlrpc_req_async_args(req);
2648                         aa->oa_ei = einfo;
2649                         aa->oa_exp = exp;
2650                         aa->oa_flags  = flags;
2651                         aa->oa_upcall = upcall;
2652                         aa->oa_cookie = cookie;
2653                         aa->oa_lvb    = lvb;
2654                         aa->oa_lockh  = lockh;
2655                         aa->oa_agl    = !!agl;
2656
2657                         req->rq_interpret_reply =
2658                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2659                         if (rqset == PTLRPCD_SET)
2660                                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2661                         else
2662                                 ptlrpc_set_add_req(rqset, req);
2663                 } else if (intent) {
2664                         ptlrpc_req_finished(req);
2665                 }
2666                 RETURN(rc);
2667         }
2668
2669         rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2670         if (intent)
2671                 ptlrpc_req_finished(req);
2672
2673         RETURN(rc);
2674 }
2675
2676 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
2677                        struct ldlm_enqueue_info *einfo,
2678                        struct ptlrpc_request_set *rqset)
2679 {
2680         struct ldlm_res_id res_id;
2681         int rc;
2682         ENTRY;
2683
2684         ostid_build_res_name(&oinfo->oi_md->lsm_oi, &res_id);
2685         rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
2686                               &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
2687                               oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
2688                               oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
2689                               rqset, rqset != NULL, 0);
2690         RETURN(rc);
2691 }
2692
2693 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2694                    __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2695                    __u64 *flags, void *data, struct lustre_handle *lockh,
2696                    int unref)
2697 {
2698         struct obd_device *obd = exp->exp_obd;
2699         __u64 lflags = *flags;
2700         ldlm_mode_t rc;
2701         ENTRY;
2702
2703         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2704                 RETURN(-EIO);
2705
2706         /* Filesystem lock extents are extended to page boundaries so that
2707          * dealing with the page cache is a little smoother */
2708         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2709         policy->l_extent.end |= ~CFS_PAGE_MASK;
2710
2711         /* Next, search for already existing extent locks that will cover us */
2712         /* If we're trying to read, we also search for an existing PW lock.  The
2713          * VFS and page cache already protect us locally, so lots of readers/
2714          * writers can share a single PW lock. */
2715         rc = mode;
2716         if (mode == LCK_PR)
2717                 rc |= LCK_PW;
2718         rc = ldlm_lock_match(obd->obd_namespace, lflags,
2719                              res_id, type, policy, rc, lockh, unref);
2720         if (rc) {
2721                 if (data != NULL) {
2722                         if (!osc_set_data_with_check(lockh, data)) {
2723                                 if (!(lflags & LDLM_FL_TEST_LOCK))
2724                                         ldlm_lock_decref(lockh, rc);
2725                                 RETURN(0);
2726                         }
2727                 }
2728                 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2729                         ldlm_lock_addref(lockh, LCK_PR);
2730                         ldlm_lock_decref(lockh, LCK_PW);
2731                 }
2732                 RETURN(rc);
2733         }
2734         RETURN(rc);
2735 }
2736
2737 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2738 {
2739         ENTRY;
2740
2741         if (unlikely(mode == LCK_GROUP))
2742                 ldlm_lock_decref_and_cancel(lockh, mode);
2743         else
2744                 ldlm_lock_decref(lockh, mode);
2745
2746         RETURN(0);
2747 }
2748
2749 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2750                       __u32 mode, struct lustre_handle *lockh)
2751 {
2752         ENTRY;
2753         RETURN(osc_cancel_base(lockh, mode));
2754 }
2755
2756 static int osc_cancel_unused(struct obd_export *exp,
2757                              struct lov_stripe_md *lsm,
2758                              ldlm_cancel_flags_t flags,
2759                              void *opaque)
2760 {
2761         struct obd_device *obd = class_exp2obd(exp);
2762         struct ldlm_res_id res_id, *resp = NULL;
2763
2764         if (lsm != NULL) {
2765                 ostid_build_res_name(&lsm->lsm_oi, &res_id);
2766                 resp = &res_id;
2767         }
2768
2769         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2770 }
2771
2772 static int osc_statfs_interpret(const struct lu_env *env,
2773                                 struct ptlrpc_request *req,
2774                                 struct osc_async_args *aa, int rc)
2775 {
2776         struct obd_statfs *msfs;
2777         ENTRY;
2778
2779         if (rc == -EBADR)
2780                 /* The request has in fact never been sent
2781                  * due to issues at a higher level (LOV).
2782                  * Exit immediately since the caller is
2783                  * aware of the problem and takes care
2784                  * of the clean up */
2785                  RETURN(rc);
2786
2787         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2788             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2789                 GOTO(out, rc = 0);
2790
2791         if (rc != 0)
2792                 GOTO(out, rc);
2793
2794         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2795         if (msfs == NULL) {
2796                 GOTO(out, rc = -EPROTO);
2797         }
2798
2799         *aa->aa_oi->oi_osfs = *msfs;
2800 out:
2801         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2802         RETURN(rc);
2803 }
2804
2805 static int osc_statfs_async(struct obd_export *exp,
2806                             struct obd_info *oinfo, __u64 max_age,
2807                             struct ptlrpc_request_set *rqset)
2808 {
2809         struct obd_device     *obd = class_exp2obd(exp);
2810         struct ptlrpc_request *req;
2811         struct osc_async_args *aa;
2812         int                    rc;
2813         ENTRY;
2814
2815         /* We could possibly pass max_age in the request (as an absolute
2816          * timestamp or a "seconds.usec ago") so the target can avoid doing
2817          * extra calls into the filesystem if that isn't necessary (e.g.
2818          * during mount that would help a bit).  Having relative timestamps
2819          * is not so great if request processing is slow, while absolute
2820          * timestamps are not ideal because they need time synchronization. */
2821         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2822         if (req == NULL)
2823                 RETURN(-ENOMEM);
2824
2825         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2826         if (rc) {
2827                 ptlrpc_request_free(req);
2828                 RETURN(rc);
2829         }
2830         ptlrpc_request_set_replen(req);
2831         req->rq_request_portal = OST_CREATE_PORTAL;
2832         ptlrpc_at_set_req_timeout(req);
2833
2834         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2835                 /* procfs requests not want stat in wait for avoid deadlock */
2836                 req->rq_no_resend = 1;
2837                 req->rq_no_delay = 1;
2838         }
2839
2840         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2841         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2842         aa = ptlrpc_req_async_args(req);
2843         aa->aa_oi = oinfo;
2844
2845         ptlrpc_set_add_req(rqset, req);
2846         RETURN(0);
2847 }
2848
2849 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2850                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2851 {
2852         struct obd_device     *obd = class_exp2obd(exp);
2853         struct obd_statfs     *msfs;
2854         struct ptlrpc_request *req;
2855         struct obd_import     *imp = NULL;
2856         int rc;
2857         ENTRY;
2858
2859         /*Since the request might also come from lprocfs, so we need
2860          *sync this with client_disconnect_export Bug15684*/
2861         down_read(&obd->u.cli.cl_sem);
2862         if (obd->u.cli.cl_import)
2863                 imp = class_import_get(obd->u.cli.cl_import);
2864         up_read(&obd->u.cli.cl_sem);
2865         if (!imp)
2866                 RETURN(-ENODEV);
2867
2868         /* We could possibly pass max_age in the request (as an absolute
2869          * timestamp or a "seconds.usec ago") so the target can avoid doing
2870          * extra calls into the filesystem if that isn't necessary (e.g.
2871          * during mount that would help a bit).  Having relative timestamps
2872          * is not so great if request processing is slow, while absolute
2873          * timestamps are not ideal because they need time synchronization. */
2874         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2875
2876         class_import_put(imp);
2877
2878         if (req == NULL)
2879                 RETURN(-ENOMEM);
2880
2881         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2882         if (rc) {
2883                 ptlrpc_request_free(req);
2884                 RETURN(rc);
2885         }
2886         ptlrpc_request_set_replen(req);
2887         req->rq_request_portal = OST_CREATE_PORTAL;
2888         ptlrpc_at_set_req_timeout(req);
2889
2890         if (flags & OBD_STATFS_NODELAY) {
2891                 /* procfs requests not want stat in wait for avoid deadlock */
2892                 req->rq_no_resend = 1;
2893                 req->rq_no_delay = 1;
2894         }
2895
2896         rc = ptlrpc_queue_wait(req);
2897         if (rc)
2898                 GOTO(out, rc);
2899
2900         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2901         if (msfs == NULL) {
2902                 GOTO(out, rc = -EPROTO);
2903         }
2904
2905         *osfs = *msfs;
2906
2907         EXIT;
2908  out:
2909         ptlrpc_req_finished(req);
2910         return rc;
2911 }
2912
2913 /* Retrieve object striping information.
2914  *
2915  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2916  * the maximum number of OST indices which will fit in the user buffer.
2917  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2918  */
2919 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2920 {
2921         /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2922         struct lov_user_md_v3 lum, *lumk;
2923         struct lov_user_ost_data_v1 *lmm_objects;
2924         int rc = 0, lum_size;
2925         ENTRY;
2926
2927         if (!lsm)
2928                 RETURN(-ENODATA);
2929
2930         /* we only need the header part from user space to get lmm_magic and
2931          * lmm_stripe_count, (the header part is common to v1 and v3) */
2932         lum_size = sizeof(struct lov_user_md_v1);
2933         if (copy_from_user(&lum, lump, lum_size))
2934                 RETURN(-EFAULT);
2935
2936         if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2937             (lum.lmm_magic != LOV_USER_MAGIC_V3))
2938                 RETURN(-EINVAL);
2939
2940         /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2941         LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2942         LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2943         LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2944
2945         /* we can use lov_mds_md_size() to compute lum_size
2946          * because lov_user_md_vX and lov_mds_md_vX have the same size */
2947         if (lum.lmm_stripe_count > 0) {
2948                 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2949                 OBD_ALLOC(lumk, lum_size);
2950                 if (!lumk)
2951                         RETURN(-ENOMEM);
2952
2953                 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2954                         lmm_objects =
2955                             &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2956                 else
2957                         lmm_objects = &(lumk->lmm_objects[0]);
2958                 lmm_objects->l_ost_oi = lsm->lsm_oi;
2959         } else {
2960                 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2961                 lumk = &lum;
2962         }
2963
2964         lumk->lmm_oi = lsm->lsm_oi;
2965         lumk->lmm_stripe_count = 1;
2966
2967         if (copy_to_user(lump, lumk, lum_size))
2968                 rc = -EFAULT;
2969
2970         if (lumk != &lum)
2971                 OBD_FREE(lumk, lum_size);
2972
2973         RETURN(rc);
2974 }
2975
2976
2977 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2978                          void *karg, void *uarg)
2979 {
2980         struct obd_device *obd = exp->exp_obd;
2981         struct obd_ioctl_data *data = karg;
2982         int err = 0;
2983         ENTRY;
2984
2985         if (!try_module_get(THIS_MODULE)) {
2986                 CERROR("Can't get module. Is it alive?");
2987                 return -EINVAL;
2988         }
2989         switch (cmd) {
2990         case OBD_IOC_LOV_GET_CONFIG: {
2991                 char *buf;
2992                 struct lov_desc *desc;
2993                 struct obd_uuid uuid;
2994
2995                 buf = NULL;
2996                 len = 0;
2997                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2998                         GOTO(out, err = -EINVAL);
2999
3000                 data = (struct obd_ioctl_data *)buf;
3001
3002                 if (sizeof(*desc) > data->ioc_inllen1) {
3003                         obd_ioctl_freedata(buf, len);
3004                         GOTO(out, err = -EINVAL);
3005                 }
3006
3007                 if (data->ioc_inllen2 < sizeof(uuid)) {
3008                         obd_ioctl_freedata(buf, len);
3009                         GOTO(out, err = -EINVAL);
3010                 }
3011
3012                 desc = (struct lov_desc *)data->ioc_inlbuf1;
3013                 desc->ld_tgt_count = 1;
3014                 desc->ld_active_tgt_count = 1;
3015                 desc->ld_default_stripe_count = 1;
3016                 desc->ld_default_stripe_size = 0;
3017                 desc->ld_default_stripe_offset = 0;
3018                 desc->ld_pattern = 0;
3019                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3020
3021                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3022
3023                 err = copy_to_user((void *)uarg, buf, len);
3024                 if (err)
3025                         err = -EFAULT;
3026                 obd_ioctl_freedata(buf, len);
3027                 GOTO(out, err);
3028         }
3029         case LL_IOC_LOV_SETSTRIPE:
3030                 err = obd_alloc_memmd(exp, karg);
3031                 if (err > 0)
3032                         err = 0;
3033                 GOTO(out, err);
3034         case LL_IOC_LOV_GETSTRIPE:
3035                 err = osc_getstripe(karg, uarg);
3036                 GOTO(out, err);
3037         case OBD_IOC_CLIENT_RECOVER:
3038                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3039                                             data->ioc_inlbuf1, 0);
3040                 if (err > 0)
3041                         err = 0;
3042                 GOTO(out, err);
3043         case IOC_OSC_SET_ACTIVE:
3044                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3045                                                data->ioc_offset);
3046                 GOTO(out, err);
3047         case OBD_IOC_POLL_QUOTACHECK:
3048                 err = osc_quota_poll_check(exp, (struct if_quotacheck *)karg);
3049                 GOTO(out, err);
3050         case OBD_IOC_PING_TARGET:
3051                 err = ptlrpc_obd_ping(obd);
3052                 GOTO(out, err);
3053         default:
3054                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3055                        cmd, current_comm());
3056                 GOTO(out, err = -ENOTTY);
3057         }
3058 out:
3059         module_put(THIS_MODULE);
3060         return err;
3061 }
3062
3063 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
3064                         obd_count keylen, void *key, __u32 *vallen, void *val,
3065                         struct lov_stripe_md *lsm)
3066 {
3067         ENTRY;
3068         if (!vallen || !val)
3069                 RETURN(-EFAULT);
3070
3071         if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3072                 __u32 *stripe = val;
3073                 *vallen = sizeof(*stripe);
3074                 *stripe = 0;
3075                 RETURN(0);
3076         } else if (KEY_IS(KEY_LAST_ID)) {
3077                 struct ptlrpc_request *req;
3078                 obd_id                *reply;
3079                 char                  *tmp;
3080                 int                    rc;
3081
3082                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3083                                            &RQF_OST_GET_INFO_LAST_ID);
3084                 if (req == NULL)
3085                         RETURN(-ENOMEM);
3086
3087                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3088                                      RCL_CLIENT, keylen);
3089                 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3090                 if (rc) {
3091                         ptlrpc_request_free(req);
3092                         RETURN(rc);
3093                 }
3094
3095                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3096                 memcpy(tmp, key, keylen);
3097
3098                 req->rq_no_delay = req->rq_no_resend = 1;
3099                 ptlrpc_request_set_replen(req);
3100                 rc = ptlrpc_queue_wait(req);
3101                 if (rc)
3102                         GOTO(out, rc);
3103
3104                 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3105                 if (reply == NULL)
3106                         GOTO(out, rc = -EPROTO);
3107
3108                 *((obd_id *)val) = *reply;
3109         out:
3110                 ptlrpc_req_finished(req);
3111                 RETURN(rc);
3112         } else if (KEY_IS(KEY_FIEMAP)) {
3113                 struct ll_fiemap_info_key *fm_key =
3114                                 (struct ll_fiemap_info_key *)key;
3115                 struct ldlm_res_id       res_id;
3116                 ldlm_policy_data_t       policy;
3117                 struct lustre_handle     lockh;
3118                 ldlm_mode_t              mode = 0;
3119                 struct ptlrpc_request   *req;
3120                 struct ll_user_fiemap   *reply;
3121                 char                    *tmp;
3122                 int                      rc;
3123
3124                 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
3125                         goto skip_locking;
3126
3127                 policy.l_extent.start = fm_key->fiemap.fm_start &
3128                                                 CFS_PAGE_MASK;
3129
3130                 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
3131                     fm_key->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
3132                         policy.l_extent.end = OBD_OBJECT_EOF;
3133                 else
3134                         policy.l_extent.end = (fm_key->fiemap.fm_start +
3135                                 fm_key->fiemap.fm_length +
3136                                 PAGE_CACHE_SIZE - 1) & CFS_PAGE_MASK;
3137
3138                 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
3139                 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
3140                                        LDLM_FL_BLOCK_GRANTED |
3141                                        LDLM_FL_LVB_READY,
3142                                        &res_id, LDLM_EXTENT, &policy,
3143                                        LCK_PR | LCK_PW, &lockh, 0);
3144                 if (mode) { /* lock is cached on client */
3145                         if (mode != LCK_PR) {
3146                                 ldlm_lock_addref(&lockh, LCK_PR);
3147                                 ldlm_lock_decref(&lockh, LCK_PW);
3148                         }
3149                 } else { /* no cached lock, needs acquire lock on server side */
3150                         fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
3151                         fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
3152                 }
3153
3154 skip_locking:
3155                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3156                                            &RQF_OST_GET_INFO_FIEMAP);
3157                 if (req == NULL)
3158                         GOTO(drop_lock, rc = -ENOMEM);
3159
3160                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3161                                      RCL_CLIENT, keylen);
3162                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3163                                      RCL_CLIENT, *vallen);
3164                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3165                                      RCL_SERVER, *vallen);
3166
3167                 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3168                 if (rc) {
3169                         ptlrpc_request_free(req);
3170                         GOTO(drop_lock, rc);
3171                 }
3172
3173                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3174                 memcpy(tmp, key, keylen);
3175                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3176                 memcpy(tmp, val, *vallen);
3177
3178                 ptlrpc_request_set_replen(req);
3179                 rc = ptlrpc_queue_wait(req);
3180                 if (rc)
3181                         GOTO(fini_req, rc);
3182
3183                 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3184                 if (reply == NULL)
3185                         GOTO(fini_req, rc = -EPROTO);
3186
3187                 memcpy(val, reply, *vallen);
3188 fini_req:
3189                 ptlrpc_req_finished(req);
3190 drop_lock:
3191                 if (mode)
3192                         ldlm_lock_decref(&lockh, LCK_PR);
3193                 RETURN(rc);
3194         }
3195
3196         RETURN(-EINVAL);
3197 }
3198
3199 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3200                               obd_count keylen, void *key, obd_count vallen,
3201                               void *val, struct ptlrpc_request_set *set)
3202 {
3203         struct ptlrpc_request *req;
3204         struct obd_device     *obd = exp->exp_obd;
3205         struct obd_import     *imp = class_exp2cliimp(exp);
3206         char                  *tmp;
3207         int                    rc;
3208         ENTRY;
3209
3210         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3211
3212         if (KEY_IS(KEY_CHECKSUM)) {
3213                 if (vallen != sizeof(int))
3214                         RETURN(-EINVAL);
3215                 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3216                 RETURN(0);
3217         }
3218
3219         if (KEY_IS(KEY_SPTLRPC_CONF)) {
3220                 sptlrpc_conf_client_adapt(obd);
3221                 RETURN(0);
3222         }
3223
3224         if (KEY_IS(KEY_FLUSH_CTX)) {
3225                 sptlrpc_import_flush_my_ctx(imp);
3226                 RETURN(0);
3227         }
3228
3229         if (KEY_IS(KEY_CACHE_SET)) {
3230                 struct client_obd *cli = &obd->u.cli;
3231
3232                 LASSERT(cli->cl_cache == NULL); /* only once */
3233                 cli->cl_cache = (struct cl_client_cache *)val;
3234                 cfs_atomic_inc(&cli->cl_cache->ccc_users);
3235                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
3236
3237                 /* add this osc into entity list */
3238                 LASSERT(cfs_list_empty(&cli->cl_lru_osc));
3239                 spin_lock(&cli->cl_cache->ccc_lru_lock);
3240                 cfs_list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
3241                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3242
3243                 RETURN(0);
3244         }
3245
3246         if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3247                 struct client_obd *cli = &obd->u.cli;
3248                 int nr = cfs_atomic_read(&cli->cl_lru_in_list) >> 1;
3249                 int target = *(int *)val;
3250
3251                 nr = osc_lru_shrink(env, cli, min(nr, target), true);
3252                 *(int *)val -= nr;
3253                 RETURN(0);
3254         }
3255
3256         if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3257                 RETURN(-EINVAL);
3258
3259         /* We pass all other commands directly to OST. Since nobody calls osc
3260            methods directly and everybody is supposed to go through LOV, we
3261            assume lov checked invalid values for us.
3262            The only recognised values so far are evict_by_nid and mds_conn.
3263            Even if something bad goes through, we'd get a -EINVAL from OST
3264            anyway. */
3265
3266         req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3267                                                 &RQF_OST_SET_GRANT_INFO :
3268                                                 &RQF_OBD_SET_INFO);
3269         if (req == NULL)
3270                 RETURN(-ENOMEM);
3271
3272         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3273                              RCL_CLIENT, keylen);
3274         if (!KEY_IS(KEY_GRANT_SHRINK))
3275                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3276                                      RCL_CLIENT, vallen);
3277         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3278         if (rc) {
3279                 ptlrpc_request_free(req);
3280                 RETURN(rc);
3281         }
3282
3283         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3284         memcpy(tmp, key, keylen);
3285         tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3286                                                         &RMF_OST_BODY :
3287                                                         &RMF_SETINFO_VAL);
3288         memcpy(tmp, val, vallen);
3289
3290         if (KEY_IS(KEY_GRANT_SHRINK)) {
3291                 struct osc_grant_args *aa;
3292                 struct obdo *oa;
3293
3294                 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3295                 aa = ptlrpc_req_async_args(req);
3296                 OBDO_ALLOC(oa);
3297                 if (!oa) {
3298                         ptlrpc_req_finished(req);
3299                         RETURN(-ENOMEM);
3300                 }
3301                 *oa = ((struct ost_body *)val)->oa;
3302                 aa->aa_oa = oa;
3303                 req->rq_interpret_reply = osc_shrink_grant_interpret;
3304         }
3305
3306         ptlrpc_request_set_replen(req);
3307         if (!KEY_IS(KEY_GRANT_SHRINK)) {
3308                 LASSERT(set != NULL);
3309                 ptlrpc_set_add_req(set, req);
3310                 ptlrpc_check_set(NULL, set);
3311         } else
3312                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3313
3314         RETURN(0);
3315 }
3316
3317
3318 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
3319                          struct obd_device *disk_obd, int *index)
3320 {
3321         /* this code is not supposed to be used with LOD/OSP
3322          * to be removed soon */
3323         LBUG();
3324         return 0;
3325 }
3326
3327 static int osc_llog_finish(struct obd_device *obd, int count)
3328 {
3329         struct llog_ctxt *ctxt;
3330
3331         ENTRY;
3332
3333         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
3334         if (ctxt) {
3335                 llog_cat_close(NULL, ctxt->loc_handle);
3336                 llog_cleanup(NULL, ctxt);
3337         }
3338
3339         ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3340         if (ctxt)
3341                 llog_cleanup(NULL, ctxt);
3342         RETURN(0);
3343 }
3344
3345 static int osc_reconnect(const struct lu_env *env,
3346                          struct obd_export *exp, struct obd_device *obd,
3347                          struct obd_uuid *cluuid,
3348                          struct obd_connect_data *data,
3349                          void *localdata)
3350 {
3351         struct client_obd *cli = &obd->u.cli;
3352
3353         if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3354                 long lost_grant;
3355
3356                 client_obd_list_lock(&cli->cl_loi_list_lock);
3357                 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
3358                                 2 * cli_brw_size(obd);
3359                 lost_grant = cli->cl_lost_grant;
3360                 cli->cl_lost_grant = 0;
3361                 client_obd_list_unlock(&cli->cl_loi_list_lock);
3362
3363                 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
3364                        " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3365                        data->ocd_version, data->ocd_grant, lost_grant);
3366         }
3367
3368         RETURN(0);
3369 }
3370
3371 static int osc_disconnect(struct obd_export *exp)
3372 {
3373         struct obd_device *obd = class_exp2obd(exp);
3374         struct llog_ctxt  *ctxt;
3375         int rc;
3376
3377         ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
3378         if (ctxt) {
3379                 if (obd->u.cli.cl_conn_count == 1) {
3380                         /* Flush any remaining cancel messages out to the
3381                          * target */
3382                         llog_sync(ctxt, exp, 0);
3383                 }
3384                 llog_ctxt_put(ctxt);
3385         } else {
3386                 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
3387                        obd);
3388         }
3389
3390         rc = client_disconnect_export(exp);
3391         /**
3392          * Initially we put del_shrink_grant before disconnect_export, but it
3393          * causes the following problem if setup (connect) and cleanup
3394          * (disconnect) are tangled together.
3395          *      connect p1                     disconnect p2
3396          *   ptlrpc_connect_import
3397          *     ...............               class_manual_cleanup
3398          *                                     osc_disconnect
3399          *                                     del_shrink_grant
3400          *   ptlrpc_connect_interrupt
3401          *     init_grant_shrink
3402          *   add this client to shrink list
3403          *                                      cleanup_osc
3404          * Bang! pinger trigger the shrink.
3405          * So the osc should be disconnected from the shrink list, after we
3406          * are sure the import has been destroyed. BUG18662
3407          */
3408         if (obd->u.cli.cl_import == NULL)
3409                 osc_del_shrink_grant(&obd->u.cli);
3410         return rc;
3411 }
3412
3413 static int osc_import_event(struct obd_device *obd,
3414                             struct obd_import *imp,
3415                             enum obd_import_event event)
3416 {
3417         struct client_obd *cli;
3418         int rc = 0;
3419
3420         ENTRY;
3421         LASSERT(imp->imp_obd == obd);
3422
3423         switch (event) {
3424         case IMP_EVENT_DISCON: {
3425                 cli = &obd->u.cli;
3426                 client_obd_list_lock(&cli->cl_loi_list_lock);
3427                 cli->cl_avail_grant = 0;
3428                 cli->cl_lost_grant = 0;
3429                 client_obd_list_unlock(&cli->cl_loi_list_lock);
3430                 break;
3431         }
3432         case IMP_EVENT_INACTIVE: {
3433                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3434                 break;
3435         }
3436         case IMP_EVENT_INVALIDATE: {
3437                 struct ldlm_namespace *ns = obd->obd_namespace;
3438                 struct lu_env         *env;
3439                 int                    refcheck;
3440
3441                 env = cl_env_get(&refcheck);
3442                 if (!IS_ERR(env)) {
3443                         /* Reset grants */
3444                         cli = &obd->u.cli;
3445                         /* all pages go to failing rpcs due to the invalid
3446                          * import */
3447                         osc_io_unplug(env, cli, NULL, PDL_POLICY_ROUND);
3448
3449                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3450                         cl_env_put(env, &refcheck);
3451                 } else
3452                         rc = PTR_ERR(env);
3453                 break;
3454         }
3455         case IMP_EVENT_ACTIVE: {
3456                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3457                 break;
3458         }
3459         case IMP_EVENT_OCD: {
3460                 struct obd_connect_data *ocd = &imp->imp_connect_data;
3461
3462                 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3463                         osc_init_grant(&obd->u.cli, ocd);
3464
3465                 /* See bug 7198 */
3466                 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3467                         imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3468
3469                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3470                 break;
3471         }
3472         case IMP_EVENT_DEACTIVATE: {
3473                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3474                 break;
3475         }
3476         case IMP_EVENT_ACTIVATE: {
3477                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3478                 break;
3479         }
3480         default:
3481                 CERROR("Unknown import event %d\n", event);
3482                 LBUG();
3483         }
3484         RETURN(rc);
3485 }
3486
3487 /**
3488  * Determine whether the lock can be canceled before replaying the lock
3489  * during recovery, see bug16774 for detailed information.
3490  *
3491  * \retval zero the lock can't be canceled
3492  * \retval other ok to cancel
3493  */
3494 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
3495 {
3496         /*
3497          * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
3498          *
3499          * XXX as a future improvement, we can also cancel unused write lock
3500          * if it doesn't have dirty data and active mmaps.
3501          */
3502         if (lock->l_resource->lr_type == LDLM_EXTENT &&
3503             (lock->l_granted_mode == LCK_PR || lock->l_granted_mode == LCK_CR)&&
3504             osc_ldlm_weigh_ast(lock) == 0)
3505                 RETURN(1);
3506
3507         RETURN(0);
3508 }
3509
3510 static int brw_queue_work(const struct lu_env *env, void *data)
3511 {
3512         struct client_obd *cli = data;
3513
3514         CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3515
3516         osc_io_unplug(env, cli, NULL, PDL_POLICY_SAME);
3517         RETURN(0);
3518 }
3519
3520 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3521 {
3522         struct lprocfs_static_vars lvars = { 0 };
3523         struct client_obd          *cli = &obd->u.cli;
3524         void                       *handler;
3525         int                        rc;
3526         ENTRY;
3527
3528         rc = ptlrpcd_addref();
3529         if (rc)
3530                 RETURN(rc);
3531
3532         rc = client_obd_setup(obd, lcfg);
3533         if (rc)
3534                 GOTO(out_ptlrpcd, rc);
3535
3536         handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3537         if (IS_ERR(handler))
3538                 GOTO(out_client_setup, rc = PTR_ERR(handler));
3539         cli->cl_writeback_work = handler;
3540
3541         handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3542         if (IS_ERR(handler))
3543                 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3544         cli->cl_lru_work = handler;
3545
3546         rc = osc_quota_setup(obd);
3547         if (rc)
3548                 GOTO(out_ptlrpcd_work, rc);
3549
3550         cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3551         lprocfs_osc_init_vars(&lvars);
3552         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
3553                 lproc_osc_attach_seqstat(obd);
3554                 sptlrpc_lprocfs_cliobd_attach(obd);
3555                 ptlrpc_lprocfs_register_obd(obd);
3556         }
3557
3558         /* We need to allocate a few requests more, because
3559          * brw_interpret tries to create new requests before freeing
3560          * previous ones, Ideally we want to have 2x max_rpcs_in_flight
3561          * reserved, but I'm afraid that might be too much wasted RAM
3562          * in fact, so 2 is just my guess and still should work. */
3563         cli->cl_import->imp_rq_pool =
3564                 ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
3565                                     OST_MAXREQSIZE,
3566                                     ptlrpc_add_rqs_to_pool);
3567
3568         CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3569         ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
3570         RETURN(rc);
3571
3572 out_ptlrpcd_work:
3573         if (cli->cl_writeback_work != NULL) {
3574                 ptlrpcd_destroy_work(cli->cl_writeback_work);
3575                 cli->cl_writeback_work = NULL;
3576         }
3577         if (cli->cl_lru_work != NULL) {
3578                 ptlrpcd_destroy_work(cli->cl_lru_work);
3579                 cli->cl_lru_work = NULL;
3580         }
3581 out_client_setup:
3582         client_obd_cleanup(obd);
3583 out_ptlrpcd:
3584         ptlrpcd_decref();
3585         RETURN(rc);
3586 }
3587
3588 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3589 {
3590         int rc = 0;
3591         ENTRY;
3592
3593         switch (stage) {
3594         case OBD_CLEANUP_EARLY: {
3595                 struct obd_import *imp;
3596                 imp = obd->u.cli.cl_import;
3597                 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3598                 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3599                 ptlrpc_deactivate_import(imp);
3600                 spin_lock(&imp->imp_lock);
3601                 imp->imp_pingable = 0;
3602                 spin_unlock(&imp->imp_lock);
3603                 break;
3604         }
3605         case OBD_CLEANUP_EXPORTS: {
3606                 struct client_obd *cli = &obd->u.cli;
3607                 /* LU-464
3608                  * for echo client, export may be on zombie list, wait for
3609                  * zombie thread to cull it, because cli.cl_import will be
3610                  * cleared in client_disconnect_export():
3611                  *   class_export_destroy() -> obd_cleanup() ->
3612                  *   echo_device_free() -> echo_client_cleanup() ->
3613                  *   obd_disconnect() -> osc_disconnect() ->
3614                  *   client_disconnect_export()
3615                  */
3616                 obd_zombie_barrier();
3617                 if (cli->cl_writeback_work) {
3618                         ptlrpcd_destroy_work(cli->cl_writeback_work);
3619                         cli->cl_writeback_work = NULL;
3620                 }
3621                 if (cli->cl_lru_work) {
3622                         ptlrpcd_destroy_work(cli->cl_lru_work);
3623                         cli->cl_lru_work = NULL;
3624                 }
3625                 obd_cleanup_client_import(obd);
3626                 ptlrpc_lprocfs_unregister_obd(obd);
3627                 lprocfs_obd_cleanup(obd);
3628                 rc = obd_llog_finish(obd, 0);
3629                 if (rc != 0)
3630                         CERROR("failed to cleanup llogging subsystems\n");
3631                 break;
3632                 }
3633         }
3634         RETURN(rc);
3635 }
3636
3637 int osc_cleanup(struct obd_device *obd)
3638 {
3639         struct client_obd *cli = &obd->u.cli;
3640         int rc;
3641
3642         ENTRY;
3643
3644         /* lru cleanup */
3645         if (cli->cl_cache != NULL) {
3646                 LASSERT(cfs_atomic_read(&cli->cl_cache->ccc_users) > 0);
3647                 spin_lock(&cli->cl_cache->ccc_lru_lock);
3648                 cfs_list_del_init(&cli->cl_lru_osc);
3649                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3650                 cli->cl_lru_left = NULL;
3651                 cfs_atomic_dec(&cli->cl_cache->ccc_users);
3652                 cli->cl_cache = NULL;
3653         }
3654
3655         /* free memory of osc quota cache */
3656         osc_quota_cleanup(obd);
3657
3658