Whamcloud - gitweb
22bf7d9d219294e2b7d02d2ce40a8314e7424d4a
[fs/lustre-release.git] / lustre / osc / osc_request.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
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 #ifndef EXPORT_SYMTAB
38 # define EXPORT_SYMTAB
39 #endif
40 #define DEBUG_SUBSYSTEM S_OSC
41
42 #include <libcfs/libcfs.h>
43
44 #ifndef __KERNEL__
45 # include <liblustre.h>
46 #endif
47
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre/lustre_user.h>
51 #include <obd_cksum.h>
52 #include <obd_ost.h>
53 #include <obd_lov.h>
54
55 #ifdef  __CYGWIN__
56 # include <ctype.h>
57 #endif
58
59 #include <lustre_ha.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_debug.h>
63 #include <lustre_param.h>
64 #include "osc_internal.h"
65
66 static quota_interface_t *quota_interface = NULL;
67 extern quota_interface_t osc_quota_interface;
68
69 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
70 static int brw_interpret(const struct lu_env *env,
71                          struct ptlrpc_request *req, void *data, int rc);
72 int osc_cleanup(struct obd_device *obd);
73
74 /* Pack OSC object metadata for disk storage (LE byte order). */
75 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
76                       struct lov_stripe_md *lsm)
77 {
78         int lmm_size;
79         ENTRY;
80
81         lmm_size = sizeof(**lmmp);
82         if (!lmmp)
83                 RETURN(lmm_size);
84
85         if (*lmmp && !lsm) {
86                 OBD_FREE(*lmmp, lmm_size);
87                 *lmmp = NULL;
88                 RETURN(0);
89         }
90
91         if (!*lmmp) {
92                 OBD_ALLOC(*lmmp, lmm_size);
93                 if (!*lmmp)
94                         RETURN(-ENOMEM);
95         }
96
97         if (lsm) {
98                 LASSERT(lsm->lsm_object_id);
99                 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
100                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101                 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
102         }
103
104         RETURN(lmm_size);
105 }
106
107 /* Unpack OSC object metadata from disk storage (LE byte order). */
108 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
109                         struct lov_mds_md *lmm, int lmm_bytes)
110 {
111         int lsm_size;
112         ENTRY;
113
114         if (lmm != NULL) {
115                 if (lmm_bytes < sizeof (*lmm)) {
116                         CERROR("lov_mds_md too small: %d, need %d\n",
117                                lmm_bytes, (int)sizeof(*lmm));
118                         RETURN(-EINVAL);
119                 }
120                 /* XXX LOV_MAGIC etc check? */
121
122                 if (lmm->lmm_object_id == 0) {
123                         CERROR("lov_mds_md: zero lmm_object_id\n");
124                         RETURN(-EINVAL);
125                 }
126         }
127
128         lsm_size = lov_stripe_md_size(1);
129         if (lsmp == NULL)
130                 RETURN(lsm_size);
131
132         if (*lsmp != NULL && lmm == NULL) {
133                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
134                 OBD_FREE(*lsmp, lsm_size);
135                 *lsmp = NULL;
136                 RETURN(0);
137         }
138
139         if (*lsmp == NULL) {
140                 OBD_ALLOC(*lsmp, lsm_size);
141                 if (*lsmp == NULL)
142                         RETURN(-ENOMEM);
143                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
144                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
145                         OBD_FREE(*lsmp, lsm_size);
146                         RETURN(-ENOMEM);
147                 }
148                 loi_init((*lsmp)->lsm_oinfo[0]);
149         }
150
151         if (lmm != NULL) {
152                 /* XXX zero *lsmp? */
153                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
154                 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
155                 LASSERT((*lsmp)->lsm_object_id);
156                 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
157         }
158
159         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
160
161         RETURN(lsm_size);
162 }
163
164 static inline void osc_pack_capa(struct ptlrpc_request *req,
165                                  struct ost_body *body, void *capa)
166 {
167         struct obd_capa *oc = (struct obd_capa *)capa;
168         struct lustre_capa *c;
169
170         if (!capa)
171                 return;
172
173         c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
174         LASSERT(c);
175         capa_cpy(c, oc);
176         body->oa.o_valid |= OBD_MD_FLOSSCAPA;
177         DEBUG_CAPA(D_SEC, c, "pack");
178 }
179
180 static inline void osc_pack_req_body(struct ptlrpc_request *req,
181                                      struct obd_info *oinfo)
182 {
183         struct ost_body *body;
184
185         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
186         LASSERT(body);
187
188         lustre_set_wire_obdo(&body->oa, 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(aa->aa_oi->oi_oa, &body->oa);
217
218                 /* This should really be sent by the OST */
219                 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
220                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
221         } else {
222                 CDEBUG(D_INFO, "can't unpack ost_body\n");
223                 rc = -EPROTO;
224                 aa->aa_oi->oi_oa->o_valid = 0;
225         }
226 out:
227         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
228         RETURN(rc);
229 }
230
231 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
232                              struct ptlrpc_request_set *set)
233 {
234         struct ptlrpc_request *req;
235         struct osc_async_args *aa;
236         int                    rc;
237         ENTRY;
238
239         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
240         if (req == NULL)
241                 RETURN(-ENOMEM);
242
243         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
244         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
245         if (rc) {
246                 ptlrpc_request_free(req);
247                 RETURN(rc);
248         }
249
250         osc_pack_req_body(req, oinfo);
251
252         ptlrpc_request_set_replen(req);
253         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
254
255         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
256         aa = ptlrpc_req_async_args(req);
257         aa->aa_oi = oinfo;
258
259         ptlrpc_set_add_req(set, req);
260         RETURN(0);
261 }
262
263 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
264 {
265         struct ptlrpc_request *req;
266         struct ost_body       *body;
267         int                    rc;
268         ENTRY;
269
270         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
271         if (req == NULL)
272                 RETURN(-ENOMEM);
273
274         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
275         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
276         if (rc) {
277                 ptlrpc_request_free(req);
278                 RETURN(rc);
279         }
280
281         osc_pack_req_body(req, oinfo);
282
283         ptlrpc_request_set_replen(req);
284
285         rc = ptlrpc_queue_wait(req);
286         if (rc)
287                 GOTO(out, rc);
288
289         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
290         if (body == NULL)
291                 GOTO(out, rc = -EPROTO);
292
293         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
294         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
295
296         /* This should really be sent by the OST */
297         oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
298         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
299
300         EXIT;
301  out:
302         ptlrpc_req_finished(req);
303         return rc;
304 }
305
306 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
307                        struct obd_trans_info *oti)
308 {
309         struct ptlrpc_request *req;
310         struct ost_body       *body;
311         int                    rc;
312         ENTRY;
313
314         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
315
316         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
317         if (req == NULL)
318                 RETURN(-ENOMEM);
319
320         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
321         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
322         if (rc) {
323                 ptlrpc_request_free(req);
324                 RETURN(rc);
325         }
326
327         osc_pack_req_body(req, oinfo);
328
329         ptlrpc_request_set_replen(req);
330
331         rc = ptlrpc_queue_wait(req);
332         if (rc)
333                 GOTO(out, rc);
334
335         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
336         if (body == NULL)
337                 GOTO(out, rc = -EPROTO);
338
339         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
340
341         EXIT;
342 out:
343         ptlrpc_req_finished(req);
344         RETURN(rc);
345 }
346
347 static int osc_setattr_interpret(const struct lu_env *env,
348                                  struct ptlrpc_request *req,
349                                  struct osc_setattr_args *sa, int rc)
350 {
351         struct ost_body *body;
352         ENTRY;
353
354         if (rc != 0)
355                 GOTO(out, rc);
356
357         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
358         if (body == NULL)
359                 GOTO(out, rc = -EPROTO);
360
361         lustre_get_wire_obdo(sa->sa_oa, &body->oa);
362 out:
363         rc = sa->sa_upcall(sa->sa_cookie, rc);
364         RETURN(rc);
365 }
366
367 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
368                            struct obd_trans_info *oti,
369                            obd_enqueue_update_f upcall, void *cookie,
370                            struct ptlrpc_request_set *rqset)
371 {
372         struct ptlrpc_request   *req;
373         struct osc_setattr_args *sa;
374         int                      rc;
375         ENTRY;
376
377         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
378         if (req == NULL)
379                 RETURN(-ENOMEM);
380
381         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
382         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
383         if (rc) {
384                 ptlrpc_request_free(req);
385                 RETURN(rc);
386         }
387
388         if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
389                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
390
391         osc_pack_req_body(req, oinfo);
392
393         ptlrpc_request_set_replen(req);
394
395         /* do mds to ost setattr asynchronously */
396         if (!rqset) {
397                 /* Do not wait for response. */
398                 ptlrpcd_add_req(req, PSCOPE_OTHER);
399         } else {
400                 req->rq_interpret_reply =
401                         (ptlrpc_interpterer_t)osc_setattr_interpret;
402
403                 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
404                 sa = ptlrpc_req_async_args(req);
405                 sa->sa_oa = oinfo->oi_oa;
406                 sa->sa_upcall = upcall;
407                 sa->sa_cookie = cookie;
408
409                 if (rqset == PTLRPCD_SET)
410                         ptlrpcd_add_req(req, PSCOPE_OTHER);
411                 else
412                         ptlrpc_set_add_req(rqset, req);
413         }
414
415         RETURN(0);
416 }
417
418 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
419                              struct obd_trans_info *oti,
420                              struct ptlrpc_request_set *rqset)
421 {
422         return osc_setattr_async_base(exp, oinfo, oti,
423                                       oinfo->oi_cb_up, oinfo, rqset);
424 }
425
426 int osc_real_create(struct obd_export *exp, struct obdo *oa,
427                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
428 {
429         struct ptlrpc_request *req;
430         struct ost_body       *body;
431         struct lov_stripe_md  *lsm;
432         int                    rc;
433         ENTRY;
434
435         LASSERT(oa);
436         LASSERT(ea);
437
438         lsm = *ea;
439         if (!lsm) {
440                 rc = obd_alloc_memmd(exp, &lsm);
441                 if (rc < 0)
442                         RETURN(rc);
443         }
444
445         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
446         if (req == NULL)
447                 GOTO(out, rc = -ENOMEM);
448
449         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
450         if (rc) {
451                 ptlrpc_request_free(req);
452                 GOTO(out, rc);
453         }
454
455         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
456         LASSERT(body);
457         lustre_set_wire_obdo(&body->oa, oa);
458
459         ptlrpc_request_set_replen(req);
460
461         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
462             oa->o_flags == OBD_FL_DELORPHAN) {
463                 DEBUG_REQ(D_HA, req,
464                           "delorphan from OST integration");
465                 /* Don't resend the delorphan req */
466                 req->rq_no_resend = req->rq_no_delay = 1;
467         }
468
469         rc = ptlrpc_queue_wait(req);
470         if (rc)
471                 GOTO(out_req, rc);
472
473         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
474         if (body == NULL)
475                 GOTO(out_req, rc = -EPROTO);
476
477         lustre_get_wire_obdo(oa, &body->oa);
478
479         /* This should really be sent by the OST */
480         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
481         oa->o_valid |= OBD_MD_FLBLKSZ;
482
483         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
484          * have valid lsm_oinfo data structs, so don't go touching that.
485          * This needs to be fixed in a big way.
486          */
487         lsm->lsm_object_id = oa->o_id;
488         lsm->lsm_object_seq = oa->o_seq;
489         *ea = lsm;
490
491         if (oti != NULL) {
492                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
493
494                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
495                         if (!oti->oti_logcookies)
496                                 oti_alloc_cookies(oti, 1);
497                         *oti->oti_logcookies = oa->o_lcookie;
498                 }
499         }
500
501         CDEBUG(D_HA, "transno: "LPD64"\n",
502                lustre_msg_get_transno(req->rq_repmsg));
503 out_req:
504         ptlrpc_req_finished(req);
505 out:
506         if (rc && !*ea)
507                 obd_free_memmd(exp, &lsm);
508         RETURN(rc);
509 }
510
511 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
512                    obd_enqueue_update_f upcall, void *cookie,
513                    struct ptlrpc_request_set *rqset)
514 {
515         struct ptlrpc_request   *req;
516         struct osc_setattr_args *sa;
517         struct ost_body         *body;
518         int                      rc;
519         ENTRY;
520
521         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
522         if (req == NULL)
523                 RETURN(-ENOMEM);
524
525         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
526         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
527         if (rc) {
528                 ptlrpc_request_free(req);
529                 RETURN(rc);
530         }
531         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
532         ptlrpc_at_set_req_timeout(req);
533
534         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
535         LASSERT(body);
536         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
537         osc_pack_capa(req, body, oinfo->oi_capa);
538
539         ptlrpc_request_set_replen(req);
540
541
542         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
543         CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
544         sa = ptlrpc_req_async_args(req);
545         sa->sa_oa     = oinfo->oi_oa;
546         sa->sa_upcall = upcall;
547         sa->sa_cookie = cookie;
548         if (rqset == PTLRPCD_SET)
549                 ptlrpcd_add_req(req, PSCOPE_OTHER);
550         else
551                 ptlrpc_set_add_req(rqset, req);
552
553         RETURN(0);
554 }
555
556 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
557                      struct obd_trans_info *oti,
558                      struct ptlrpc_request_set *rqset)
559 {
560         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
561         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
562         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
563         return osc_punch_base(exp, oinfo,
564                               oinfo->oi_cb_up, oinfo, rqset);
565 }
566
567 static int osc_sync(struct obd_export *exp, struct obdo *oa,
568                     struct lov_stripe_md *md, obd_size start, obd_size end,
569                     void *capa)
570 {
571         struct ptlrpc_request *req;
572         struct ost_body       *body;
573         int                    rc;
574         ENTRY;
575
576         if (!oa) {
577                 CDEBUG(D_INFO, "oa NULL\n");
578                 RETURN(-EINVAL);
579         }
580
581         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
582         if (req == NULL)
583                 RETURN(-ENOMEM);
584
585         osc_set_capa_size(req, &RMF_CAPA1, capa);
586         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
587         if (rc) {
588                 ptlrpc_request_free(req);
589                 RETURN(rc);
590         }
591
592         /* overload the size and blocks fields in the oa with start/end */
593         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
594         LASSERT(body);
595         lustre_set_wire_obdo(&body->oa, oa);
596         body->oa.o_size = start;
597         body->oa.o_blocks = end;
598         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
599         osc_pack_capa(req, body, capa);
600
601         ptlrpc_request_set_replen(req);
602
603         rc = ptlrpc_queue_wait(req);
604         if (rc)
605                 GOTO(out, rc);
606
607         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
608         if (body == NULL)
609                 GOTO(out, rc = -EPROTO);
610
611         lustre_get_wire_obdo(oa, &body->oa);
612
613         EXIT;
614  out:
615         ptlrpc_req_finished(req);
616         return rc;
617 }
618
619 /* Find and cancel locally locks matched by @mode in the resource found by
620  * @objid. Found locks are added into @cancel list. Returns the amount of
621  * locks added to @cancels list. */
622 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
623                                    cfs_list_t *cancels,
624                                    ldlm_mode_t mode, int lock_flags)
625 {
626         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
627         struct ldlm_res_id res_id;
628         struct ldlm_resource *res;
629         int count;
630         ENTRY;
631
632         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
633         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
634         if (res == NULL)
635                 RETURN(0);
636
637         LDLM_RESOURCE_ADDREF(res);
638         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
639                                            lock_flags, 0, NULL);
640         LDLM_RESOURCE_DELREF(res);
641         ldlm_resource_putref(res);
642         RETURN(count);
643 }
644
645 static int osc_destroy_interpret(const struct lu_env *env,
646                                  struct ptlrpc_request *req, void *data,
647                                  int rc)
648 {
649         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
650
651         cfs_atomic_dec(&cli->cl_destroy_in_flight);
652         cfs_waitq_signal(&cli->cl_destroy_waitq);
653         return 0;
654 }
655
656 static int osc_can_send_destroy(struct client_obd *cli)
657 {
658         if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
659             cli->cl_max_rpcs_in_flight) {
660                 /* The destroy request can be sent */
661                 return 1;
662         }
663         if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
664             cli->cl_max_rpcs_in_flight) {
665                 /*
666                  * The counter has been modified between the two atomic
667                  * operations.
668                  */
669                 cfs_waitq_signal(&cli->cl_destroy_waitq);
670         }
671         return 0;
672 }
673
674 /* Destroy requests can be async always on the client, and we don't even really
675  * care about the return code since the client cannot do anything at all about
676  * a destroy failure.
677  * When the MDS is unlinking a filename, it saves the file objects into a
678  * recovery llog, and these object records are cancelled when the OST reports
679  * they were destroyed and sync'd to disk (i.e. transaction committed).
680  * If the client dies, or the OST is down when the object should be destroyed,
681  * the records are not cancelled, and when the OST reconnects to the MDS next,
682  * it will retrieve the llog unlink logs and then sends the log cancellation
683  * cookies to the MDS after committing destroy transactions. */
684 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
685                        struct lov_stripe_md *ea, struct obd_trans_info *oti,
686                        struct obd_export *md_export, void *capa)
687 {
688         struct client_obd     *cli = &exp->exp_obd->u.cli;
689         struct ptlrpc_request *req;
690         struct ost_body       *body;
691         CFS_LIST_HEAD(cancels);
692         int rc, count;
693         ENTRY;
694
695         if (!oa) {
696                 CDEBUG(D_INFO, "oa NULL\n");
697                 RETURN(-EINVAL);
698         }
699
700         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
701                                         LDLM_FL_DISCARD_DATA);
702
703         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
704         if (req == NULL) {
705                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
706                 RETURN(-ENOMEM);
707         }
708
709         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
710         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
711                                0, &cancels, count);
712         if (rc) {
713                 ptlrpc_request_free(req);
714                 RETURN(rc);
715         }
716
717         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
718         ptlrpc_at_set_req_timeout(req);
719
720         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
721                 oa->o_lcookie = *oti->oti_logcookies;
722         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
723         LASSERT(body);
724         lustre_set_wire_obdo(&body->oa, oa);
725
726         osc_pack_capa(req, body, (struct obd_capa *)capa);
727         ptlrpc_request_set_replen(req);
728
729         /* don't throttle destroy RPCs for the MDT */
730         if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
731                 req->rq_interpret_reply = osc_destroy_interpret;
732                 if (!osc_can_send_destroy(cli)) {
733                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
734                                                           NULL);
735
736                         /*
737                          * Wait until the number of on-going destroy RPCs drops
738                          * under max_rpc_in_flight
739                          */
740                         l_wait_event_exclusive(cli->cl_destroy_waitq,
741                                                osc_can_send_destroy(cli), &lwi);
742                 }
743         }
744
745         /* Do not wait for response */
746         ptlrpcd_add_req(req, PSCOPE_OTHER);
747         RETURN(0);
748 }
749
750 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
751                                 long writing_bytes)
752 {
753         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
754
755         LASSERT(!(oa->o_valid & bits));
756
757         oa->o_valid |= bits;
758         client_obd_list_lock(&cli->cl_loi_list_lock);
759         oa->o_dirty = cli->cl_dirty;
760         if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
761                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
762                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
763                 oa->o_undirty = 0;
764         } else if (cfs_atomic_read(&obd_dirty_pages) -
765                    cfs_atomic_read(&obd_dirty_transit_pages) >
766                    obd_max_dirty_pages + 1){
767                 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
768                  * not covered by a lock thus they may safely race and trip
769                  * this CERROR() unless we add in a small fudge factor (+1). */
770                 CERROR("dirty %d - %d > system dirty_max %d\n",
771                        cfs_atomic_read(&obd_dirty_pages),
772                        cfs_atomic_read(&obd_dirty_transit_pages),
773                        obd_max_dirty_pages);
774                 oa->o_undirty = 0;
775         } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
776                 CERROR("dirty %lu - dirty_max %lu too big???\n",
777                        cli->cl_dirty, cli->cl_dirty_max);
778                 oa->o_undirty = 0;
779         } else {
780                 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
781                                 (cli->cl_max_rpcs_in_flight + 1);
782                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
783         }
784         oa->o_grant = cli->cl_avail_grant;
785         oa->o_dropped = cli->cl_lost_grant;
786         cli->cl_lost_grant = 0;
787         client_obd_list_unlock(&cli->cl_loi_list_lock);
788         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
789                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
790
791 }
792
793 static void osc_update_next_shrink(struct client_obd *cli)
794 {
795         cli->cl_next_shrink_grant =
796                 cfs_time_shift(cli->cl_grant_shrink_interval);
797         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
798                cli->cl_next_shrink_grant);
799 }
800
801 /* caller must hold loi_list_lock */
802 static void osc_consume_write_grant(struct client_obd *cli,
803                                     struct brw_page *pga)
804 {
805         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
806         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
807         cfs_atomic_inc(&obd_dirty_pages);
808         cli->cl_dirty += CFS_PAGE_SIZE;
809         cli->cl_avail_grant -= CFS_PAGE_SIZE;
810         pga->flag |= OBD_BRW_FROM_GRANT;
811         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
812                CFS_PAGE_SIZE, pga, pga->pg);
813         LASSERT(cli->cl_avail_grant >= 0);
814         osc_update_next_shrink(cli);
815 }
816
817 /* the companion to osc_consume_write_grant, called when a brw has completed.
818  * must be called with the loi lock held. */
819 static void osc_release_write_grant(struct client_obd *cli,
820                                     struct brw_page *pga, int sent)
821 {
822         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
823         ENTRY;
824
825         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
826         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
827                 EXIT;
828                 return;
829         }
830
831         pga->flag &= ~OBD_BRW_FROM_GRANT;
832         cfs_atomic_dec(&obd_dirty_pages);
833         cli->cl_dirty -= CFS_PAGE_SIZE;
834         if (pga->flag & OBD_BRW_NOCACHE) {
835                 pga->flag &= ~OBD_BRW_NOCACHE;
836                 cfs_atomic_dec(&obd_dirty_transit_pages);
837                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
838         }
839         if (!sent) {
840                 cli->cl_lost_grant += CFS_PAGE_SIZE;
841                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
842                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
843         } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
844                 /* For short writes we shouldn't count parts of pages that
845                  * span a whole block on the OST side, or our accounting goes
846                  * wrong.  Should match the code in filter_grant_check. */
847                 int offset = pga->off & ~CFS_PAGE_MASK;
848                 int count = pga->count + (offset & (blocksize - 1));
849                 int end = (offset + pga->count) & (blocksize - 1);
850                 if (end)
851                         count += blocksize - end;
852
853                 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
854                 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
855                        CFS_PAGE_SIZE - count, cli->cl_lost_grant,
856                        cli->cl_avail_grant, cli->cl_dirty);
857         }
858
859         EXIT;
860 }
861
862 static unsigned long rpcs_in_flight(struct client_obd *cli)
863 {
864         return cli->cl_r_in_flight + cli->cl_w_in_flight;
865 }
866
867 /* caller must hold loi_list_lock */
868 void osc_wake_cache_waiters(struct client_obd *cli)
869 {
870         cfs_list_t *l, *tmp;
871         struct osc_cache_waiter *ocw;
872
873         ENTRY;
874         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
875                 /* if we can't dirty more, we must wait until some is written */
876                 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
877                    (cfs_atomic_read(&obd_dirty_pages) + 1 >
878                     obd_max_dirty_pages)) {
879                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
880                                "osc max %ld, sys max %d\n", cli->cl_dirty,
881                                cli->cl_dirty_max, obd_max_dirty_pages);
882                         return;
883                 }
884
885                 /* if still dirty cache but no grant wait for pending RPCs that
886                  * may yet return us some grant before doing sync writes */
887                 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
888                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
889                                cli->cl_w_in_flight);
890                         return;
891                 }
892
893                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
894                 cfs_list_del_init(&ocw->ocw_entry);
895                 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
896                         /* no more RPCs in flight to return grant, do sync IO */
897                         ocw->ocw_rc = -EDQUOT;
898                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
899                 } else {
900                         osc_consume_write_grant(cli,
901                                                 &ocw->ocw_oap->oap_brw_page);
902                 }
903
904                 cfs_waitq_signal(&ocw->ocw_waitq);
905         }
906
907         EXIT;
908 }
909
910 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
911 {
912         client_obd_list_lock(&cli->cl_loi_list_lock);
913         cli->cl_avail_grant += grant;
914         client_obd_list_unlock(&cli->cl_loi_list_lock);
915 }
916
917 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
918 {
919         if (body->oa.o_valid & OBD_MD_FLGRANT) {
920                 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
921                 __osc_update_grant(cli, body->oa.o_grant);
922         }
923 }
924
925 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
926                               void *key, obd_count vallen, void *val,
927                               struct ptlrpc_request_set *set);
928
929 static int osc_shrink_grant_interpret(const struct lu_env *env,
930                                       struct ptlrpc_request *req,
931                                       void *aa, int rc)
932 {
933         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
934         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
935         struct ost_body *body;
936
937         if (rc != 0) {
938                 __osc_update_grant(cli, oa->o_grant);
939                 GOTO(out, rc);
940         }
941
942         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
943         LASSERT(body);
944         osc_update_grant(cli, body);
945 out:
946         OBD_FREE_PTR(oa);
947         return rc;
948 }
949
950 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
951 {
952         client_obd_list_lock(&cli->cl_loi_list_lock);
953         oa->o_grant = cli->cl_avail_grant / 4;
954         cli->cl_avail_grant -= oa->o_grant;
955         client_obd_list_unlock(&cli->cl_loi_list_lock);
956         oa->o_flags |= OBD_FL_SHRINK_GRANT;
957         osc_update_next_shrink(cli);
958 }
959
960 /* Shrink the current grant, either from some large amount to enough for a
961  * full set of in-flight RPCs, or if we have already shrunk to that limit
962  * then to enough for a single RPC.  This avoids keeping more grant than
963  * needed, and avoids shrinking the grant piecemeal. */
964 static int osc_shrink_grant(struct client_obd *cli)
965 {
966         long target = (cli->cl_max_rpcs_in_flight + 1) *
967                       cli->cl_max_pages_per_rpc;
968
969         client_obd_list_lock(&cli->cl_loi_list_lock);
970         if (cli->cl_avail_grant <= target)
971                 target = cli->cl_max_pages_per_rpc;
972         client_obd_list_unlock(&cli->cl_loi_list_lock);
973
974         return osc_shrink_grant_to_target(cli, target);
975 }
976
977 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
978 {
979         int    rc = 0;
980         struct ost_body     *body;
981         ENTRY;
982
983         client_obd_list_lock(&cli->cl_loi_list_lock);
984         /* Don't shrink if we are already above or below the desired limit
985          * We don't want to shrink below a single RPC, as that will negatively
986          * impact block allocation and long-term performance. */
987         if (target < cli->cl_max_pages_per_rpc)
988                 target = cli->cl_max_pages_per_rpc;
989
990         if (target >= cli->cl_avail_grant) {
991                 client_obd_list_unlock(&cli->cl_loi_list_lock);
992                 RETURN(0);
993         }
994         client_obd_list_unlock(&cli->cl_loi_list_lock);
995
996         OBD_ALLOC_PTR(body);
997         if (!body)
998                 RETURN(-ENOMEM);
999
1000         osc_announce_cached(cli, &body->oa, 0);
1001
1002         client_obd_list_lock(&cli->cl_loi_list_lock);
1003         body->oa.o_grant = cli->cl_avail_grant - target;
1004         cli->cl_avail_grant = target;
1005         client_obd_list_unlock(&cli->cl_loi_list_lock);
1006         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1007         osc_update_next_shrink(cli);
1008
1009         rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1010                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1011                                 sizeof(*body), body, NULL);
1012         if (rc != 0)
1013                 __osc_update_grant(cli, body->oa.o_grant);
1014         OBD_FREE_PTR(body);
1015         RETURN(rc);
1016 }
1017
1018 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1019 static int osc_should_shrink_grant(struct client_obd *client)
1020 {
1021         cfs_time_t time = cfs_time_current();
1022         cfs_time_t next_shrink = client->cl_next_shrink_grant;
1023         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1024                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1025                     client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1026                         return 1;
1027                 else
1028                         osc_update_next_shrink(client);
1029         }
1030         return 0;
1031 }
1032
1033 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1034 {
1035         struct client_obd *client;
1036
1037         cfs_list_for_each_entry(client, &item->ti_obd_list,
1038                                 cl_grant_shrink_list) {
1039                 if (osc_should_shrink_grant(client))
1040                         osc_shrink_grant(client);
1041         }
1042         return 0;
1043 }
1044
1045 static int osc_add_shrink_grant(struct client_obd *client)
1046 {
1047         int rc;
1048
1049         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1050                                        TIMEOUT_GRANT,
1051                                        osc_grant_shrink_grant_cb, NULL,
1052                                        &client->cl_grant_shrink_list);
1053         if (rc) {
1054                 CERROR("add grant client %s error %d\n",
1055                         client->cl_import->imp_obd->obd_name, rc);
1056                 return rc;
1057         }
1058         CDEBUG(D_CACHE, "add grant client %s \n",
1059                client->cl_import->imp_obd->obd_name);
1060         osc_update_next_shrink(client);
1061         return 0;
1062 }
1063
1064 static int osc_del_shrink_grant(struct client_obd *client)
1065 {
1066         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1067                                          TIMEOUT_GRANT);
1068 }
1069
1070 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1071 {
1072         /*
1073          * ocd_grant is the total grant amount we're expect to hold: if we've
1074          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1075          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1076          *
1077          * race is tolerable here: if we're evicted, but imp_state already
1078          * left EVICTED state, then cl_dirty must be 0 already.
1079          */
1080         client_obd_list_lock(&cli->cl_loi_list_lock);
1081         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1082                 cli->cl_avail_grant = ocd->ocd_grant;
1083         else
1084                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1085         client_obd_list_unlock(&cli->cl_loi_list_lock);
1086
1087         CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1088                cli->cl_avail_grant, cli->cl_lost_grant);
1089         LASSERT(cli->cl_avail_grant >= 0);
1090
1091         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1092             cfs_list_empty(&cli->cl_grant_shrink_list))
1093                 osc_add_shrink_grant(cli);
1094 }
1095
1096 /* We assume that the reason this OSC got a short read is because it read
1097  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1098  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1099  * this stripe never got written at or beyond this stripe offset yet. */
1100 static void handle_short_read(int nob_read, obd_count page_count,
1101                               struct brw_page **pga)
1102 {
1103         char *ptr;
1104         int i = 0;
1105
1106         /* skip bytes read OK */
1107         while (nob_read > 0) {
1108                 LASSERT (page_count > 0);
1109
1110                 if (pga[i]->count > nob_read) {
1111                         /* EOF inside this page */
1112                         ptr = cfs_kmap(pga[i]->pg) +
1113                                 (pga[i]->off & ~CFS_PAGE_MASK);
1114                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1115                         cfs_kunmap(pga[i]->pg);
1116                         page_count--;
1117                         i++;
1118                         break;
1119                 }
1120
1121                 nob_read -= pga[i]->count;
1122                 page_count--;
1123                 i++;
1124         }
1125
1126         /* zero remaining pages */
1127         while (page_count-- > 0) {
1128                 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1129                 memset(ptr, 0, pga[i]->count);
1130                 cfs_kunmap(pga[i]->pg);
1131                 i++;
1132         }
1133 }
1134
1135 static int check_write_rcs(struct ptlrpc_request *req,
1136                            int requested_nob, int niocount,
1137                            obd_count page_count, struct brw_page **pga)
1138 {
1139         int     i;
1140         __u32   *remote_rcs;
1141
1142         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1143                                                   sizeof(*remote_rcs) *
1144                                                   niocount);
1145         if (remote_rcs == NULL) {
1146                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1147                 return(-EPROTO);
1148         }
1149
1150         /* return error if any niobuf was in error */
1151         for (i = 0; i < niocount; i++) {
1152                 if (remote_rcs[i] < 0)
1153                         return(remote_rcs[i]);
1154
1155                 if (remote_rcs[i] != 0) {
1156                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1157                                 i, remote_rcs[i], req);
1158                         return(-EPROTO);
1159                 }
1160         }
1161
1162         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1163                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1164                        req->rq_bulk->bd_nob_transferred, requested_nob);
1165                 return(-EPROTO);
1166         }
1167
1168         return (0);
1169 }
1170
1171 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1172 {
1173         if (p1->flag != p2->flag) {
1174                 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1175                                   OBD_BRW_NOCACHE|OBD_BRW_SYNC);
1176
1177                 /* warn if we try to combine flags that we don't know to be
1178                  * safe to combine */
1179                 if ((p1->flag & mask) != (p2->flag & mask))
1180                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
1181                                "same brw?\n", p1->flag, p2->flag);
1182                 return 0;
1183         }
1184
1185         return (p1->off + p1->count == p2->off);
1186 }
1187
1188 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1189                                    struct brw_page **pga, int opc,
1190                                    cksum_type_t cksum_type)
1191 {
1192         __u32 cksum;
1193         int i = 0;
1194
1195         LASSERT (pg_count > 0);
1196         cksum = init_checksum(cksum_type);
1197         while (nob > 0 && pg_count > 0) {
1198                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1199                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1200                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1201
1202                 /* corrupt the data before we compute the checksum, to
1203                  * simulate an OST->client data error */
1204                 if (i == 0 && opc == OST_READ &&
1205                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1206                         memcpy(ptr + off, "bad1", min(4, nob));
1207                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1208                 cfs_kunmap(pga[i]->pg);
1209                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1210                                off, cksum);
1211
1212                 nob -= pga[i]->count;
1213                 pg_count--;
1214                 i++;
1215         }
1216         /* For sending we only compute the wrong checksum instead
1217          * of corrupting the data so it is still correct on a redo */
1218         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1219                 cksum++;
1220
1221         return cksum;
1222 }
1223
1224 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1225                                 struct lov_stripe_md *lsm, obd_count page_count,
1226                                 struct brw_page **pga,
1227                                 struct ptlrpc_request **reqp,
1228                                 struct obd_capa *ocapa, int reserve)
1229 {
1230         struct ptlrpc_request   *req;
1231         struct ptlrpc_bulk_desc *desc;
1232         struct ost_body         *body;
1233         struct obd_ioobj        *ioobj;
1234         struct niobuf_remote    *niobuf;
1235         int niocount, i, requested_nob, opc, rc;
1236         struct osc_brw_async_args *aa;
1237         struct req_capsule      *pill;
1238         struct brw_page *pg_prev;
1239
1240         ENTRY;
1241         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1242                 RETURN(-ENOMEM); /* Recoverable */
1243         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1244                 RETURN(-EINVAL); /* Fatal */
1245
1246         if ((cmd & OBD_BRW_WRITE) != 0) {
1247                 opc = OST_WRITE;
1248                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1249                                                 cli->cl_import->imp_rq_pool,
1250                                                 &RQF_OST_BRW_WRITE);
1251         } else {
1252                 opc = OST_READ;
1253                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1254         }
1255         if (req == NULL)
1256                 RETURN(-ENOMEM);
1257
1258         for (niocount = i = 1; i < page_count; i++) {
1259                 if (!can_merge_pages(pga[i - 1], pga[i]))
1260                         niocount++;
1261         }
1262
1263         pill = &req->rq_pill;
1264         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1265                              sizeof(*ioobj));
1266         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1267                              niocount * sizeof(*niobuf));
1268         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1269
1270         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1271         if (rc) {
1272                 ptlrpc_request_free(req);
1273                 RETURN(rc);
1274         }
1275         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1276         ptlrpc_at_set_req_timeout(req);
1277
1278         if (opc == OST_WRITE)
1279                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1280                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1281         else
1282                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1283                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1284
1285         if (desc == NULL)
1286                 GOTO(out, rc = -ENOMEM);
1287         /* NB request now owns desc and will free it when it gets freed */
1288
1289         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1290         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1291         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1292         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1293
1294         lustre_set_wire_obdo(&body->oa, oa);
1295
1296         obdo_to_ioobj(oa, ioobj);
1297         ioobj->ioo_bufcnt = niocount;
1298         osc_pack_capa(req, body, ocapa);
1299         LASSERT (page_count > 0);
1300         pg_prev = pga[0];
1301         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1302                 struct brw_page *pg = pga[i];
1303
1304                 LASSERT(pg->count > 0);
1305                 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1306                          "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1307                          pg->off, pg->count);
1308 #ifdef __linux__
1309                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1310                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1311                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1312                          i, page_count,
1313                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1314                          pg_prev->pg, page_private(pg_prev->pg),
1315                          pg_prev->pg->index, pg_prev->off);
1316 #else
1317                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1318                          "i %d p_c %u\n", i, page_count);
1319 #endif
1320                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1321                         (pg->flag & OBD_BRW_SRVLOCK));
1322
1323                 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1324                                       pg->count);
1325                 requested_nob += pg->count;
1326
1327                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1328                         niobuf--;
1329                         niobuf->len += pg->count;
1330                 } else {
1331                         niobuf->offset = pg->off;
1332                         niobuf->len    = pg->count;
1333                         niobuf->flags  = pg->flag;
1334                 }
1335                 pg_prev = pg;
1336         }
1337
1338         LASSERTF((void *)(niobuf - niocount) ==
1339                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1340                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1341                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1342
1343         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1344         if (osc_should_shrink_grant(cli))
1345                 osc_shrink_grant_local(cli, &body->oa);
1346
1347         /* size[REQ_REC_OFF] still sizeof (*body) */
1348         if (opc == OST_WRITE) {
1349                 if (unlikely(cli->cl_checksum) &&
1350                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1351                         /* store cl_cksum_type in a local variable since
1352                          * it can be changed via lprocfs */
1353                         cksum_type_t cksum_type = cli->cl_cksum_type;
1354
1355                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1356                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1357                                 body->oa.o_flags = 0;
1358                         }
1359                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1360                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1361                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1362                                                              page_count, pga,
1363                                                              OST_WRITE,
1364                                                              cksum_type);
1365                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1366                                body->oa.o_cksum);
1367                         /* save this in 'oa', too, for later checking */
1368                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1369                         oa->o_flags |= cksum_type_pack(cksum_type);
1370                 } else {
1371                         /* clear out the checksum flag, in case this is a
1372                          * resend but cl_checksum is no longer set. b=11238 */
1373                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1374                 }
1375                 oa->o_cksum = body->oa.o_cksum;
1376                 /* 1 RC per niobuf */
1377                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1378                                      sizeof(__u32) * niocount);
1379         } else {
1380                 if (unlikely(cli->cl_checksum) &&
1381                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1382                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1383                                 body->oa.o_flags = 0;
1384                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1385                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1386                 }
1387         }
1388         ptlrpc_request_set_replen(req);
1389
1390         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1391         aa = ptlrpc_req_async_args(req);
1392         aa->aa_oa = oa;
1393         aa->aa_requested_nob = requested_nob;
1394         aa->aa_nio_count = niocount;
1395         aa->aa_page_count = page_count;
1396         aa->aa_resends = 0;
1397         aa->aa_ppga = pga;
1398         aa->aa_cli = cli;
1399         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1400         if (ocapa && reserve)
1401                 aa->aa_ocapa = capa_get(ocapa);
1402
1403         *reqp = req;
1404         RETURN(0);
1405
1406  out:
1407         ptlrpc_req_finished(req);
1408         RETURN(rc);
1409 }
1410
1411 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1412                                 __u32 client_cksum, __u32 server_cksum, int nob,
1413                                 obd_count page_count, struct brw_page **pga,
1414                                 cksum_type_t client_cksum_type)
1415 {
1416         __u32 new_cksum;
1417         char *msg;
1418         cksum_type_t cksum_type;
1419
1420         if (server_cksum == client_cksum) {
1421                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1422                 return 0;
1423         }
1424
1425         if (oa->o_valid & OBD_MD_FLFLAGS)
1426                 cksum_type = cksum_type_unpack(oa->o_flags);
1427         else
1428                 cksum_type = OBD_CKSUM_CRC32;
1429
1430         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1431                                       cksum_type);
1432
1433         if (cksum_type != client_cksum_type)
1434                 msg = "the server did not use the checksum type specified in "
1435                       "the original request - likely a protocol problem";
1436         else if (new_cksum == server_cksum)
1437                 msg = "changed on the client after we checksummed it - "
1438                       "likely false positive due to mmap IO (bug 11742)";
1439         else if (new_cksum == client_cksum)
1440                 msg = "changed in transit before arrival at OST";
1441         else
1442                 msg = "changed in transit AND doesn't match the original - "
1443                       "likely false positive due to mmap IO (bug 11742)";
1444
1445         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1446                            " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1447                            msg, libcfs_nid2str(peer->nid),
1448                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1449                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1450                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1451                            oa->o_id,
1452                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1453                            pga[0]->off,
1454                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1455         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1456                "client csum now %x\n", client_cksum, client_cksum_type,
1457                server_cksum, cksum_type, new_cksum);
1458         return 1;
1459 }
1460
1461 /* Note rc enters this function as number of bytes transferred */
1462 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1463 {
1464         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1465         const lnet_process_id_t *peer =
1466                         &req->rq_import->imp_connection->c_peer;
1467         struct client_obd *cli = aa->aa_cli;
1468         struct ost_body *body;
1469         __u32 client_cksum = 0;
1470         ENTRY;
1471
1472         if (rc < 0 && rc != -EDQUOT) {
1473                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1474                 RETURN(rc);
1475         }
1476
1477         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1478         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1479         if (body == NULL) {
1480                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1481                 RETURN(-EPROTO);
1482         }
1483
1484 #ifdef HAVE_QUOTA_SUPPORT
1485         /* set/clear over quota flag for a uid/gid */
1486         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1487             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1488                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1489
1490                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1491                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1492                        body->oa.o_flags);
1493                 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1494                              body->oa.o_flags);
1495         }
1496 #endif
1497
1498         if (rc < 0)
1499                 RETURN(rc);
1500
1501         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1502                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1503
1504         osc_update_grant(cli, body);
1505
1506         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1507                 if (rc > 0) {
1508                         CERROR("Unexpected +ve rc %d\n", rc);
1509                         RETURN(-EPROTO);
1510                 }
1511                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1512
1513                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1514                         RETURN(-EAGAIN);
1515
1516                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1517                     check_write_checksum(&body->oa, peer, client_cksum,
1518                                          body->oa.o_cksum, aa->aa_requested_nob,
1519                                          aa->aa_page_count, aa->aa_ppga,
1520                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1521                         RETURN(-EAGAIN);
1522
1523                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1524                                      aa->aa_page_count, aa->aa_ppga);
1525                 GOTO(out, rc);
1526         }
1527
1528         /* The rest of this function executes only for OST_READs */
1529
1530         /* if unwrap_bulk failed, return -EAGAIN to retry */
1531         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1532         if (rc < 0)
1533                 GOTO(out, rc = -EAGAIN);
1534
1535         if (rc > aa->aa_requested_nob) {
1536                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1537                        aa->aa_requested_nob);
1538                 RETURN(-EPROTO);
1539         }
1540
1541         if (rc != req->rq_bulk->bd_nob_transferred) {
1542                 CERROR ("Unexpected rc %d (%d transferred)\n",
1543                         rc, req->rq_bulk->bd_nob_transferred);
1544                 return (-EPROTO);
1545         }
1546
1547         if (rc < aa->aa_requested_nob)
1548                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1549
1550         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1551                 static int cksum_counter;
1552                 __u32      server_cksum = body->oa.o_cksum;
1553                 char      *via;
1554                 char      *router;
1555                 cksum_type_t cksum_type;
1556
1557                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1558                         cksum_type = cksum_type_unpack(body->oa.o_flags);
1559                 else
1560                         cksum_type = OBD_CKSUM_CRC32;
1561                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1562                                                  aa->aa_ppga, OST_READ,
1563                                                  cksum_type);
1564
1565                 if (peer->nid == req->rq_bulk->bd_sender) {
1566                         via = router = "";
1567                 } else {
1568                         via = " via ";
1569                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1570                 }
1571
1572                 if (server_cksum == ~0 && rc > 0) {
1573                         CERROR("Protocol error: server %s set the 'checksum' "
1574                                "bit, but didn't send a checksum.  Not fatal, "
1575                                "but please notify on http://bugzilla.lustre.org/\n",
1576                                libcfs_nid2str(peer->nid));
1577                 } else if (server_cksum != client_cksum) {
1578                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1579                                            "%s%s%s inode "DFID" object "
1580                                            LPU64"/"LPU64" extent "
1581                                            "["LPU64"-"LPU64"]\n",
1582                                            req->rq_import->imp_obd->obd_name,
1583                                            libcfs_nid2str(peer->nid),
1584                                            via, router,
1585                                            body->oa.o_valid & OBD_MD_FLFID ?
1586                                                 body->oa.o_parent_seq : (__u64)0,
1587                                            body->oa.o_valid & OBD_MD_FLFID ?
1588                                                 body->oa.o_parent_oid : 0,
1589                                            body->oa.o_valid & OBD_MD_FLFID ?
1590                                                 body->oa.o_parent_ver : 0,
1591                                            body->oa.o_id,
1592                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1593                                                 body->oa.o_seq : (__u64)0,
1594                                            aa->aa_ppga[0]->off,
1595                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1596                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1597                                                                         1);
1598                         CERROR("client %x, server %x, cksum_type %x\n",
1599                                client_cksum, server_cksum, cksum_type);
1600                         cksum_counter = 0;
1601                         aa->aa_oa->o_cksum = client_cksum;
1602                         rc = -EAGAIN;
1603                 } else {
1604                         cksum_counter++;
1605                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1606                         rc = 0;
1607                 }
1608         } else if (unlikely(client_cksum)) {
1609                 static int cksum_missed;
1610
1611                 cksum_missed++;
1612                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1613                         CERROR("Checksum %u requested from %s but not sent\n",
1614                                cksum_missed, libcfs_nid2str(peer->nid));
1615         } else {
1616                 rc = 0;
1617         }
1618 out:
1619         if (rc >= 0)
1620                 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1621
1622         RETURN(rc);
1623 }
1624
1625 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1626                             struct lov_stripe_md *lsm,
1627                             obd_count page_count, struct brw_page **pga,
1628                             struct obd_capa *ocapa)
1629 {
1630         struct ptlrpc_request *req;
1631         int                    rc;
1632         cfs_waitq_t            waitq;
1633         int                    resends = 0;
1634         struct l_wait_info     lwi;
1635
1636         ENTRY;
1637
1638         cfs_waitq_init(&waitq);
1639
1640 restart_bulk:
1641         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1642                                   page_count, pga, &req, ocapa, 0);
1643         if (rc != 0)
1644                 return (rc);
1645
1646         rc = ptlrpc_queue_wait(req);
1647
1648         if (rc == -ETIMEDOUT && req->rq_resend) {
1649                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1650                 ptlrpc_req_finished(req);
1651                 goto restart_bulk;
1652         }
1653
1654         rc = osc_brw_fini_request(req, rc);
1655
1656         ptlrpc_req_finished(req);
1657         if (osc_recoverable_error(rc)) {
1658                 resends++;
1659                 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1660                         CERROR("too many resend retries, returning error\n");
1661                         RETURN(-EIO);
1662                 }
1663
1664                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1665                 l_wait_event(waitq, 0, &lwi);
1666
1667                 goto restart_bulk;
1668         }
1669
1670         RETURN (rc);
1671 }
1672
1673 int osc_brw_redo_request(struct ptlrpc_request *request,
1674                          struct osc_brw_async_args *aa)
1675 {
1676         struct ptlrpc_request *new_req;
1677         struct ptlrpc_request_set *set = request->rq_set;
1678         struct osc_brw_async_args *new_aa;
1679         struct osc_async_page *oap;
1680         int rc = 0;
1681         ENTRY;
1682
1683         if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1684                 CERROR("too many resend retries, returning error\n");
1685                 RETURN(-EIO);
1686         }
1687
1688         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1689
1690         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1691                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1692                                   aa->aa_cli, aa->aa_oa,
1693                                   NULL /* lsm unused by osc currently */,
1694                                   aa->aa_page_count, aa->aa_ppga,
1695                                   &new_req, aa->aa_ocapa, 0);
1696         if (rc)
1697                 RETURN(rc);
1698
1699         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1700
1701         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1702                 if (oap->oap_request != NULL) {
1703                         LASSERTF(request == oap->oap_request,
1704                                  "request %p != oap_request %p\n",
1705                                  request, oap->oap_request);
1706                         if (oap->oap_interrupted) {
1707                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1708                                 ptlrpc_req_finished(new_req);
1709                                 RETURN(-EINTR);
1710                         }
1711                 }
1712         }
1713         /* New request takes over pga and oaps from old request.
1714          * Note that copying a list_head doesn't work, need to move it... */
1715         aa->aa_resends++;
1716         new_req->rq_interpret_reply = request->rq_interpret_reply;
1717         new_req->rq_async_args = request->rq_async_args;
1718         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1719
1720         new_aa = ptlrpc_req_async_args(new_req);
1721
1722         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1723         cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1724         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1725
1726         cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1727                 if (oap->oap_request) {
1728                         ptlrpc_req_finished(oap->oap_request);
1729                         oap->oap_request = ptlrpc_request_addref(new_req);
1730                 }
1731         }
1732
1733         new_aa->aa_ocapa = aa->aa_ocapa;
1734         aa->aa_ocapa = NULL;
1735
1736         /* use ptlrpc_set_add_req is safe because interpret functions work
1737          * in check_set context. only one way exist with access to request
1738          * from different thread got -EINTR - this way protected with
1739          * cl_loi_list_lock */
1740         ptlrpc_set_add_req(set, new_req);
1741
1742         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1743
1744         DEBUG_REQ(D_INFO, new_req, "new request");
1745         RETURN(0);
1746 }
1747
1748 /*
1749  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1750  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1751  * fine for our small page arrays and doesn't require allocation.  its an
1752  * insertion sort that swaps elements that are strides apart, shrinking the
1753  * stride down until its '1' and the array is sorted.
1754  */
1755 static void sort_brw_pages(struct brw_page **array, int num)
1756 {
1757         int stride, i, j;
1758         struct brw_page *tmp;
1759
1760         if (num == 1)
1761                 return;
1762         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1763                 ;
1764
1765         do {
1766                 stride /= 3;
1767                 for (i = stride ; i < num ; i++) {
1768                         tmp = array[i];
1769                         j = i;
1770                         while (j >= stride && array[j - stride]->off > tmp->off) {
1771                                 array[j] = array[j - stride];
1772                                 j -= stride;
1773                         }
1774                         array[j] = tmp;
1775                 }
1776         } while (stride > 1);
1777 }
1778
1779 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1780 {
1781         int count = 1;
1782         int offset;
1783         int i = 0;
1784
1785         LASSERT (pages > 0);
1786         offset = pg[i]->off & ~CFS_PAGE_MASK;
1787
1788         for (;;) {
1789                 pages--;
1790                 if (pages == 0)         /* that's all */
1791                         return count;
1792
1793                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1794                         return count;   /* doesn't end on page boundary */
1795
1796                 i++;
1797                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1798                 if (offset != 0)        /* doesn't start on page boundary */
1799                         return count;
1800
1801                 count++;
1802         }
1803 }
1804
1805 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1806 {
1807         struct brw_page **ppga;
1808         int i;
1809
1810         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1811         if (ppga == NULL)
1812                 return NULL;
1813
1814         for (i = 0; i < count; i++)
1815                 ppga[i] = pga + i;
1816         return ppga;
1817 }
1818
1819 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1820 {
1821         LASSERT(ppga != NULL);
1822         OBD_FREE(ppga, sizeof(*ppga) * count);
1823 }
1824
1825 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1826                    obd_count page_count, struct brw_page *pga,
1827                    struct obd_trans_info *oti)
1828 {
1829         struct obdo *saved_oa = NULL;
1830         struct brw_page **ppga, **orig;
1831         struct obd_import *imp = class_exp2cliimp(exp);
1832         struct client_obd *cli;
1833         int rc, page_count_orig;
1834         ENTRY;
1835
1836         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1837         cli = &imp->imp_obd->u.cli;
1838
1839         if (cmd & OBD_BRW_CHECK) {
1840                 /* The caller just wants to know if there's a chance that this
1841                  * I/O can succeed */
1842
1843                 if (imp->imp_invalid)
1844                         RETURN(-EIO);
1845                 RETURN(0);
1846         }
1847
1848         /* test_brw with a failed create can trip this, maybe others. */
1849         LASSERT(cli->cl_max_pages_per_rpc);
1850
1851         rc = 0;
1852
1853         orig = ppga = osc_build_ppga(pga, page_count);
1854         if (ppga == NULL)
1855                 RETURN(-ENOMEM);
1856         page_count_orig = page_count;
1857
1858         sort_brw_pages(ppga, page_count);
1859         while (page_count) {
1860                 obd_count pages_per_brw;
1861
1862                 if (page_count > cli->cl_max_pages_per_rpc)
1863                         pages_per_brw = cli->cl_max_pages_per_rpc;
1864                 else
1865                         pages_per_brw = page_count;
1866
1867                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1868
1869                 if (saved_oa != NULL) {
1870                         /* restore previously saved oa */
1871                         *oinfo->oi_oa = *saved_oa;
1872                 } else if (page_count > pages_per_brw) {
1873                         /* save a copy of oa (brw will clobber it) */
1874                         OBDO_ALLOC(saved_oa);
1875                         if (saved_oa == NULL)
1876                                 GOTO(out, rc = -ENOMEM);
1877                         *saved_oa = *oinfo->oi_oa;
1878                 }
1879
1880                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1881                                       pages_per_brw, ppga, oinfo->oi_capa);
1882
1883                 if (rc != 0)
1884                         break;
1885
1886                 page_count -= pages_per_brw;
1887                 ppga += pages_per_brw;
1888         }
1889
1890 out:
1891         osc_release_ppga(orig, page_count_orig);
1892
1893         if (saved_oa != NULL)
1894                 OBDO_FREE(saved_oa);
1895
1896         RETURN(rc);
1897 }
1898
1899 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1900  * the dirty accounting.  Writeback completes or truncate happens before
1901  * writing starts.  Must be called with the loi lock held. */
1902 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1903                            int sent)
1904 {
1905         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1906 }
1907
1908
1909 /* This maintains the lists of pending pages to read/write for a given object
1910  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1911  * to quickly find objects that are ready to send an RPC. */
1912 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1913                          int cmd)
1914 {
1915         int optimal;
1916         ENTRY;
1917
1918         if (lop->lop_num_pending == 0)
1919                 RETURN(0);
1920
1921         /* if we have an invalid import we want to drain the queued pages
1922          * by forcing them through rpcs that immediately fail and complete
1923          * the pages.  recovery relies on this to empty the queued pages
1924          * before canceling the locks and evicting down the llite pages */
1925         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1926                 RETURN(1);
1927
1928         /* stream rpcs in queue order as long as as there is an urgent page
1929          * queued.  this is our cheap solution for good batching in the case
1930          * where writepage marks some random page in the middle of the file
1931          * as urgent because of, say, memory pressure */
1932         if (!cfs_list_empty(&lop->lop_urgent)) {
1933                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1934                 RETURN(1);
1935         }
1936         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1937         optimal = cli->cl_max_pages_per_rpc;
1938         if (cmd & OBD_BRW_WRITE) {
1939                 /* trigger a write rpc stream as long as there are dirtiers
1940                  * waiting for space.  as they're waiting, they're not going to
1941                  * create more pages to coallesce with what's waiting.. */
1942                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1943                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1944                         RETURN(1);
1945                 }
1946                 /* +16 to avoid triggering rpcs that would want to include pages
1947                  * that are being queued but which can't be made ready until
1948                  * the queuer finishes with the page. this is a wart for
1949                  * llite::commit_write() */
1950                 optimal += 16;
1951         }
1952         if (lop->lop_num_pending >= optimal)
1953                 RETURN(1);
1954
1955         RETURN(0);
1956 }
1957
1958 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1959 {
1960         struct osc_async_page *oap;
1961         ENTRY;
1962
1963         if (cfs_list_empty(&lop->lop_urgent))
1964                 RETURN(0);
1965
1966         oap = cfs_list_entry(lop->lop_urgent.next,
1967                          struct osc_async_page, oap_urgent_item);
1968
1969         if (oap->oap_async_flags & ASYNC_HP) {
1970                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1971                 RETURN(1);
1972         }
1973
1974         RETURN(0);
1975 }
1976
1977 static void on_list(cfs_list_t *item, cfs_list_t *list,
1978                     int should_be_on)
1979 {
1980         if (cfs_list_empty(item) && should_be_on)
1981                 cfs_list_add_tail(item, list);
1982         else if (!cfs_list_empty(item) && !should_be_on)
1983                 cfs_list_del_init(item);
1984 }
1985
1986 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1987  * can find pages to build into rpcs quickly */
1988 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1989 {
1990         if (lop_makes_hprpc(&loi->loi_write_lop) ||
1991             lop_makes_hprpc(&loi->loi_read_lop)) {
1992                 /* HP rpc */
1993                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1994                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1995         } else {
1996                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1997                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
1998                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
1999                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2000         }
2001
2002         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2003                 loi->loi_write_lop.lop_num_pending);
2004
2005         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2006                 loi->loi_read_lop.lop_num_pending);
2007 }
2008
2009 static void lop_update_pending(struct client_obd *cli,
2010                                struct loi_oap_pages *lop, int cmd, int delta)
2011 {
2012         lop->lop_num_pending += delta;
2013         if (cmd & OBD_BRW_WRITE)
2014                 cli->cl_pending_w_pages += delta;
2015         else
2016                 cli->cl_pending_r_pages += delta;
2017 }
2018
2019 /**
2020  * this is called when a sync waiter receives an interruption.  Its job is to
2021  * get the caller woken as soon as possible.  If its page hasn't been put in an
2022  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2023  * desiring interruption which will forcefully complete the rpc once the rpc
2024  * has timed out.
2025  */
2026 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2027 {
2028         struct loi_oap_pages *lop;
2029         struct lov_oinfo *loi;
2030         int rc = -EBUSY;
2031         ENTRY;
2032
2033         LASSERT(!oap->oap_interrupted);
2034         oap->oap_interrupted = 1;
2035
2036         /* ok, it's been put in an rpc. only one oap gets a request reference */
2037         if (oap->oap_request != NULL) {
2038                 ptlrpc_mark_interrupted(oap->oap_request);
2039                 ptlrpcd_wake(oap->oap_request);
2040                 ptlrpc_req_finished(oap->oap_request);
2041                 oap->oap_request = NULL;
2042         }
2043
2044         /*
2045          * page completion may be called only if ->cpo_prep() method was
2046          * executed by osc_io_submit(), that also adds page the to pending list
2047          */
2048         if (!cfs_list_empty(&oap->oap_pending_item)) {
2049                 cfs_list_del_init(&oap->oap_pending_item);
2050                 cfs_list_del_init(&oap->oap_urgent_item);
2051
2052                 loi = oap->oap_loi;
2053                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2054                         &loi->loi_write_lop : &loi->loi_read_lop;
2055                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2056                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2057                 rc = oap->oap_caller_ops->ap_completion(env,
2058                                           oap->oap_caller_data,
2059                                           oap->oap_cmd, NULL, -EINTR);
2060         }
2061
2062         RETURN(rc);
2063 }
2064
2065 /* this is trying to propogate async writeback errors back up to the
2066  * application.  As an async write fails we record the error code for later if
2067  * the app does an fsync.  As long as errors persist we force future rpcs to be
2068  * sync so that the app can get a sync error and break the cycle of queueing
2069  * pages for which writeback will fail. */
2070 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2071                            int rc)
2072 {
2073         if (rc) {
2074                 if (!ar->ar_rc)
2075                         ar->ar_rc = rc;
2076
2077                 ar->ar_force_sync = 1;
2078                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2079                 return;
2080
2081         }
2082
2083         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2084                 ar->ar_force_sync = 0;
2085 }
2086
2087 void osc_oap_to_pending(struct osc_async_page *oap)
2088 {
2089         struct loi_oap_pages *lop;
2090
2091         if (oap->oap_cmd & OBD_BRW_WRITE)
2092                 lop = &oap->oap_loi->loi_write_lop;
2093         else
2094                 lop = &oap->oap_loi->loi_read_lop;
2095
2096         if (oap->oap_async_flags & ASYNC_HP)
2097                 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2098         else if (oap->oap_async_flags & ASYNC_URGENT)
2099                 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2100         cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2101         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2102 }
2103
2104 /* this must be called holding the loi list lock to give coverage to exit_cache,
2105  * async_flag maintenance, and oap_request */
2106 static void osc_ap_completion(const struct lu_env *env,
2107                               struct client_obd *cli, struct obdo *oa,
2108                               struct osc_async_page *oap, int sent, int rc)
2109 {
2110         __u64 xid = 0;
2111
2112         ENTRY;
2113         if (oap->oap_request != NULL) {
2114                 xid = ptlrpc_req_xid(oap->oap_request);
2115                 ptlrpc_req_finished(oap->oap_request);
2116                 oap->oap_request = NULL;
2117         }
2118
2119         cfs_spin_lock(&oap->oap_lock);
2120         oap->oap_async_flags = 0;
2121         cfs_spin_unlock(&oap->oap_lock);
2122         oap->oap_interrupted = 0;
2123
2124         if (oap->oap_cmd & OBD_BRW_WRITE) {
2125                 osc_process_ar(&cli->cl_ar, xid, rc);
2126                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2127         }
2128
2129         if (rc == 0 && oa != NULL) {
2130                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2131                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2132                 if (oa->o_valid & OBD_MD_FLMTIME)
2133                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2134                 if (oa->o_valid & OBD_MD_FLATIME)
2135                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2136                 if (oa->o_valid & OBD_MD_FLCTIME)
2137                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2138         }
2139
2140         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2141                                                 oap->oap_cmd, oa, rc);
2142
2143         /* ll_ap_completion (from llite) drops PG_locked. so, a new
2144          * I/O on the page could start, but OSC calls it under lock
2145          * and thus we can add oap back to pending safely */
2146         if (rc)
2147                 /* upper layer wants to leave the page on pending queue */
2148                 osc_oap_to_pending(oap);
2149         else
2150                 osc_exit_cache(cli, oap, sent);
2151         EXIT;
2152 }
2153
2154 static int brw_interpret(const struct lu_env *env,
2155                          struct ptlrpc_request *req, void *data, int rc)
2156 {
2157         struct osc_brw_async_args *aa = data;
2158         struct client_obd *cli;
2159         int async;
2160         ENTRY;
2161
2162         rc = osc_brw_fini_request(req, rc);
2163         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2164         if (osc_recoverable_error(rc)) {
2165                 rc = osc_brw_redo_request(req, aa);
2166                 if (rc == 0)
2167                         RETURN(0);
2168         }
2169
2170         if (aa->aa_ocapa) {
2171                 capa_put(aa->aa_ocapa);
2172                 aa->aa_ocapa = NULL;
2173         }
2174
2175         cli = aa->aa_cli;
2176
2177         client_obd_list_lock(&cli->cl_loi_list_lock);
2178
2179         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2180          * is called so we know whether to go to sync BRWs or wait for more
2181          * RPCs to complete */
2182         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2183                 cli->cl_w_in_flight--;
2184         else
2185                 cli->cl_r_in_flight--;
2186
2187         async = cfs_list_empty(&aa->aa_oaps);
2188         if (!async) { /* from osc_send_oap_rpc() */
2189                 struct osc_async_page *oap, *tmp;
2190                 /* the caller may re-use the oap after the completion call so
2191                  * we need to clean it up a little */
2192                 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2193                                              oap_rpc_item) {
2194                         cfs_list_del_init(&oap->oap_rpc_item);
2195                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2196                 }
2197                 OBDO_FREE(aa->aa_oa);
2198         } else { /* from async_internal() */
2199                 int i;
2200                 for (i = 0; i < aa->aa_page_count; i++)
2201                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2202
2203                 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2204                         OBDO_FREE(aa->aa_oa);
2205         }
2206         osc_wake_cache_waiters(cli);
2207         osc_check_rpcs(env, cli);
2208         client_obd_list_unlock(&cli->cl_loi_list_lock);
2209         if (!async)
2210                 cl_req_completion(env, aa->aa_clerq, rc);
2211         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2212         RETURN(rc);
2213 }
2214
2215 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2216                                             struct client_obd *cli,
2217                                             cfs_list_t *rpc_list,
2218                                             int page_count, int cmd)
2219 {
2220         struct ptlrpc_request *req;
2221         struct brw_page **pga = NULL;
2222         struct osc_brw_async_args *aa;
2223         struct obdo *oa = NULL;
2224         const struct obd_async_page_ops *ops = NULL;
2225         void *caller_data = NULL;
2226         struct osc_async_page *oap;
2227         struct osc_async_page *tmp;
2228         struct ost_body *body;
2229         struct cl_req *clerq = NULL;
2230         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2231         struct ldlm_lock *lock = NULL;
2232         struct cl_req_attr crattr;
2233         int i, rc;
2234
2235         ENTRY;
2236         LASSERT(!cfs_list_empty(rpc_list));
2237
2238         memset(&crattr, 0, sizeof crattr);
2239         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2240         if (pga == NULL)
2241                 GOTO(out, req = ERR_PTR(-ENOMEM));
2242
2243         OBDO_ALLOC(oa);
2244         if (oa == NULL)
2245                 GOTO(out, req = ERR_PTR(-ENOMEM));
2246
2247         i = 0;
2248         cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2249                 struct cl_page *page = osc_oap2cl_page(oap);
2250                 if (ops == NULL) {
2251                         ops = oap->oap_caller_ops;
2252                         caller_data = oap->oap_caller_data;
2253
2254                         clerq = cl_req_alloc(env, page, crt,
2255                                              1 /* only 1-object rpcs for
2256                                                 * now */);
2257                         if (IS_ERR(clerq))
2258                                 GOTO(out, req = (void *)clerq);
2259                         lock = oap->oap_ldlm_lock;
2260                 }
2261                 pga[i] = &oap->oap_brw_page;
2262                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2263                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2264                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2265                 i++;
2266                 cl_req_page_add(env, clerq, page);
2267         }
2268
2269         /* always get the data for the obdo for the rpc */
2270         LASSERT(ops != NULL);
2271         crattr.cra_oa = oa;
2272         crattr.cra_capa = NULL;
2273         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2274         if (lock) {
2275                 oa->o_handle = lock->l_remote_handle;
2276                 oa->o_valid |= OBD_MD_FLHANDLE;
2277         }
2278
2279         rc = cl_req_prep(env, clerq);
2280         if (rc != 0) {
2281                 CERROR("cl_req_prep failed: %d\n", rc);
2282                 GOTO(out, req = ERR_PTR(rc));
2283         }
2284
2285         sort_brw_pages(pga, page_count);
2286         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2287                                   pga, &req, crattr.cra_capa, 1);
2288         if (rc != 0) {
2289                 CERROR("prep_req failed: %d\n", rc);
2290                 GOTO(out, req = ERR_PTR(rc));
2291         }
2292
2293         /* Need to update the timestamps after the request is built in case
2294          * we race with setattr (locally or in queue at OST).  If OST gets
2295          * later setattr before earlier BRW (as determined by the request xid),
2296          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2297          * way to do this in a single call.  bug 10150 */
2298         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2299         cl_req_attr_set(env, clerq, &crattr,
2300                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2301
2302         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2303         aa = ptlrpc_req_async_args(req);
2304         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2305         cfs_list_splice(rpc_list, &aa->aa_oaps);
2306         CFS_INIT_LIST_HEAD(rpc_list);
2307         aa->aa_clerq = clerq;
2308 out:
2309         capa_put(crattr.cra_capa);
2310         if (IS_ERR(req)) {
2311                 if (oa)
2312                         OBDO_FREE(oa);
2313                 if (pga)
2314                         OBD_FREE(pga, sizeof(*pga) * page_count);
2315                 /* this should happen rarely and is pretty bad, it makes the
2316                  * pending list not follow the dirty order */
2317                 client_obd_list_lock(&cli->cl_loi_list_lock);
2318                 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2319                         cfs_list_del_init(&oap->oap_rpc_item);
2320
2321                         /* queued sync pages can be torn down while the pages
2322                          * were between the pending list and the rpc */
2323                         if (oap->oap_interrupted) {
2324                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2325                                 osc_ap_completion(env, cli, NULL, oap, 0,
2326                                                   oap->oap_count);
2327                                 continue;
2328                         }
2329                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2330                 }
2331                 if (clerq && !IS_ERR(clerq))
2332                         cl_req_completion(env, clerq, PTR_ERR(req));
2333         }
2334         RETURN(req);
2335 }
2336
2337 /**
2338  * prepare pages for ASYNC io and put pages in send queue.
2339  *
2340  * \param cmd OBD_BRW_* macroses
2341  * \param lop pending pages
2342  *
2343  * \return zero if pages successfully add to send queue.
2344  * \return not zere if error occurring.
2345  */
2346 static int
2347 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2348                  struct lov_oinfo *loi,
2349                  int cmd, struct loi_oap_pages *lop)
2350 {
2351         struct ptlrpc_request *req;
2352         obd_count page_count = 0;
2353         struct osc_async_page *oap = NULL, *tmp;
2354         struct osc_brw_async_args *aa;
2355         const struct obd_async_page_ops *ops;
2356         CFS_LIST_HEAD(rpc_list);
2357         CFS_LIST_HEAD(tmp_list);
2358         unsigned int ending_offset;
2359         unsigned  starting_offset = 0;
2360         int srvlock = 0;
2361         struct cl_object *clob = NULL;
2362         ENTRY;
2363
2364         /* ASYNC_HP pages first. At present, when the lock the pages is
2365          * to be canceled, the pages covered by the lock will be sent out
2366          * with ASYNC_HP. We have to send out them as soon as possible. */
2367         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2368                 if (oap->oap_async_flags & ASYNC_HP) 
2369                         cfs_list_move(&oap->oap_pending_item, &tmp_list);
2370                 else
2371                         cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
2372                 if (++page_count >= cli->cl_max_pages_per_rpc)
2373                         break;
2374         }
2375
2376         cfs_list_splice(&tmp_list, &lop->lop_pending);
2377         page_count = 0;
2378
2379         /* first we find the pages we're allowed to work with */
2380         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2381                                      oap_pending_item) {
2382                 ops = oap->oap_caller_ops;
2383
2384                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2385                          "magic 0x%x\n", oap, oap->oap_magic);
2386
2387                 if (clob == NULL) {
2388                         /* pin object in memory, so that completion call-backs
2389                          * can be safely called under client_obd_list lock. */
2390                         clob = osc_oap2cl_page(oap)->cp_obj;
2391                         cl_object_get(clob);
2392                 }
2393
2394                 if (page_count != 0 &&
2395                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2396                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2397                                " oap %p, page %p, srvlock %u\n",
2398                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2399                         break;
2400                 }
2401
2402                 /* If there is a gap at the start of this page, it can't merge
2403                  * with any previous page, so we'll hand the network a
2404                  * "fragmented" page array that it can't transfer in 1 RDMA */
2405                 if (page_count != 0 && oap->oap_page_off != 0)
2406                         break;
2407
2408                 /* in llite being 'ready' equates to the page being locked
2409                  * until completion unlocks it.  commit_write submits a page
2410                  * as not ready because its unlock will happen unconditionally
2411                  * as the call returns.  if we race with commit_write giving
2412                  * us that page we dont' want to create a hole in the page
2413                  * stream, so we stop and leave the rpc to be fired by
2414                  * another dirtier or kupdated interval (the not ready page
2415                  * will still be on the dirty list).  we could call in
2416                  * at the end of ll_file_write to process the queue again. */
2417                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2418                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2419                                                     cmd);
2420                         if (rc < 0)
2421                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2422                                                 "instead of ready\n", oap,
2423                                                 oap->oap_page, rc);
2424                         switch (rc) {
2425                         case -EAGAIN:
2426                                 /* llite is telling us that the page is still
2427                                  * in commit_write and that we should try
2428                                  * and put it in an rpc again later.  we
2429                                  * break out of the loop so we don't create
2430                                  * a hole in the sequence of pages in the rpc
2431                                  * stream.*/
2432                                 oap = NULL;
2433                                 break;
2434                         case -EINTR:
2435                                 /* the io isn't needed.. tell the checks
2436                                  * below to complete the rpc with EINTR */
2437                                 cfs_spin_lock(&oap->oap_lock);
2438                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2439                                 cfs_spin_unlock(&oap->oap_lock);
2440                                 oap->oap_count = -EINTR;
2441                                 break;
2442                         case 0:
2443                                 cfs_spin_lock(&oap->oap_lock);
2444                                 oap->oap_async_flags |= ASYNC_READY;
2445                                 cfs_spin_unlock(&oap->oap_lock);
2446                                 break;
2447                         default:
2448                                 LASSERTF(0, "oap %p page %p returned %d "
2449                                             "from make_ready\n", oap,
2450                                             oap->oap_page, rc);
2451                                 break;
2452                         }
2453                 }
2454                 if (oap == NULL)
2455                         break;
2456                 /*
2457                  * Page submitted for IO has to be locked. Either by
2458                  * ->ap_make_ready() or by higher layers.
2459                  */
2460 #if defined(__KERNEL__) && defined(__linux__)
2461                 {
2462                         struct cl_page *page;
2463
2464                         page = osc_oap2cl_page(oap);
2465
2466                         if (page->cp_type == CPT_CACHEABLE &&
2467                             !(PageLocked(oap->oap_page) &&
2468                               (CheckWriteback(oap->oap_page, cmd)))) {
2469                                 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2470                                        oap->oap_page,
2471                                        (long)oap->oap_page->flags,
2472                                        oap->oap_async_flags);
2473                                 LBUG();
2474                         }
2475                 }
2476 #endif
2477
2478                 /* take the page out of our book-keeping */
2479                 cfs_list_del_init(&oap->oap_pending_item);
2480                 lop_update_pending(cli, lop, cmd, -1);
2481                 cfs_list_del_init(&oap->oap_urgent_item);
2482
2483                 if (page_count == 0)
2484                         starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2485                                           (PTLRPC_MAX_BRW_SIZE - 1);
2486
2487                 /* ask the caller for the size of the io as the rpc leaves. */
2488                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2489                         oap->oap_count =
2490                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2491                                                       cmd);
2492                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2493                 }
2494                 if (oap->oap_count <= 0) {
2495                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2496                                oap->oap_count);
2497                         osc_ap_completion(env, cli, NULL,
2498                                           oap, 0, oap->oap_count);
2499                         continue;
2500                 }
2501
2502                 /* now put the page back in our accounting */
2503                 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2504                 if (page_count == 0)
2505                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2506                 if (++page_count >= cli->cl_max_pages_per_rpc)
2507                         break;
2508
2509                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2510                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2511                  * have the same alignment as the initial writes that allocated
2512                  * extents on the server. */
2513                 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2514                                  oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2515                 if (ending_offset == 0)
2516                         break;
2517
2518                 /* If there is a gap at the end of this page, it can't merge
2519                  * with any subsequent pages, so we'll hand the network a
2520                  * "fragmented" page array that it can't transfer in 1 RDMA */
2521                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2522                         break;
2523         }
2524
2525         osc_wake_cache_waiters(cli);
2526
2527         loi_list_maint(cli, loi);
2528
2529         client_obd_list_unlock(&cli->cl_loi_list_lock);
2530
2531         if (clob != NULL)
2532                 cl_object_put(env, clob);
2533
2534         if (page_count == 0) {
2535                 client_obd_list_lock(&cli->cl_loi_list_lock);
2536                 RETURN(0);
2537         }
2538
2539         req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2540         if (IS_ERR(req)) {
2541                 LASSERT(cfs_list_empty(&rpc_list));
2542                 loi_list_maint(cli, loi);
2543                 RETURN(PTR_ERR(req));
2544         }
2545
2546         aa = ptlrpc_req_async_args(req);
2547
2548         if (cmd == OBD_BRW_READ) {
2549                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2550                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2551                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2552                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2553         } else {
2554                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2555                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2556                                  cli->cl_w_in_flight);
2557                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2558                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2559         }
2560         ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2561
2562         client_obd_list_lock(&cli->cl_loi_list_lock);
2563
2564         if (cmd == OBD_BRW_READ)
2565                 cli->cl_r_in_flight++;
2566         else
2567                 cli->cl_w_in_flight++;
2568
2569         /* queued sync pages can be torn down while the pages
2570          * were between the pending list and the rpc */
2571         tmp = NULL;
2572         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2573                 /* only one oap gets a request reference */
2574                 if (tmp == NULL)
2575                         tmp = oap;
2576                 if (oap->oap_interrupted && !req->rq_intr) {
2577                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2578                                oap, req);
2579                         ptlrpc_mark_interrupted(req);
2580                 }
2581         }
2582         if (tmp != NULL)
2583                 tmp->oap_request = ptlrpc_request_addref(req);
2584
2585         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2586                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2587
2588         req->rq_interpret_reply = brw_interpret;
2589         ptlrpcd_add_req(req, PSCOPE_BRW);
2590         RETURN(1);
2591 }
2592
2593 #define LOI_DEBUG(LOI, STR, args...)                                     \
2594         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2595                !cfs_list_empty(&(LOI)->loi_ready_item) ||                \
2596                !cfs_list_empty(&(LOI)->loi_hp_ready_item),               \
2597                (LOI)->loi_write_lop.lop_num_pending,                     \
2598                !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent),        \
2599                (LOI)->loi_read_lop.lop_num_pending,                      \
2600                !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent),         \
2601                args)                                                     \
2602
2603 /* This is called by osc_check_rpcs() to find which objects have pages that
2604  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2605 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2606 {
2607         ENTRY;
2608
2609         /* First return objects that have blocked locks so that they
2610          * will be flushed quickly and other clients can get the lock,
2611          * then objects which have pages ready to be stuffed into RPCs */
2612         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2613                 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2614                                       struct lov_oinfo, loi_hp_ready_item));
2615         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2616                 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2617                                       struct lov_oinfo, loi_ready_item));
2618
2619         /* then if we have cache waiters, return all objects with queued
2620          * writes.  This is especially important when many small files
2621          * have filled up the cache and not been fired into rpcs because
2622          * they don't pass the nr_pending/object threshhold */
2623         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2624             !cfs_list_empty(&cli->cl_loi_write_list))
2625                 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2626                                       struct lov_oinfo, loi_write_item));
2627
2628         /* then return all queued objects when we have an invalid import
2629          * so that they get flushed */
2630         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2631                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2632                         RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2633                                               struct lov_oinfo,
2634                                               loi_write_item));
2635                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2636                         RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2637                                               struct lov_oinfo, loi_read_item));
2638         }
2639         RETURN(NULL);
2640 }
2641
2642 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2643 {
2644         struct osc_async_page *oap;
2645         int hprpc = 0;
2646
2647         if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2648                 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2649                                      struct osc_async_page, oap_urgent_item);
2650                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2651         }
2652
2653         if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2654                 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2655                                      struct osc_async_page, oap_urgent_item);
2656                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2657         }
2658
2659         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2660 }
2661
2662 /* called with the loi list lock held */
2663 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2664 {
2665         struct lov_oinfo *loi;
2666         int rc = 0, race_counter = 0;
2667         ENTRY;
2668
2669         while ((loi = osc_next_loi(cli)) != NULL) {
2670                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2671
2672                 if (osc_max_rpc_in_flight(cli, loi))
2673                         break;
2674
2675                 /* attempt some read/write balancing by alternating between
2676                  * reads and writes in an object.  The makes_rpc checks here
2677                  * would be redundant if we were getting read/write work items
2678                  * instead of objects.  we don't want send_oap_rpc to drain a
2679                  * partial read pending queue when we're given this object to
2680                  * do io on writes while there are cache waiters */
2681                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2682                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2683                                               &loi->loi_write_lop);
2684                         if (rc < 0) {
2685                                 CERROR("Write request failed with %d\n", rc);
2686
2687                                 /* osc_send_oap_rpc failed, mostly because of
2688                                  * memory pressure.
2689                                  *
2690                                  * It can't break here, because if:
2691                                  *  - a page was submitted by osc_io_submit, so
2692                                  *    page locked;
2693                                  *  - no request in flight
2694                                  *  - no subsequent request
2695                                  * The system will be in live-lock state,
2696                                  * because there is no chance to call
2697                                  * osc_io_unplug() and osc_check_rpcs() any
2698                                  * more. pdflush can't help in this case,
2699                                  * because it might be blocked at grabbing
2700                                  * the page lock as we mentioned.
2701                                  *
2702                                  * Anyway, continue to drain pages. */
2703                                 /* break; */
2704                         }
2705
2706                         if (rc > 0)
2707                                 race_counter = 0;
2708                         else
2709                                 race_counter++;
2710                 }
2711                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2712                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2713                                               &loi->loi_read_lop);
2714                         if (rc < 0)
2715                                 CERROR("Read request failed with %d\n", rc);
2716
2717                         if (rc > 0)
2718                                 race_counter = 0;
2719                         else
2720                                 race_counter++;
2721                 }
2722
2723                 /* attempt some inter-object balancing by issueing rpcs
2724                  * for each object in turn */
2725                 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2726                         cfs_list_del_init(&loi->loi_hp_ready_item);
2727                 if (!cfs_list_empty(&loi->loi_ready_item))
2728                         cfs_list_del_init(&loi->loi_ready_item);
2729                 if (!cfs_list_empty(&loi->loi_write_item))
2730                         cfs_list_del_init(&loi->loi_write_item);
2731                 if (!cfs_list_empty(&loi->loi_read_item))
2732                         cfs_list_del_init(&loi->loi_read_item);
2733
2734                 loi_list_maint(cli, loi);
2735
2736                 /* send_oap_rpc fails with 0 when make_ready tells it to
2737                  * back off.  llite's make_ready does this when it tries
2738                  * to lock a page queued for write that is already locked.
2739                  * we want to try sending rpcs from many objects, but we
2740                  * don't want to spin failing with 0.  */
2741                 if (race_counter == 10)
2742                         break;
2743         }
2744         EXIT;
2745 }
2746
2747 /* we're trying to queue a page in the osc so we're subject to the
2748  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2749  * If the osc's queued pages are already at that limit, then we want to sleep
2750  * until there is space in the osc's queue for us.  We also may be waiting for
2751  * write credits from the OST if there are RPCs in flight that may return some
2752  * before we fall back to sync writes.
2753  *
2754  * We need this know our allocation was granted in the presence of signals */
2755 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2756 {
2757         int rc;
2758         ENTRY;
2759         client_obd_list_lock(&cli->cl_loi_list_lock);
2760         rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2761         client_obd_list_unlock(&cli->cl_loi_list_lock);
2762         RETURN(rc);
2763 };
2764
2765 /**
2766  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2767  * is available.
2768  */
2769 int osc_enter_cache_try(const struct lu_env *env,
2770                         struct client_obd *cli, struct lov_oinfo *loi,
2771                         struct osc_async_page *oap, int transient)
2772 {
2773         int has_grant;
2774
2775         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2776         if (has_grant) {
2777                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2778                 if (transient) {
2779                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2780                         cfs_atomic_inc(&obd_dirty_transit_pages);
2781                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2782                 }
2783         }
2784         return has_grant;
2785 }
2786
2787 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2788  * grant or cache space. */
2789 static int osc_enter_cache(const struct lu_env *env,
2790                            struct client_obd *cli, struct lov_oinfo *loi,
2791                            struct osc_async_page *oap)
2792 {
2793         struct osc_cache_waiter ocw;
2794         struct l_wait_info lwi = { 0 };
2795
2796         ENTRY;
2797
2798         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2799                "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2800                cli->cl_dirty_max, obd_max_dirty_pages,
2801                cli->cl_lost_grant, cli->cl_avail_grant);
2802
2803         /* force the caller to try sync io.  this can jump the list
2804          * of queued writes and create a discontiguous rpc stream */
2805         if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2806             loi->loi_ar.ar_force_sync)
2807                 RETURN(-EDQUOT);
2808
2809         /* Hopefully normal case - cache space and write credits available */
2810         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2811             cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2812             osc_enter_cache_try(env, cli, loi, oap, 0))
2813                 RETURN(0);
2814
2815         /* It is safe to block as a cache waiter as long as there is grant
2816          * space available or the hope of additional grant being returned
2817          * when an in flight write completes.  Using the write back cache
2818          * if possible is preferable to sending the data synchronously
2819          * because write pages can then be merged in to large requests.
2820          * The addition of this cache waiter will causing pending write
2821          * pages to be sent immediately. */
2822         if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2823                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2824                 cfs_waitq_init(&ocw.ocw_waitq);
2825                 ocw.ocw_oap = oap;
2826                 ocw.ocw_rc = 0;
2827
2828                 loi_list_maint(cli, loi);
2829                 osc_check_rpcs(env, cli);
2830                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2831
2832                 CDEBUG(D_CACHE, "sleeping for cache space\n");
2833                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2834
2835                 client_obd_list_lock(&cli->cl_loi_list_lock);
2836                 if (!cfs_list_empty(&ocw.ocw_entry)) {
2837                         cfs_list_del(&ocw.ocw_entry);
2838                         RETURN(-EINTR);
2839                 }
2840                 RETURN(ocw.ocw_rc);
2841         }
2842
2843         RETURN(-EDQUOT);
2844 }
2845
2846
2847 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2848                         struct lov_oinfo *loi, cfs_page_t *page,
2849                         obd_off offset, const struct obd_async_page_ops *ops,
2850                         void *data, void **res, int nocache,
2851                         struct lustre_handle *lockh)
2852 {
2853         struct osc_async_page *oap;
2854
2855         ENTRY;
2856
2857         if (!page)
2858                 return cfs_size_round(sizeof(*oap));
2859
2860         oap = *res;
2861         oap->oap_magic = OAP_MAGIC;
2862         oap->oap_cli = &exp->exp_obd->u.cli;
2863         oap->oap_loi = loi;
2864
2865         oap->oap_caller_ops = ops;
2866         oap->oap_caller_data = data;
2867
2868         oap->oap_page = page;
2869         oap->oap_obj_off = offset;
2870         if (!client_is_remote(exp) &&
2871             cfs_capable(CFS_CAP_SYS_RESOURCE))
2872                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2873
2874         LASSERT(!(offset & ~CFS_PAGE_MASK));
2875
2876         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2877         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2878         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2879         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2880
2881         cfs_spin_lock_init(&oap->oap_lock);
2882         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2883         RETURN(0);
2884 }
2885
2886 struct osc_async_page *oap_from_cookie(void *cookie)
2887 {
2888         struct osc_async_page *oap = cookie;
2889         if (oap->oap_magic != OAP_MAGIC)
2890                 return ERR_PTR(-EINVAL);
2891         return oap;
2892 };
2893
2894 int osc_queue_async_io(const struct lu_env *env,
2895                        struct obd_export *exp, struct lov_stripe_md *lsm,
2896                        struct lov_oinfo *loi, void *cookie,
2897                        int cmd, obd_off off, int count,
2898                        obd_flag brw_flags, enum async_flags async_flags)
2899 {
2900         struct client_obd *cli = &exp->exp_obd->u.cli;
2901         struct osc_async_page *oap;
2902         int rc = 0;
2903         ENTRY;
2904
2905         oap = oap_from_cookie(cookie);
2906         if (IS_ERR(oap))
2907                 RETURN(PTR_ERR(oap));
2908
2909         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2910                 RETURN(-EIO);
2911
2912         if (!cfs_list_empty(&oap->oap_pending_item) ||
2913             !cfs_list_empty(&oap->oap_urgent_item) ||
2914             !cfs_list_empty(&oap->oap_rpc_item))
2915                 RETURN(-EBUSY);
2916
2917         /* check if the file's owner/group is over quota */
2918         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2919                 struct cl_object *obj;
2920                 struct cl_attr    attr; /* XXX put attr into thread info */
2921                 unsigned int qid[MAXQUOTAS];
2922
2923                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2924
2925                 cl_object_attr_lock(obj);
2926                 rc = cl_object_attr_get(env, obj, &attr);
2927                 cl_object_attr_unlock(obj);
2928
2929                 qid[USRQUOTA] = attr.cat_uid;
2930                 qid[GRPQUOTA] = attr.cat_gid;
2931                 if (rc == 0 &&
2932                     lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2933                         rc = -EDQUOT;
2934                 if (rc)
2935                         RETURN(rc);
2936         }
2937
2938         if (loi == NULL)
2939                 loi = lsm->lsm_oinfo[0];
2940
2941         client_obd_list_lock(&cli->cl_loi_list_lock);
2942
2943         LASSERT(off + count <= CFS_PAGE_SIZE);
2944         oap->oap_cmd = cmd;
2945         oap->oap_page_off = off;
2946         oap->oap_count = count;
2947         oap->oap_brw_flags = brw_flags;
2948         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2949         if (libcfs_memory_pressure_get())
2950                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2951         cfs_spin_lock(&oap->oap_lock);
2952         oap->oap_async_flags = async_flags;
2953         cfs_spin_unlock(&oap->oap_lock);
2954
2955         if (cmd & OBD_BRW_WRITE) {
2956                 rc = osc_enter_cache(env, cli, loi, oap);
2957                 if (rc) {
2958                         client_obd_list_unlock(&cli->cl_loi_list_lock);
2959                         RETURN(rc);
2960                 }
2961         }
2962
2963         osc_oap_to_pending(oap);
2964         loi_list_maint(cli, loi);
2965
2966         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2967                   cmd);
2968
2969         osc_check_rpcs(env, cli);
2970         client_obd_list_unlock(&cli->cl_loi_list_lock);
2971
2972         RETURN(0);
2973 }
2974
2975 /* aka (~was & now & flag), but this is more clear :) */
2976 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2977
2978 int osc_set_async_flags_base(struct client_obd *cli,
2979                              struct lov_oinfo *loi, struct osc_async_page *oap,
2980                              obd_flag async_flags)
2981 {
2982         struct loi_oap_pages *lop;
2983         int flags = 0;
2984         ENTRY;
2985
2986         LASSERT(!cfs_list_empty(&oap->oap_pending_item));
2987
2988         if (oap->oap_cmd & OBD_BRW_WRITE) {
2989                 lop = &loi->loi_write_lop;
2990         } else {
2991                 lop = &loi->loi_read_lop;
2992         }
2993
2994         if ((oap->oap_async_flags & async_flags) == async_flags)
2995                 RETURN(0);
2996
2997         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2998                 flags |= ASYNC_READY;
2999
3000         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3001             cfs_list_empty(&oap->oap_rpc_item)) {
3002                 if (oap->oap_async_flags & ASYNC_HP)
3003                         cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3004                 else
3005                         cfs_list_add_tail(&oap->oap_urgent_item,
3006                                           &lop->lop_urgent);
3007                 flags |= ASYNC_URGENT;
3008                 loi_list_maint(cli, loi);
3009         }
3010         cfs_spin_lock(&oap->oap_lock);
3011         oap->oap_async_flags |= flags;
3012         cfs_spin_unlock(&oap->oap_lock);
3013
3014         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3015                         oap->oap_async_flags);
3016         RETURN(0);
3017 }
3018
3019 int osc_teardown_async_page(struct obd_export *exp,
3020                             struct lov_stripe_md *lsm,
3021                             struct lov_oinfo *loi, void *cookie)
3022 {
3023         struct client_obd *cli = &exp->exp_obd->u.cli;
3024         struct loi_oap_pages *lop;
3025         struct osc_async_page *oap;
3026         int rc = 0;
3027         ENTRY;
3028
3029         oap = oap_from_cookie(cookie);
3030         if (IS_ERR(oap))
3031                 RETURN(PTR_ERR(oap));
3032
3033         if (loi == NULL)
3034                 loi = lsm->lsm_oinfo[0];
3035
3036         if (oap->oap_cmd & OBD_BRW_WRITE) {
3037                 lop = &loi->loi_write_lop;
3038         } else {
3039                 lop = &loi->loi_read_lop;
3040         }
3041
3042         client_obd_list_lock(&cli->cl_loi_list_lock);
3043
3044         if (!cfs_list_empty(&oap->oap_rpc_item))
3045                 GOTO(out, rc = -EBUSY);
3046
3047         osc_exit_cache(cli, oap, 0);
3048         osc_wake_cache_waiters(cli);
3049
3050         if (!cfs_list_empty(&oap->oap_urgent_item)) {
3051                 cfs_list_del_init(&oap->oap_urgent_item);
3052                 cfs_spin_lock(&oap->oap_lock);
3053                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3054                 cfs_spin_unlock(&oap->oap_lock);
3055         }
3056         if (!cfs_list_empty(&oap->oap_pending_item)) {
3057                 cfs_list_del_init(&oap->oap_pending_item);
3058                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3059         }
3060         loi_list_maint(cli, loi);
3061         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3062 out:
3063         client_obd_list_unlock(&cli->cl_loi_list_lock);
3064         RETURN(rc);
3065 }
3066
3067 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3068                                          struct ldlm_enqueue_info *einfo,
3069                                          int flags)
3070 {
3071         void *data = einfo->ei_cbdata;
3072
3073         LASSERT(lock != NULL);
3074         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3075         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3076         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3077         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3078
3079         lock_res_and_lock(lock);
3080         cfs_spin_lock(&osc_ast_guard);
3081         LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3082         lock->l_ast_data = data;
3083         cfs_spin_unlock(&osc_ast_guard);
3084         unlock_res_and_lock(lock);
3085 }
3086
3087 static void osc_set_data_with_check(struct lustre_handle *lockh,
3088                                     struct ldlm_enqueue_info *einfo,
3089                                     int flags)
3090 {
3091         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3092
3093         if (lock != NULL) {
3094                 osc_set_lock_data_with_check(lock, einfo, flags);
3095                 LDLM_LOCK_PUT(lock);
3096         } else
3097                 CERROR("lockh %p, data %p - client evicted?\n",
3098                        lockh, einfo->ei_cbdata);
3099 }
3100
3101 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3102                              ldlm_iterator_t replace, void *data)
3103 {
3104         struct ldlm_res_id res_id;
3105         struct obd_device *obd = class_exp2obd(exp);
3106
3107         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3108         ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3109         return 0;
3110 }
3111
3112 /* find any ldlm lock of the inode in osc
3113  * return 0    not find
3114  *        1    find one
3115  *      < 0    error */
3116 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3117                            ldlm_iterator_t replace, void *data)
3118 {
3119         struct ldlm_res_id res_id;
3120         struct obd_device *obd = class_exp2obd(exp);
3121         int rc = 0;
3122
3123         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3124         rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3125         if (rc == LDLM_ITER_STOP)
3126                 return(1);
3127         if (rc == LDLM_ITER_CONTINUE)
3128                 return(0);
3129         return(rc);
3130 }
3131
3132 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3133                             obd_enqueue_update_f upcall, void *cookie,
3134                             int *flags, int rc)
3135 {
3136         int intent = *flags & LDLM_FL_HAS_INTENT;
3137         ENTRY;
3138
3139         if (intent) {
3140                 /* The request was created before ldlm_cli_enqueue call. */
3141                 if (rc == ELDLM_LOCK_ABORTED) {
3142                         struct ldlm_reply *rep;
3143                         rep = req_capsule_server_get(&req->rq_pill,
3144                                                      &RMF_DLM_REP);
3145
3146                         LASSERT(rep != NULL);
3147                         if (rep->lock_policy_res1)
3148                                 rc = rep->lock_policy_res1;
3149                 }
3150         }
3151
3152         if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3153                 *flags |= LDLM_FL_LVB_READY;
3154                 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3155                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3156         }
3157
3158         /* Call the update callback. */
3159         rc = (*upcall)(cookie, rc);
3160         RETURN(rc);
3161 }
3162
3163 static int osc_enqueue_interpret(const struct lu_env *env,
3164                                  struct ptlrpc_request *req,
3165                                  struct osc_enqueue_args *aa, int rc)
3166 {
3167         struct ldlm_lock *lock;
3168         struct lustre_handle handle;
3169         __u32 mode;
3170
3171         /* Make a local copy of a lock handle and a mode, because aa->oa_*
3172          * might be freed anytime after lock upcall has been called. */
3173         lustre_handle_copy(&handle, aa->oa_lockh);
3174         mode = aa->oa_ei->ei_mode;
3175
3176         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3177          * be valid. */
3178         lock = ldlm_handle2lock(&handle);
3179
3180         /* Take an additional reference so that a blocking AST that
3181          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3182          * to arrive after an upcall has been executed by
3183          * osc_enqueue_fini(). */
3184         ldlm_lock_addref(&handle, mode);
3185
3186         /* Complete obtaining the lock procedure. */
3187         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3188                                    mode, aa->oa_flags, aa->oa_lvb,
3189                                    sizeof(*aa->oa_lvb), &handle, rc);
3190         /* Complete osc stuff. */
3191         rc = osc_enqueue_fini(req, aa->oa_lvb,
3192                               aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3193
3194         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3195
3196         /* Release the lock for async request. */
3197         if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3198                 /*
3199                  * Releases a reference taken by ldlm_cli_enqueue(), if it is
3200                  * not already released by
3201                  * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3202                  */
3203                 ldlm_lock_decref(&handle, mode);
3204
3205         LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3206                  aa->oa_lockh, req, aa);
3207         ldlm_lock_decref(&handle, mode);
3208         LDLM_LOCK_PUT(lock);
3209         return rc;
3210 }
3211
3212 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3213                         struct lov_oinfo *loi, int flags,
3214                         struct ost_lvb *lvb, __u32 mode, int rc)
3215 {
3216         if (rc == ELDLM_OK) {
3217                 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3218                 __u64 tmp;
3219
3220                 LASSERT(lock != NULL);
3221                 loi->loi_lvb = *lvb;
3222                 tmp = loi->loi_lvb.lvb_size;
3223                 /* Extend KMS up to the end of this lock and no further
3224                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3225                 if (tmp > lock->l_policy_data.l_extent.end)
3226                         tmp = lock->l_policy_data.l_extent.end + 1;
3227                 if (tmp >= loi->loi_kms) {
3228                         LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3229                                    ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3230                         loi_kms_set(loi, tmp);
3231                 } else {
3232                         LDLM_DEBUG(lock, "lock acquired, setting rss="
3233                                    LPU64"; leaving kms="LPU64", end="LPU64,
3234                                    loi->loi_lvb.lvb_size, loi->loi_kms,
3235                                    lock->l_policy_data.l_extent.end);
3236                 }
3237                 ldlm_lock_allow_match(lock);
3238                 LDLM_LOCK_PUT(lock);
3239         } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3240                 loi->loi_lvb = *lvb;
3241                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3242                        " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3243                 rc = ELDLM_OK;
3244         }
3245 }
3246 EXPORT_SYMBOL(osc_update_enqueue);
3247
3248 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3249
3250 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3251  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3252  * other synchronous requests, however keeping some locks and trying to obtain
3253  * others may take a considerable amount of time in a case of ost failure; and
3254  * when other sync requests do not get released lock from a client, the client
3255  * is excluded from the cluster -- such scenarious make the life difficult, so
3256  * release locks just after they are obtained. */
3257 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3258                      int *flags, ldlm_policy_data_t *policy,
3259                      struct ost_lvb *lvb, int kms_valid,
3260                      obd_enqueue_update_f upcall, void *cookie,
3261                      struct ldlm_enqueue_info *einfo,
3262                      struct lustre_handle *lockh,
3263                      struct ptlrpc_request_set *rqset, int async)
3264 {
3265         struct obd_device *obd = exp->exp_obd;
3266         struct ptlrpc_request *req = NULL;
3267         int intent = *flags & LDLM_FL_HAS_INTENT;
3268         ldlm_mode_t mode;
3269         int rc;
3270         ENTRY;
3271
3272         /* Filesystem lock extents are extended to page boundaries so that
3273          * dealing with the page cache is a little smoother.  */
3274         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3275         policy->l_extent.end |= ~CFS_PAGE_MASK;
3276
3277         /*
3278          * kms is not valid when either object is completely fresh (so that no
3279          * locks are cached), or object was evicted. In the latter case cached
3280          * lock cannot be used, because it would prime inode state with
3281          * potentially stale LVB.
3282          */
3283         if (!kms_valid)
3284                 goto no_match;
3285
3286         /* Next, search for already existing extent locks that will cover us */
3287         /* If we're trying to read, we also search for an existing PW lock.  The
3288          * VFS and page cache already protect us locally, so lots of readers/
3289          * writers can share a single PW lock.
3290          *
3291          * There are problems with conversion deadlocks, so instead of
3292          * converting a read lock to a write lock, we'll just enqueue a new
3293          * one.
3294          *
3295          * At some point we should cancel the read lock instead of making them
3296          * send us a blocking callback, but there are problems with canceling
3297          * locks out from other users right now, too. */
3298         mode = einfo->ei_mode;
3299         if (einfo->ei_mode == LCK_PR)
3300                 mode |= LCK_PW;
3301         mode = ldlm_lock_match(obd->obd_namespace,
3302                                *flags | LDLM_FL_LVB_READY, res_id,
3303                                einfo->ei_type, policy, mode, lockh, 0);
3304         if (mode) {
3305                 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3306
3307                 if (matched->l_ast_data == NULL ||
3308                     matched->l_ast_data == einfo->ei_cbdata) {
3309                         /* addref the lock only if not async requests and PW
3310                          * lock is matched whereas we asked for PR. */
3311                         if (!rqset && einfo->ei_mode != mode)
3312                                 ldlm_lock_addref(lockh, LCK_PR);
3313                         osc_set_lock_data_with_check(matched, einfo, *flags);
3314                         if (intent) {
3315                                 /* I would like to be able to ASSERT here that
3316                                  * rss <= kms, but I can't, for reasons which
3317                                  * are explained in lov_enqueue() */
3318                         }
3319
3320                         /* We already have a lock, and it's referenced */
3321                         (*upcall)(cookie, ELDLM_OK);
3322
3323                         /* For async requests, decref the lock. */
3324                         if (einfo->ei_mode != mode)
3325                                 ldlm_lock_decref(lockh, LCK_PW);
3326                         else if (rqset)
3327                                 ldlm_lock_decref(lockh, einfo->ei_mode);
3328                         LDLM_LOCK_PUT(matched);
3329                         RETURN(ELDLM_OK);
3330                 } else
3331                         ldlm_lock_decref(lockh, mode);
3332                 LDLM_LOCK_PUT(matched);
3333         }
3334
3335  no_match:
3336         if (intent) {
3337                 CFS_LIST_HEAD(cancels);
3338                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3339                                            &RQF_LDLM_ENQUEUE_LVB);
3340                 if (req == NULL)
3341                         RETURN(-ENOMEM);
3342
3343                 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3344                 if (rc)
3345                         RETURN(rc);
3346
3347                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3348                                      sizeof *lvb);
3349                 ptlrpc_request_set_replen(req);
3350         }
3351
3352         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3353         *flags &= ~LDLM_FL_BLOCK_GRANTED;
3354
3355         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3356                               sizeof(*lvb), lockh, async);
3357         if (rqset) {
3358                 if (!rc) {
3359                         struct osc_enqueue_args *aa;
3360                         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3361                         aa = ptlrpc_req_async_args(req);
3362                         aa->oa_ei = einfo;
3363                         aa->oa_exp = exp;
3364                         aa->oa_flags  = flags;
3365                         aa->oa_upcall = upcall;
3366                         aa->oa_cookie = cookie;
3367                         aa->oa_lvb    = lvb;
3368                         aa->oa_lockh  = lockh;
3369
3370                         req->rq_interpret_reply =
3371                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3372                         if (rqset == PTLRPCD_SET)
3373                                 ptlrpcd_add_req(req, PSCOPE_OTHER);
3374                         else
3375                                 ptlrpc_set_add_req(rqset, req);
3376                 } else if (intent) {
3377                         ptlrpc_req_finished(req);
3378                 }
3379                 RETURN(rc);
3380         }
3381
3382         rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3383         if (intent)
3384                 ptlrpc_req_finished(req);
3385
3386         RETURN(rc);
3387 }
3388
3389 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3390                        struct ldlm_enqueue_info *einfo,
3391                        struct ptlrpc_request_set *rqset)
3392 {
3393         struct ldlm_res_id res_id;
3394         int rc;
3395         ENTRY;
3396
3397         osc_build_res_name(oinfo->oi_md->lsm_object_id,
3398                            oinfo->oi_md->lsm_object_seq, &res_id);
3399
3400         rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3401                               &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3402                               oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3403                               oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3404                               rqset, rqset != NULL);
3405         RETURN(rc);
3406 }
3407
3408 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3409                    __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3410                    int *flags, void *data, struct lustre_handle *lockh,
3411                    int unref)
3412 {
3413         struct obd_device *obd = exp->exp_obd;
3414         int lflags = *flags;
3415         ldlm_mode_t rc;
3416         ENTRY;
3417
3418         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3419                 RETURN(-EIO);
3420
3421         /* Filesystem lock extents are extended to page boundaries so that
3422          * dealing with the page cache is a little smoother */
3423         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3424         policy->l_extent.end |= ~CFS_PAGE_MASK;
3425
3426         /* Next, search for already existing extent locks that will cover us */
3427         /* If we're trying to read, we also search for an existing PW lock.  The
3428          * VFS and page cache already protect us locally, so lots of readers/
3429          * writers can share a single PW lock. */
3430         rc = mode;
3431         if (mode == LCK_PR)
3432                 rc |= LCK_PW;
3433         rc = ldlm_lock_match(obd->obd_namespace, lflags,
3434                              res_id, type, policy, rc, lockh, unref);
3435         if (rc) {
3436                 if (data != NULL)
3437                         osc_set_data_with_check(lockh, data, lflags);
3438                 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3439                         ldlm_lock_addref(lockh, LCK_PR);
3440                         ldlm_lock_decref(lockh, LCK_PW);
3441                 }
3442                 RETURN(rc);
3443         }
3444         RETURN(rc);
3445 }
3446
3447 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3448 {
3449         ENTRY;
3450
3451         if (unlikely(mode == LCK_GROUP))
3452                 ldlm_lock_decref_and_cancel(lockh, mode);
3453         else
3454                 ldlm_lock_decref(lockh, mode);
3455
3456         RETURN(0);
3457 }
3458
3459 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3460                       __u32 mode, struct lustre_handle *lockh)
3461 {
3462         ENTRY;
3463         RETURN(osc_cancel_base(lockh, mode));
3464 }
3465
3466 static int osc_cancel_unused(struct obd_export *exp,
3467                              struct lov_stripe_md *lsm,
3468                              ldlm_cancel_flags_t flags,
3469                              void *opaque)
3470 {
3471         struct obd_device *obd = class_exp2obd(exp);
3472         struct ldlm_res_id res_id, *resp = NULL;
3473
3474         if (lsm != NULL) {
3475                 resp = osc_build_res_name(lsm->lsm_object_id,
3476                                           lsm->lsm_object_seq, &res_id);
3477         }
3478
3479         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3480 }
3481
3482 static int osc_statfs_interpret(const struct lu_env *env,
3483                                 struct ptlrpc_request *req,
3484                                 struct osc_async_args *aa, int rc)
3485 {
3486         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3487         struct obd_statfs *msfs;
3488         __u64 used;
3489         ENTRY;
3490
3491         if (rc == -EBADR)
3492                 /* The request has in fact never been sent
3493                  * due to issues at a higher level (LOV).
3494                  * Exit immediately since the caller is
3495                  * aware of the problem and takes care
3496                  * of the clean up */
3497                  RETURN(rc);
3498
3499         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3500             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3501                 GOTO(out, rc = 0);
3502
3503         if (rc != 0)
3504                 GOTO(out, rc);
3505
3506         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3507         if (msfs == NULL) {
3508                 GOTO(out, rc = -EPROTO);
3509         }
3510
3511         /* Reinitialize the RDONLY and DEGRADED flags at the client
3512          * on each statfs, so they don't stay set permanently. */
3513         cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3514
3515         if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3516                 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3517         else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3518                 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3519
3520         if (unlikely(msfs->os_state & OS_STATE_READONLY))
3521                 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3522         else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3523                 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3524
3525         /* Add a bit of hysteresis so this flag isn't continually flapping,
3526          * and ensure that new files don't get extremely fragmented due to
3527          * only a small amount of available space in the filesystem.
3528          * We want to set the NOSPC flag when there is less than ~0.1% free
3529          * and clear it when there is at least ~0.2% free space, so:
3530          *                   avail < ~0.1% max          max = avail + used
3531          *            1025 * avail < avail + used       used = blocks - free
3532          *            1024 * avail < used
3533          *            1024 * avail < blocks - free                      
3534          *                   avail < ((blocks - free) >> 10)    
3535          *
3536          * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3537          * lose that amount of space so in those cases we report no space left
3538          * if their is less than 1 GB left.                             */
3539         used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3540         if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3541                      ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3542                 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3543         else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3544                 (msfs->os_ffree > 64) && (msfs->os_bavail > (used << 1))))
3545                         cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_NOSPC;
3546
3547         cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3548
3549         *aa->aa_oi->oi_osfs = *msfs;
3550 out:
3551         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3552         RETURN(rc);
3553 }
3554
3555 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3556                             __u64 max_age, struct ptlrpc_request_set *rqset)
3557 {
3558         struct ptlrpc_request *req;
3559         struct osc_async_args *aa;
3560         int                    rc;
3561         ENTRY;
3562
3563         /* We could possibly pass max_age in the request (as an absolute
3564          * timestamp or a "seconds.usec ago") so the target can avoid doing
3565          * extra calls into the filesystem if that isn't necessary (e.g.
3566          * during mount that would help a bit).  Having relative timestamps
3567          * is not so great if request processing is slow, while absolute
3568          * timestamps are not ideal because they need time synchronization. */
3569         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3570         if (req == NULL)
3571                 RETURN(-ENOMEM);
3572
3573         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3574         if (rc) {
3575                 ptlrpc_request_free(req);
3576                 RETURN(rc);
3577         }
3578         ptlrpc_request_set_replen(req);
3579         req->rq_request_portal = OST_CREATE_PORTAL;
3580         ptlrpc_at_set_req_timeout(req);
3581
3582         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3583                 /* procfs requests not want stat in wait for avoid deadlock */
3584                 req->rq_no_resend = 1;
3585                 req->rq_no_delay = 1;
3586         }
3587
3588         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3589         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3590         aa = ptlrpc_req_async_args(req);
3591         aa->aa_oi = oinfo;
3592
3593         ptlrpc_set_add_req(rqset, req);
3594         RETURN(0);
3595 }
3596
3597 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3598                       __u64 max_age, __u32 flags)
3599 {
3600         struct obd_statfs     *msfs;
3601         struct ptlrpc_request *req;
3602         struct obd_import     *imp = NULL;
3603         int rc;
3604         ENTRY;
3605
3606         /*Since the request might also come from lprocfs, so we need
3607          *sync this with client_disconnect_export Bug15684*/
3608         cfs_down_read(&obd->u.cli.cl_sem);
3609         if (obd->u.cli.cl_import)
3610                 imp = class_import_get(obd->u.cli.cl_import);
3611         cfs_up_read(&obd->u.cli.cl_sem);
3612         if (!imp)
3613                 RETURN(-ENODEV);
3614
3615         /* We could possibly pass max_age in the request (as an absolute
3616          * timestamp or a "seconds.usec ago") so the target can avoid doing
3617          * extra calls into the filesystem if that isn't necessary (e.g.
3618          * during mount that would help a bit).  Having relative timestamps
3619          * is not so great if request processing is slow, while absolute
3620          * timestamps are not ideal because they need time synchronization. */
3621         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3622
3623         class_import_put(imp);
3624
3625         if (req == NULL)
3626                 RETURN(-ENOMEM);
3627
3628         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3629         if (rc) {
3630                 ptlrpc_request_free(req);
3631                 RETURN(rc);
3632         }
3633         ptlrpc_request_set_replen(req);
3634         req->rq_request_portal = OST_CREATE_PORTAL;
3635         ptlrpc_at_set_req_timeout(req);
3636
3637         if (flags & OBD_STATFS_NODELAY) {
3638                 /* procfs requests not want stat in wait for avoid deadlock */
3639                 req->rq_no_resend = 1;
3640                 req->rq_no_delay = 1;
3641         }
3642
3643         rc = ptlrpc_queue_wait(req);
3644         if (rc)
3645                 GOTO(out, rc);
3646
3647         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3648         if (msfs == NULL) {
3649         &