Whamcloud - gitweb
1efb9a210fb1ad66c2499cb8267c92b1dd5e0481
[fs/lustre-release.git] / lustre / osc / osc_request.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
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 void osc_release_ppga(struct brw_page **ppga, obd_count count);
67 static int brw_interpret(const struct lu_env *env,
68                          struct ptlrpc_request *req, void *data, int rc);
69 static void osc_check_rpcs0(const struct lu_env *env, struct client_obd *cli,
70                             int ptlrpc);
71 int osc_cleanup(struct obd_device *obd);
72
73 /* Pack OSC object metadata for disk storage (LE byte order). */
74 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
75                       struct lov_stripe_md *lsm)
76 {
77         int lmm_size;
78         ENTRY;
79
80         lmm_size = sizeof(**lmmp);
81         if (!lmmp)
82                 RETURN(lmm_size);
83
84         if (*lmmp && !lsm) {
85                 OBD_FREE(*lmmp, lmm_size);
86                 *lmmp = NULL;
87                 RETURN(0);
88         }
89
90         if (!*lmmp) {
91                 OBD_ALLOC(*lmmp, lmm_size);
92                 if (!*lmmp)
93                         RETURN(-ENOMEM);
94         }
95
96         if (lsm) {
97                 LASSERT(lsm->lsm_object_id);
98                 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
99                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
100                 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
101         }
102
103         RETURN(lmm_size);
104 }
105
106 /* Unpack OSC object metadata from disk storage (LE byte order). */
107 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
108                         struct lov_mds_md *lmm, int lmm_bytes)
109 {
110         int lsm_size;
111         struct obd_import *imp = class_exp2cliimp(exp);
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         if (imp != NULL &&
160             (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
161                 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
162         else
163                 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
164
165         RETURN(lsm_size);
166 }
167
168 static inline void osc_pack_capa(struct ptlrpc_request *req,
169                                  struct ost_body *body, void *capa)
170 {
171         struct obd_capa *oc = (struct obd_capa *)capa;
172         struct lustre_capa *c;
173
174         if (!capa)
175                 return;
176
177         c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
178         LASSERT(c);
179         capa_cpy(c, oc);
180         body->oa.o_valid |= OBD_MD_FLOSSCAPA;
181         DEBUG_CAPA(D_SEC, c, "pack");
182 }
183
184 static inline void osc_pack_req_body(struct ptlrpc_request *req,
185                                      struct obd_info *oinfo)
186 {
187         struct ost_body *body;
188
189         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
190         LASSERT(body);
191
192         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
193         osc_pack_capa(req, body, oinfo->oi_capa);
194 }
195
196 static inline void osc_set_capa_size(struct ptlrpc_request *req,
197                                      const struct req_msg_field *field,
198                                      struct obd_capa *oc)
199 {
200         if (oc == NULL)
201                 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
202         else
203                 /* it is already calculated as sizeof struct obd_capa */
204                 ;
205 }
206
207 static int osc_getattr_interpret(const struct lu_env *env,
208                                  struct ptlrpc_request *req,
209                                  struct osc_async_args *aa, int rc)
210 {
211         struct ost_body *body;
212         ENTRY;
213
214         if (rc != 0)
215                 GOTO(out, rc);
216
217         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
218         if (body) {
219                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
220                 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
221
222                 /* This should really be sent by the OST */
223                 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
224                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
225         } else {
226                 CDEBUG(D_INFO, "can't unpack ost_body\n");
227                 rc = -EPROTO;
228                 aa->aa_oi->oi_oa->o_valid = 0;
229         }
230 out:
231         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
232         RETURN(rc);
233 }
234
235 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
236                              struct ptlrpc_request_set *set)
237 {
238         struct ptlrpc_request *req;
239         struct osc_async_args *aa;
240         int                    rc;
241         ENTRY;
242
243         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
244         if (req == NULL)
245                 RETURN(-ENOMEM);
246
247         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
248         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
249         if (rc) {
250                 ptlrpc_request_free(req);
251                 RETURN(rc);
252         }
253
254         osc_pack_req_body(req, oinfo);
255
256         ptlrpc_request_set_replen(req);
257         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
258
259         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
260         aa = ptlrpc_req_async_args(req);
261         aa->aa_oi = oinfo;
262
263         ptlrpc_set_add_req(set, req);
264         RETURN(0);
265 }
266
267 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
268                        struct obd_info *oinfo)
269 {
270         struct ptlrpc_request *req;
271         struct ost_body       *body;
272         int                    rc;
273         ENTRY;
274
275         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
276         if (req == NULL)
277                 RETURN(-ENOMEM);
278
279         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
280         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
281         if (rc) {
282                 ptlrpc_request_free(req);
283                 RETURN(rc);
284         }
285
286         osc_pack_req_body(req, oinfo);
287
288         ptlrpc_request_set_replen(req);
289
290         rc = ptlrpc_queue_wait(req);
291         if (rc)
292                 GOTO(out, rc);
293
294         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
295         if (body == NULL)
296                 GOTO(out, rc = -EPROTO);
297
298         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
299         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
300
301         /* This should really be sent by the OST */
302         oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
303         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
304
305         EXIT;
306  out:
307         ptlrpc_req_finished(req);
308         return rc;
309 }
310
311 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
312                        struct obd_info *oinfo, struct obd_trans_info *oti)
313 {
314         struct ptlrpc_request *req;
315         struct ost_body       *body;
316         int                    rc;
317         ENTRY;
318
319         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
320
321         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
322         if (req == NULL)
323                 RETURN(-ENOMEM);
324
325         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
326         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
327         if (rc) {
328                 ptlrpc_request_free(req);
329                 RETURN(rc);
330         }
331
332         osc_pack_req_body(req, oinfo);
333
334         ptlrpc_request_set_replen(req);
335
336         rc = ptlrpc_queue_wait(req);
337         if (rc)
338                 GOTO(out, rc);
339
340         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
341         if (body == NULL)
342                 GOTO(out, rc = -EPROTO);
343
344         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
345
346         EXIT;
347 out:
348         ptlrpc_req_finished(req);
349         RETURN(rc);
350 }
351
352 static int osc_setattr_interpret(const struct lu_env *env,
353                                  struct ptlrpc_request *req,
354                                  struct osc_setattr_args *sa, int rc)
355 {
356         struct ost_body *body;
357         ENTRY;
358
359         if (rc != 0)
360                 GOTO(out, rc);
361
362         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
363         if (body == NULL)
364                 GOTO(out, rc = -EPROTO);
365
366         lustre_get_wire_obdo(sa->sa_oa, &body->oa);
367 out:
368         rc = sa->sa_upcall(sa->sa_cookie, rc);
369         RETURN(rc);
370 }
371
372 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
373                            struct obd_trans_info *oti,
374                            obd_enqueue_update_f upcall, void *cookie,
375                            struct ptlrpc_request_set *rqset)
376 {
377         struct ptlrpc_request   *req;
378         struct osc_setattr_args *sa;
379         int                      rc;
380         ENTRY;
381
382         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
383         if (req == NULL)
384                 RETURN(-ENOMEM);
385
386         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
387         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
388         if (rc) {
389                 ptlrpc_request_free(req);
390                 RETURN(rc);
391         }
392
393         if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
394                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
395
396         osc_pack_req_body(req, oinfo);
397
398         ptlrpc_request_set_replen(req);
399
400         /* do mds to ost setattr asynchronously */
401         if (!rqset) {
402                 /* Do not wait for response. */
403                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
404         } else {
405                 req->rq_interpret_reply =
406                         (ptlrpc_interpterer_t)osc_setattr_interpret;
407
408                 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
409                 sa = ptlrpc_req_async_args(req);
410                 sa->sa_oa = oinfo->oi_oa;
411                 sa->sa_upcall = upcall;
412                 sa->sa_cookie = cookie;
413
414                 if (rqset == PTLRPCD_SET)
415                         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
416                 else
417                         ptlrpc_set_add_req(rqset, req);
418         }
419
420         RETURN(0);
421 }
422
423 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
424                              struct obd_trans_info *oti,
425                              struct ptlrpc_request_set *rqset)
426 {
427         return osc_setattr_async_base(exp, oinfo, oti,
428                                       oinfo->oi_cb_up, oinfo, rqset);
429 }
430
431 int osc_real_create(struct obd_export *exp, struct obdo *oa,
432                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
433 {
434         struct ptlrpc_request *req;
435         struct ost_body       *body;
436         struct lov_stripe_md  *lsm;
437         int                    rc;
438         ENTRY;
439
440         LASSERT(oa);
441         LASSERT(ea);
442
443         lsm = *ea;
444         if (!lsm) {
445                 rc = obd_alloc_memmd(exp, &lsm);
446                 if (rc < 0)
447                         RETURN(rc);
448         }
449
450         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
451         if (req == NULL)
452                 GOTO(out, rc = -ENOMEM);
453
454         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
455         if (rc) {
456                 ptlrpc_request_free(req);
457                 GOTO(out, rc);
458         }
459
460         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
461         LASSERT(body);
462         lustre_set_wire_obdo(&body->oa, oa);
463
464         ptlrpc_request_set_replen(req);
465
466         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
467             oa->o_flags == OBD_FL_DELORPHAN) {
468                 DEBUG_REQ(D_HA, req,
469                           "delorphan from OST integration");
470                 /* Don't resend the delorphan req */
471                 req->rq_no_resend = req->rq_no_delay = 1;
472         }
473
474         rc = ptlrpc_queue_wait(req);
475         if (rc)
476                 GOTO(out_req, rc);
477
478         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
479         if (body == NULL)
480                 GOTO(out_req, rc = -EPROTO);
481
482         lustre_get_wire_obdo(oa, &body->oa);
483
484         /* This should really be sent by the OST */
485         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
486         oa->o_valid |= OBD_MD_FLBLKSZ;
487
488         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
489          * have valid lsm_oinfo data structs, so don't go touching that.
490          * This needs to be fixed in a big way.
491          */
492         lsm->lsm_object_id = oa->o_id;
493         lsm->lsm_object_seq = oa->o_seq;
494         *ea = lsm;
495
496         if (oti != NULL) {
497                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
498
499                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
500                         if (!oti->oti_logcookies)
501                                 oti_alloc_cookies(oti, 1);
502                         *oti->oti_logcookies = oa->o_lcookie;
503                 }
504         }
505
506         CDEBUG(D_HA, "transno: "LPD64"\n",
507                lustre_msg_get_transno(req->rq_repmsg));
508 out_req:
509         ptlrpc_req_finished(req);
510 out:
511         if (rc && !*ea)
512                 obd_free_memmd(exp, &lsm);
513         RETURN(rc);
514 }
515
516 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
517                    obd_enqueue_update_f upcall, void *cookie,
518                    struct ptlrpc_request_set *rqset)
519 {
520         struct ptlrpc_request   *req;
521         struct osc_setattr_args *sa;
522         struct ost_body         *body;
523         int                      rc;
524         ENTRY;
525
526         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
527         if (req == NULL)
528                 RETURN(-ENOMEM);
529
530         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
531         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
532         if (rc) {
533                 ptlrpc_request_free(req);
534                 RETURN(rc);
535         }
536         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
537         ptlrpc_at_set_req_timeout(req);
538
539         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
540         LASSERT(body);
541         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
542         osc_pack_capa(req, body, oinfo->oi_capa);
543
544         ptlrpc_request_set_replen(req);
545
546
547         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
548         CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
549         sa = ptlrpc_req_async_args(req);
550         sa->sa_oa     = oinfo->oi_oa;
551         sa->sa_upcall = upcall;
552         sa->sa_cookie = cookie;
553         if (rqset == PTLRPCD_SET)
554                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
555         else
556                 ptlrpc_set_add_req(rqset, req);
557
558         RETURN(0);
559 }
560
561 static int osc_punch(const struct lu_env *env, struct obd_export *exp,
562                      struct obd_info *oinfo, struct obd_trans_info *oti,
563                      struct ptlrpc_request_set *rqset)
564 {
565         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
566         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
567         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
568         return osc_punch_base(exp, oinfo,
569                               oinfo->oi_cb_up, oinfo, rqset);
570 }
571
572 static int osc_sync_interpret(const struct lu_env *env,
573                               struct ptlrpc_request *req,
574                               void *arg, int rc)
575 {
576         struct osc_async_args *aa = arg;
577         struct ost_body *body;
578         ENTRY;
579
580         if (rc)
581                 GOTO(out, rc);
582
583         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
584         if (body == NULL) {
585                 CERROR ("can't unpack ost_body\n");
586                 GOTO(out, rc = -EPROTO);
587         }
588
589         *aa->aa_oi->oi_oa = body->oa;
590 out:
591         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
592         RETURN(rc);
593 }
594
595 static int osc_sync(const struct lu_env *env, struct obd_export *exp,
596                     struct obd_info *oinfo, obd_size start, obd_size end,
597                     struct ptlrpc_request_set *set)
598 {
599         struct ptlrpc_request *req;
600         struct ost_body       *body;
601         struct osc_async_args *aa;
602         int                    rc;
603         ENTRY;
604
605         if (!oinfo->oi_oa) {
606                 CDEBUG(D_INFO, "oa NULL\n");
607                 RETURN(-EINVAL);
608         }
609
610         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
611         if (req == NULL)
612                 RETURN(-ENOMEM);
613
614         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
615         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
616         if (rc) {
617                 ptlrpc_request_free(req);
618                 RETURN(rc);
619         }
620
621         /* overload the size and blocks fields in the oa with start/end */
622         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
623         LASSERT(body);
624         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
625         body->oa.o_size = start;
626         body->oa.o_blocks = end;
627         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
628         osc_pack_capa(req, body, oinfo->oi_capa);
629
630         ptlrpc_request_set_replen(req);
631         req->rq_interpret_reply = osc_sync_interpret;
632
633         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
634         aa = ptlrpc_req_async_args(req);
635         aa->aa_oi = oinfo;
636
637         ptlrpc_set_add_req(set, req);
638         RETURN (0);
639 }
640
641 /* Find and cancel locally locks matched by @mode in the resource found by
642  * @objid. Found locks are added into @cancel list. Returns the amount of
643  * locks added to @cancels list. */
644 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
645                                    cfs_list_t *cancels,
646                                    ldlm_mode_t mode, int lock_flags)
647 {
648         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
649         struct ldlm_res_id res_id;
650         struct ldlm_resource *res;
651         int count;
652         ENTRY;
653
654         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
655         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
656         if (res == NULL)
657                 RETURN(0);
658
659         LDLM_RESOURCE_ADDREF(res);
660         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
661                                            lock_flags, 0, NULL);
662         LDLM_RESOURCE_DELREF(res);
663         ldlm_resource_putref(res);
664         RETURN(count);
665 }
666
667 static int osc_destroy_interpret(const struct lu_env *env,
668                                  struct ptlrpc_request *req, void *data,
669                                  int rc)
670 {
671         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
672
673         cfs_atomic_dec(&cli->cl_destroy_in_flight);
674         cfs_waitq_signal(&cli->cl_destroy_waitq);
675         return 0;
676 }
677
678 static int osc_can_send_destroy(struct client_obd *cli)
679 {
680         if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
681             cli->cl_max_rpcs_in_flight) {
682                 /* The destroy request can be sent */
683                 return 1;
684         }
685         if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
686             cli->cl_max_rpcs_in_flight) {
687                 /*
688                  * The counter has been modified between the two atomic
689                  * operations.
690                  */
691                 cfs_waitq_signal(&cli->cl_destroy_waitq);
692         }
693         return 0;
694 }
695
696 /* Destroy requests can be async always on the client, and we don't even really
697  * care about the return code since the client cannot do anything at all about
698  * a destroy failure.
699  * When the MDS is unlinking a filename, it saves the file objects into a
700  * recovery llog, and these object records are cancelled when the OST reports
701  * they were destroyed and sync'd to disk (i.e. transaction committed).
702  * If the client dies, or the OST is down when the object should be destroyed,
703  * the records are not cancelled, and when the OST reconnects to the MDS next,
704  * it will retrieve the llog unlink logs and then sends the log cancellation
705  * cookies to the MDS after committing destroy transactions. */
706 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
707                        struct obdo *oa, struct lov_stripe_md *ea,
708                        struct obd_trans_info *oti, struct obd_export *md_export,
709                        void *capa)
710 {
711         struct client_obd     *cli = &exp->exp_obd->u.cli;
712         struct ptlrpc_request *req;
713         struct ost_body       *body;
714         CFS_LIST_HEAD(cancels);
715         int rc, count;
716         ENTRY;
717
718         if (!oa) {
719                 CDEBUG(D_INFO, "oa NULL\n");
720                 RETURN(-EINVAL);
721         }
722
723         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
724                                         LDLM_FL_DISCARD_DATA);
725
726         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
727         if (req == NULL) {
728                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
729                 RETURN(-ENOMEM);
730         }
731
732         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
733         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
734                                0, &cancels, count);
735         if (rc) {
736                 ptlrpc_request_free(req);
737                 RETURN(rc);
738         }
739
740         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
741         ptlrpc_at_set_req_timeout(req);
742
743         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
744                 oa->o_lcookie = *oti->oti_logcookies;
745         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
746         LASSERT(body);
747         lustre_set_wire_obdo(&body->oa, oa);
748
749         osc_pack_capa(req, body, (struct obd_capa *)capa);
750         ptlrpc_request_set_replen(req);
751
752         /* don't throttle destroy RPCs for the MDT */
753         if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
754                 req->rq_interpret_reply = osc_destroy_interpret;
755                 if (!osc_can_send_destroy(cli)) {
756                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
757                                                           NULL);
758
759                         /*
760                          * Wait until the number of on-going destroy RPCs drops
761                          * under max_rpc_in_flight
762                          */
763                         l_wait_event_exclusive(cli->cl_destroy_waitq,
764                                                osc_can_send_destroy(cli), &lwi);
765                 }
766         }
767
768         /* Do not wait for response */
769         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
770         RETURN(0);
771 }
772
773 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
774                                 long writing_bytes)
775 {
776         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
777
778         LASSERT(!(oa->o_valid & bits));
779
780         oa->o_valid |= bits;
781         client_obd_list_lock(&cli->cl_loi_list_lock);
782         oa->o_dirty = cli->cl_dirty;
783         if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
784                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
785                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
786                 oa->o_undirty = 0;
787         } else if (cfs_atomic_read(&obd_dirty_pages) -
788                    cfs_atomic_read(&obd_dirty_transit_pages) >
789                    obd_max_dirty_pages + 1){
790                 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
791                  * not covered by a lock thus they may safely race and trip
792                  * this CERROR() unless we add in a small fudge factor (+1). */
793                 CERROR("dirty %d - %d > system dirty_max %d\n",
794                        cfs_atomic_read(&obd_dirty_pages),
795                        cfs_atomic_read(&obd_dirty_transit_pages),
796                        obd_max_dirty_pages);
797                 oa->o_undirty = 0;
798         } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
799                 CERROR("dirty %lu - dirty_max %lu too big???\n",
800                        cli->cl_dirty, cli->cl_dirty_max);
801                 oa->o_undirty = 0;
802         } else {
803                 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
804                                 (cli->cl_max_rpcs_in_flight + 1);
805                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
806         }
807         oa->o_grant = cli->cl_avail_grant;
808         oa->o_dropped = cli->cl_lost_grant;
809         cli->cl_lost_grant = 0;
810         client_obd_list_unlock(&cli->cl_loi_list_lock);
811         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
812                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
813
814 }
815
816 static void osc_update_next_shrink(struct client_obd *cli)
817 {
818         cli->cl_next_shrink_grant =
819                 cfs_time_shift(cli->cl_grant_shrink_interval);
820         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
821                cli->cl_next_shrink_grant);
822 }
823
824 /* caller must hold loi_list_lock */
825 static void osc_consume_write_grant(struct client_obd *cli,
826                                     struct brw_page *pga)
827 {
828         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
829         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
830         cfs_atomic_inc(&obd_dirty_pages);
831         cli->cl_dirty += CFS_PAGE_SIZE;
832         cli->cl_avail_grant -= CFS_PAGE_SIZE;
833         pga->flag |= OBD_BRW_FROM_GRANT;
834         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
835                CFS_PAGE_SIZE, pga, pga->pg);
836         LASSERT(cli->cl_avail_grant >= 0);
837         osc_update_next_shrink(cli);
838 }
839
840 /* the companion to osc_consume_write_grant, called when a brw has completed.
841  * must be called with the loi lock held. */
842 static void osc_release_write_grant(struct client_obd *cli,
843                                     struct brw_page *pga, int sent)
844 {
845         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
846         ENTRY;
847
848         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
849         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
850                 EXIT;
851                 return;
852         }
853
854         pga->flag &= ~OBD_BRW_FROM_GRANT;
855         cfs_atomic_dec(&obd_dirty_pages);
856         cli->cl_dirty -= CFS_PAGE_SIZE;
857         if (pga->flag & OBD_BRW_NOCACHE) {
858                 pga->flag &= ~OBD_BRW_NOCACHE;
859                 cfs_atomic_dec(&obd_dirty_transit_pages);
860                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
861         }
862         if (!sent) {
863                 /* Reclaim grant from truncated pages. This is used to solve
864                  * write-truncate and grant all gone(to lost_grant) problem.
865                  * For a vfs write this problem can be easily solved by a sync
866                  * write, however, this is not an option for page_mkwrite()
867                  * because grant has to be allocated before a page becomes
868                  * dirty. */
869                 if (cli->cl_avail_grant < PTLRPC_MAX_BRW_SIZE)
870                         cli->cl_avail_grant += CFS_PAGE_SIZE;
871                 else
872                         cli->cl_lost_grant += CFS_PAGE_SIZE;
873                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
874                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
875         } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
876                 /* For short writes we shouldn't count parts of pages that
877                  * span a whole block on the OST side, or our accounting goes
878                  * wrong.  Should match the code in filter_grant_check. */
879                 int offset = pga->off & ~CFS_PAGE_MASK;
880                 int count = pga->count + (offset & (blocksize - 1));
881                 int end = (offset + pga->count) & (blocksize - 1);
882                 if (end)
883                         count += blocksize - end;
884
885                 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
886                 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
887                        CFS_PAGE_SIZE - count, cli->cl_lost_grant,
888                        cli->cl_avail_grant, cli->cl_dirty);
889         }
890
891         EXIT;
892 }
893
894 static unsigned long rpcs_in_flight(struct client_obd *cli)
895 {
896         return cli->cl_r_in_flight + cli->cl_w_in_flight;
897 }
898
899 /* caller must hold loi_list_lock */
900 void osc_wake_cache_waiters(struct client_obd *cli)
901 {
902         cfs_list_t *l, *tmp;
903         struct osc_cache_waiter *ocw;
904
905         ENTRY;
906         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
907                 /* if we can't dirty more, we must wait until some is written */
908                 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
909                    (cfs_atomic_read(&obd_dirty_pages) + 1 >
910                     obd_max_dirty_pages)) {
911                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
912                                "osc max %ld, sys max %d\n", cli->cl_dirty,
913                                cli->cl_dirty_max, obd_max_dirty_pages);
914                         return;
915                 }
916
917                 /* if still dirty cache but no grant wait for pending RPCs that
918                  * may yet return us some grant before doing sync writes */
919                 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
920                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
921                                cli->cl_w_in_flight);
922                         return;
923                 }
924
925                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
926                 cfs_list_del_init(&ocw->ocw_entry);
927                 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
928                         /* no more RPCs in flight to return grant, do sync IO */
929                         ocw->ocw_rc = -EDQUOT;
930                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
931                 } else {
932                         osc_consume_write_grant(cli,
933                                                 &ocw->ocw_oap->oap_brw_page);
934                 }
935
936                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld\n",
937                        ocw, ocw->ocw_oap, cli->cl_avail_grant);
938
939                 cfs_waitq_signal(&ocw->ocw_waitq);
940         }
941
942         EXIT;
943 }
944
945 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
946 {
947         client_obd_list_lock(&cli->cl_loi_list_lock);
948         cli->cl_avail_grant += grant;
949         client_obd_list_unlock(&cli->cl_loi_list_lock);
950 }
951
952 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
953 {
954         if (body->oa.o_valid & OBD_MD_FLGRANT) {
955                 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
956                 __osc_update_grant(cli, body->oa.o_grant);
957         }
958 }
959
960 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
961                               obd_count keylen, void *key, obd_count vallen,
962                               void *val, struct ptlrpc_request_set *set);
963
964 static int osc_shrink_grant_interpret(const struct lu_env *env,
965                                       struct ptlrpc_request *req,
966                                       void *aa, int rc)
967 {
968         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
969         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
970         struct ost_body *body;
971
972         if (rc != 0) {
973                 __osc_update_grant(cli, oa->o_grant);
974                 GOTO(out, rc);
975         }
976
977         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
978         LASSERT(body);
979         osc_update_grant(cli, body);
980 out:
981         OBDO_FREE(oa);
982         return rc;
983 }
984
985 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
986 {
987         client_obd_list_lock(&cli->cl_loi_list_lock);
988         oa->o_grant = cli->cl_avail_grant / 4;
989         cli->cl_avail_grant -= oa->o_grant;
990         client_obd_list_unlock(&cli->cl_loi_list_lock);
991         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
992                 oa->o_valid |= OBD_MD_FLFLAGS;
993                 oa->o_flags = 0;
994         }
995         oa->o_flags |= OBD_FL_SHRINK_GRANT;
996         osc_update_next_shrink(cli);
997 }
998
999 /* Shrink the current grant, either from some large amount to enough for a
1000  * full set of in-flight RPCs, or if we have already shrunk to that limit
1001  * then to enough for a single RPC.  This avoids keeping more grant than
1002  * needed, and avoids shrinking the grant piecemeal. */
1003 static int osc_shrink_grant(struct client_obd *cli)
1004 {
1005         long target = (cli->cl_max_rpcs_in_flight + 1) *
1006                       cli->cl_max_pages_per_rpc;
1007
1008         client_obd_list_lock(&cli->cl_loi_list_lock);
1009         if (cli->cl_avail_grant <= target)
1010                 target = cli->cl_max_pages_per_rpc;
1011         client_obd_list_unlock(&cli->cl_loi_list_lock);
1012
1013         return osc_shrink_grant_to_target(cli, target);
1014 }
1015
1016 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
1017 {
1018         int    rc = 0;
1019         struct ost_body     *body;
1020         ENTRY;
1021
1022         client_obd_list_lock(&cli->cl_loi_list_lock);
1023         /* Don't shrink if we are already above or below the desired limit
1024          * We don't want to shrink below a single RPC, as that will negatively
1025          * impact block allocation and long-term performance. */
1026         if (target < cli->cl_max_pages_per_rpc)
1027                 target = cli->cl_max_pages_per_rpc;
1028
1029         if (target >= cli->cl_avail_grant) {
1030                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1031                 RETURN(0);
1032         }
1033         client_obd_list_unlock(&cli->cl_loi_list_lock);
1034
1035         OBD_ALLOC_PTR(body);
1036         if (!body)
1037                 RETURN(-ENOMEM);
1038
1039         osc_announce_cached(cli, &body->oa, 0);
1040
1041         client_obd_list_lock(&cli->cl_loi_list_lock);
1042         body->oa.o_grant = cli->cl_avail_grant - target;
1043         cli->cl_avail_grant = target;
1044         client_obd_list_unlock(&cli->cl_loi_list_lock);
1045         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
1046                 body->oa.o_valid |= OBD_MD_FLFLAGS;
1047                 body->oa.o_flags = 0;
1048         }
1049         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1050         osc_update_next_shrink(cli);
1051
1052         rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
1053                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1054                                 sizeof(*body), body, NULL);
1055         if (rc != 0)
1056                 __osc_update_grant(cli, body->oa.o_grant);
1057         OBD_FREE_PTR(body);
1058         RETURN(rc);
1059 }
1060
1061 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1062 static int osc_should_shrink_grant(struct client_obd *client)
1063 {
1064         cfs_time_t time = cfs_time_current();
1065         cfs_time_t next_shrink = client->cl_next_shrink_grant;
1066
1067         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1068              OBD_CONNECT_GRANT_SHRINK) == 0)
1069                 return 0;
1070
1071         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1072                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1073                     client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1074                         return 1;
1075                 else
1076                         osc_update_next_shrink(client);
1077         }
1078         return 0;
1079 }
1080
1081 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1082 {
1083         struct client_obd *client;
1084
1085         cfs_list_for_each_entry(client, &item->ti_obd_list,
1086                                 cl_grant_shrink_list) {
1087                 if (osc_should_shrink_grant(client))
1088                         osc_shrink_grant(client);
1089         }
1090         return 0;
1091 }
1092
1093 static int osc_add_shrink_grant(struct client_obd *client)
1094 {
1095         int rc;
1096
1097         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1098                                        TIMEOUT_GRANT,
1099                                        osc_grant_shrink_grant_cb, NULL,
1100                                        &client->cl_grant_shrink_list);
1101         if (rc) {
1102                 CERROR("add grant client %s error %d\n",
1103                         client->cl_import->imp_obd->obd_name, rc);
1104                 return rc;
1105         }
1106         CDEBUG(D_CACHE, "add grant client %s \n",
1107                client->cl_import->imp_obd->obd_name);
1108         osc_update_next_shrink(client);
1109         return 0;
1110 }
1111
1112 static int osc_del_shrink_grant(struct client_obd *client)
1113 {
1114         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1115                                          TIMEOUT_GRANT);
1116 }
1117
1118 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1119 {
1120         /*
1121          * ocd_grant is the total grant amount we're expect to hold: if we've
1122          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1123          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1124          *
1125          * race is tolerable here: if we're evicted, but imp_state already
1126          * left EVICTED state, then cl_dirty must be 0 already.
1127          */
1128         client_obd_list_lock(&cli->cl_loi_list_lock);
1129         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1130                 cli->cl_avail_grant = ocd->ocd_grant;
1131         else
1132                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1133
1134         if (cli->cl_avail_grant < 0) {
1135                 CWARN("%s: available grant < 0, the OSS is probably not running"
1136                       " with patch from bug20278 (%ld) \n",
1137                       cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1138                 /* workaround for 1.6 servers which do not have
1139                  * the patch from bug20278 */
1140                 cli->cl_avail_grant = ocd->ocd_grant;
1141         }
1142
1143         client_obd_list_unlock(&cli->cl_loi_list_lock);
1144
1145         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1146                cli->cl_import->imp_obd->obd_name,
1147                cli->cl_avail_grant, cli->cl_lost_grant);
1148
1149         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1150             cfs_list_empty(&cli->cl_grant_shrink_list))
1151                 osc_add_shrink_grant(cli);
1152 }
1153
1154 /* We assume that the reason this OSC got a short read is because it read
1155  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1156  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1157  * this stripe never got written at or beyond this stripe offset yet. */
1158 static void handle_short_read(int nob_read, obd_count page_count,
1159                               struct brw_page **pga)
1160 {
1161         char *ptr;
1162         int i = 0;
1163
1164         /* skip bytes read OK */
1165         while (nob_read > 0) {
1166                 LASSERT (page_count > 0);
1167
1168                 if (pga[i]->count > nob_read) {
1169                         /* EOF inside this page */
1170                         ptr = cfs_kmap(pga[i]->pg) +
1171                                 (pga[i]->off & ~CFS_PAGE_MASK);
1172                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1173                         cfs_kunmap(pga[i]->pg);
1174                         page_count--;
1175                         i++;
1176                         break;
1177                 }
1178
1179                 nob_read -= pga[i]->count;
1180                 page_count--;
1181                 i++;
1182         }
1183
1184         /* zero remaining pages */
1185         while (page_count-- > 0) {
1186                 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1187                 memset(ptr, 0, pga[i]->count);
1188                 cfs_kunmap(pga[i]->pg);
1189                 i++;
1190         }
1191 }
1192
1193 static int check_write_rcs(struct ptlrpc_request *req,
1194                            int requested_nob, int niocount,
1195                            obd_count page_count, struct brw_page **pga)
1196 {
1197         int     i;
1198         __u32   *remote_rcs;
1199
1200         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1201                                                   sizeof(*remote_rcs) *
1202                                                   niocount);
1203         if (remote_rcs == NULL) {
1204                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1205                 return(-EPROTO);
1206         }
1207
1208         /* return error if any niobuf was in error */
1209         for (i = 0; i < niocount; i++) {
1210                 if ((int)remote_rcs[i] < 0)
1211                         return(remote_rcs[i]);
1212
1213                 if (remote_rcs[i] != 0) {
1214                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1215                                 i, remote_rcs[i], req);
1216                         return(-EPROTO);
1217                 }
1218         }
1219
1220         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1221                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1222                        req->rq_bulk->bd_nob_transferred, requested_nob);
1223                 return(-EPROTO);
1224         }
1225
1226         return (0);
1227 }
1228
1229 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1230 {
1231         if (p1->flag != p2->flag) {
1232                 unsigned mask = ~(OBD_BRW_FROM_GRANT| OBD_BRW_NOCACHE|
1233                                   OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1234
1235                 /* warn if we try to combine flags that we don't know to be
1236                  * safe to combine */
1237                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1238                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1239                               "report this at http://bugs.whamcloud.com/\n",
1240                               p1->flag, p2->flag);
1241                 }
1242                 return 0;
1243         }
1244
1245         return (p1->off + p1->count == p2->off);
1246 }
1247
1248 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1249                                    struct brw_page **pga, int opc,
1250                                    cksum_type_t cksum_type)
1251 {
1252         __u32 cksum;
1253         int i = 0;
1254
1255         LASSERT (pg_count > 0);
1256         cksum = init_checksum(cksum_type);
1257         while (nob > 0 && pg_count > 0) {
1258                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1259                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1260                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1261
1262                 /* corrupt the data before we compute the checksum, to
1263                  * simulate an OST->client data error */
1264                 if (i == 0 && opc == OST_READ &&
1265                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1266                         memcpy(ptr + off, "bad1", min(4, nob));
1267                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1268                 cfs_kunmap(pga[i]->pg);
1269                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1270                                off, cksum);
1271
1272                 nob -= pga[i]->count;
1273                 pg_count--;
1274                 i++;
1275         }
1276         /* For sending we only compute the wrong checksum instead
1277          * of corrupting the data so it is still correct on a redo */
1278         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1279                 cksum++;
1280
1281         return fini_checksum(cksum, cksum_type);
1282 }
1283
1284 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1285                                 struct lov_stripe_md *lsm, obd_count page_count,
1286                                 struct brw_page **pga,
1287                                 struct ptlrpc_request **reqp,
1288                                 struct obd_capa *ocapa, int reserve,
1289                                 int resend)
1290 {
1291         struct ptlrpc_request   *req;
1292         struct ptlrpc_bulk_desc *desc;
1293         struct ost_body         *body;
1294         struct obd_ioobj        *ioobj;
1295         struct niobuf_remote    *niobuf;
1296         int niocount, i, requested_nob, opc, rc;
1297         struct osc_brw_async_args *aa;
1298         struct req_capsule      *pill;
1299         struct brw_page *pg_prev;
1300
1301         ENTRY;
1302         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1303                 RETURN(-ENOMEM); /* Recoverable */
1304         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1305                 RETURN(-EINVAL); /* Fatal */
1306
1307         if ((cmd & OBD_BRW_WRITE) != 0) {
1308                 opc = OST_WRITE;
1309                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1310                                                 cli->cl_import->imp_rq_pool,
1311                                                 &RQF_OST_BRW_WRITE);
1312         } else {
1313                 opc = OST_READ;
1314                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1315         }
1316         if (req == NULL)
1317                 RETURN(-ENOMEM);
1318
1319         for (niocount = i = 1; i < page_count; i++) {
1320                 if (!can_merge_pages(pga[i - 1], pga[i]))
1321                         niocount++;
1322         }
1323
1324         pill = &req->rq_pill;
1325         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1326                              sizeof(*ioobj));
1327         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1328                              niocount * sizeof(*niobuf));
1329         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1330
1331         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1332         if (rc) {
1333                 ptlrpc_request_free(req);
1334                 RETURN(rc);
1335         }
1336         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1337         ptlrpc_at_set_req_timeout(req);
1338
1339         if (opc == OST_WRITE)
1340                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1341                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1342         else
1343                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1344                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1345
1346         if (desc == NULL)
1347                 GOTO(out, rc = -ENOMEM);
1348         /* NB request now owns desc and will free it when it gets freed */
1349
1350         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1351         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1352         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1353         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1354
1355         lustre_set_wire_obdo(&body->oa, oa);
1356
1357         obdo_to_ioobj(oa, ioobj);
1358         ioobj->ioo_bufcnt = niocount;
1359         osc_pack_capa(req, body, ocapa);
1360         LASSERT (page_count > 0);
1361         pg_prev = pga[0];
1362         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1363                 struct brw_page *pg = pga[i];
1364                 int poff = pg->off & ~CFS_PAGE_MASK;
1365
1366                 LASSERT(pg->count > 0);
1367                 /* make sure there is no gap in the middle of page array */
1368                 LASSERTF(page_count == 1 ||
1369                          (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1370                           ergo(i > 0 && i < page_count - 1,
1371                                poff == 0 && pg->count == CFS_PAGE_SIZE)   &&
1372                           ergo(i == page_count - 1, poff == 0)),
1373                          "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1374                          i, page_count, pg, pg->off, pg->count);
1375 #ifdef __linux__
1376                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1377                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1378                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1379                          i, page_count,
1380                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1381                          pg_prev->pg, page_private(pg_prev->pg),
1382                          pg_prev->pg->index, pg_prev->off);
1383 #else
1384                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1385                          "i %d p_c %u\n", i, page_count);
1386 #endif
1387                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1388                         (pg->flag & OBD_BRW_SRVLOCK));
1389
1390                 ptlrpc_prep_bulk_page(desc, pg->pg, poff, pg->count);
1391                 requested_nob += pg->count;
1392
1393                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1394                         niobuf--;
1395                         niobuf->len += pg->count;
1396                 } else {
1397                         niobuf->offset = pg->off;
1398                         niobuf->len    = pg->count;
1399                         niobuf->flags  = pg->flag;
1400                 }
1401                 pg_prev = pg;
1402         }
1403
1404         LASSERTF((void *)(niobuf - niocount) ==
1405                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1406                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1407                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1408
1409         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1410         if (resend) {
1411                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1412                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1413                         body->oa.o_flags = 0;
1414                 }
1415                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1416         }
1417
1418         if (osc_should_shrink_grant(cli))
1419                 osc_shrink_grant_local(cli, &body->oa);
1420
1421         /* size[REQ_REC_OFF] still sizeof (*body) */
1422         if (opc == OST_WRITE) {
1423                 if (cli->cl_checksum &&
1424                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1425                         /* store cl_cksum_type in a local variable since
1426                          * it can be changed via lprocfs */
1427                         cksum_type_t cksum_type = cli->cl_cksum_type;
1428
1429                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1430                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1431                                 body->oa.o_flags = 0;
1432                         }
1433                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1434                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1435                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1436                                                              page_count, pga,
1437                                                              OST_WRITE,
1438                                                              cksum_type);
1439                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1440                                body->oa.o_cksum);
1441                         /* save this in 'oa', too, for later checking */
1442                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1443                         oa->o_flags |= cksum_type_pack(cksum_type);
1444                 } else {
1445                         /* clear out the checksum flag, in case this is a
1446                          * resend but cl_checksum is no longer set. b=11238 */
1447                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1448                 }
1449                 oa->o_cksum = body->oa.o_cksum;
1450                 /* 1 RC per niobuf */
1451                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1452                                      sizeof(__u32) * niocount);
1453         } else {
1454                 if (cli->cl_checksum &&
1455                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1456                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1457                                 body->oa.o_flags = 0;
1458                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1459                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1460                 }
1461         }
1462         ptlrpc_request_set_replen(req);
1463
1464         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1465         aa = ptlrpc_req_async_args(req);
1466         aa->aa_oa = oa;
1467         aa->aa_requested_nob = requested_nob;
1468         aa->aa_nio_count = niocount;
1469         aa->aa_page_count = page_count;
1470         aa->aa_resends = 0;
1471         aa->aa_ppga = pga;
1472         aa->aa_cli = cli;
1473         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1474         if (ocapa && reserve)
1475                 aa->aa_ocapa = capa_get(ocapa);
1476
1477         *reqp = req;
1478         RETURN(0);
1479
1480  out:
1481         ptlrpc_req_finished(req);
1482         RETURN(rc);
1483 }
1484
1485 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1486                                 __u32 client_cksum, __u32 server_cksum, int nob,
1487                                 obd_count page_count, struct brw_page **pga,
1488                                 cksum_type_t client_cksum_type)
1489 {
1490         __u32 new_cksum;
1491         char *msg;
1492         cksum_type_t cksum_type;
1493
1494         if (server_cksum == client_cksum) {
1495                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1496                 return 0;
1497         }
1498
1499         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1500                                        oa->o_flags : 0);
1501         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1502                                       cksum_type);
1503
1504         if (cksum_type != client_cksum_type)
1505                 msg = "the server did not use the checksum type specified in "
1506                       "the original request - likely a protocol problem";
1507         else if (new_cksum == server_cksum)
1508                 msg = "changed on the client after we checksummed it - "
1509                       "likely false positive due to mmap IO (bug 11742)";
1510         else if (new_cksum == client_cksum)
1511                 msg = "changed in transit before arrival at OST";
1512         else
1513                 msg = "changed in transit AND doesn't match the original - "
1514                       "likely false positive due to mmap IO (bug 11742)";
1515
1516         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1517                            " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1518                            msg, libcfs_nid2str(peer->nid),
1519                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1520                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1521                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1522                            oa->o_id,
1523                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1524                            pga[0]->off,
1525                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1526         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1527                "client csum now %x\n", client_cksum, client_cksum_type,
1528                server_cksum, cksum_type, new_cksum);
1529         return 1;
1530 }
1531
1532 /* Note rc enters this function as number of bytes transferred */
1533 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1534 {
1535         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1536         const lnet_process_id_t *peer =
1537                         &req->rq_import->imp_connection->c_peer;
1538         struct client_obd *cli = aa->aa_cli;
1539         struct ost_body *body;
1540         __u32 client_cksum = 0;
1541         ENTRY;
1542
1543         if (rc < 0 && rc != -EDQUOT) {
1544                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1545                 RETURN(rc);
1546         }
1547
1548         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1549         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1550         if (body == NULL) {
1551                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1552                 RETURN(-EPROTO);
1553         }
1554
1555         /* set/clear over quota flag for a uid/gid */
1556         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1557             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1558                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1559
1560                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1561                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1562                        body->oa.o_flags);
1563                 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1564         }
1565
1566         osc_update_grant(cli, body);
1567
1568         if (rc < 0)
1569                 RETURN(rc);
1570
1571         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1572                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1573
1574         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1575                 if (rc > 0) {
1576                         CERROR("Unexpected +ve rc %d\n", rc);
1577                         RETURN(-EPROTO);
1578                 }
1579                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1580
1581                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1582                         RETURN(-EAGAIN);
1583
1584                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1585                     check_write_checksum(&body->oa, peer, client_cksum,
1586                                          body->oa.o_cksum, aa->aa_requested_nob,
1587                                          aa->aa_page_count, aa->aa_ppga,
1588                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1589                         RETURN(-EAGAIN);
1590
1591                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1592                                      aa->aa_page_count, aa->aa_ppga);
1593                 GOTO(out, rc);
1594         }
1595
1596         /* The rest of this function executes only for OST_READs */
1597
1598         /* if unwrap_bulk failed, return -EAGAIN to retry */
1599         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1600         if (rc < 0)
1601                 GOTO(out, rc = -EAGAIN);
1602
1603         if (rc > aa->aa_requested_nob) {
1604                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1605                        aa->aa_requested_nob);
1606                 RETURN(-EPROTO);
1607         }
1608
1609         if (rc != req->rq_bulk->bd_nob_transferred) {
1610                 CERROR ("Unexpected rc %d (%d transferred)\n",
1611                         rc, req->rq_bulk->bd_nob_transferred);
1612                 return (-EPROTO);
1613         }
1614
1615         if (rc < aa->aa_requested_nob)
1616                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1617
1618         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1619                 static int cksum_counter;
1620                 __u32      server_cksum = body->oa.o_cksum;
1621                 char      *via;
1622                 char      *router;
1623                 cksum_type_t cksum_type;
1624
1625                 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1626                                                body->oa.o_flags : 0);
1627                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1628                                                  aa->aa_ppga, OST_READ,
1629                                                  cksum_type);
1630
1631                 if (peer->nid == req->rq_bulk->bd_sender) {
1632                         via = router = "";
1633                 } else {
1634                         via = " via ";
1635                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1636                 }
1637
1638                 if (server_cksum == ~0 && rc > 0) {
1639                         CERROR("Protocol error: server %s set the 'checksum' "
1640                                "bit, but didn't send a checksum.  Not fatal, "
1641                                "but please notify on http://bugs.whamcloud.com/\n",
1642                                libcfs_nid2str(peer->nid));
1643                 } else if (server_cksum != client_cksum) {
1644                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1645                                            "%s%s%s inode "DFID" object "
1646                                            LPU64"/"LPU64" extent "
1647                                            "["LPU64"-"LPU64"]\n",
1648                                            req->rq_import->imp_obd->obd_name,
1649                                            libcfs_nid2str(peer->nid),
1650                                            via, router,
1651                                            body->oa.o_valid & OBD_MD_FLFID ?
1652                                                 body->oa.o_parent_seq : (__u64)0,
1653                                            body->oa.o_valid & OBD_MD_FLFID ?
1654                                                 body->oa.o_parent_oid : 0,
1655                                            body->oa.o_valid & OBD_MD_FLFID ?
1656                                                 body->oa.o_parent_ver : 0,
1657                                            body->oa.o_id,
1658                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1659                                                 body->oa.o_seq : (__u64)0,
1660                                            aa->aa_ppga[0]->off,
1661                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1662                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1663                                                                         1);
1664                         CERROR("client %x, server %x, cksum_type %x\n",
1665                                client_cksum, server_cksum, cksum_type);
1666                         cksum_counter = 0;
1667                         aa->aa_oa->o_cksum = client_cksum;
1668                         rc = -EAGAIN;
1669                 } else {
1670                         cksum_counter++;
1671                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1672                         rc = 0;
1673                 }
1674         } else if (unlikely(client_cksum)) {
1675                 static int cksum_missed;
1676
1677                 cksum_missed++;
1678                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1679                         CERROR("Checksum %u requested from %s but not sent\n",
1680                                cksum_missed, libcfs_nid2str(peer->nid));
1681         } else {
1682                 rc = 0;
1683         }
1684 out:
1685         if (rc >= 0)
1686                 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1687
1688         RETURN(rc);
1689 }
1690
1691 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1692                             struct lov_stripe_md *lsm,
1693                             obd_count page_count, struct brw_page **pga,
1694                             struct obd_capa *ocapa)
1695 {
1696         struct ptlrpc_request *req;
1697         int                    rc;
1698         cfs_waitq_t            waitq;
1699         int                    generation, resends = 0;
1700         struct l_wait_info     lwi;
1701
1702         ENTRY;
1703
1704         cfs_waitq_init(&waitq);
1705         generation = exp->exp_obd->u.cli.cl_import->imp_generation;
1706
1707 restart_bulk:
1708         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1709                                   page_count, pga, &req, ocapa, 0, resends);
1710         if (rc != 0)
1711                 return (rc);
1712
1713         if (resends) {
1714                 req->rq_generation_set = 1;
1715                 req->rq_import_generation = generation;
1716                 req->rq_sent = cfs_time_current_sec() + resends;
1717         }
1718
1719         rc = ptlrpc_queue_wait(req);
1720
1721         if (rc == -ETIMEDOUT && req->rq_resend) {
1722                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1723                 ptlrpc_req_finished(req);
1724                 goto restart_bulk;
1725         }
1726
1727         rc = osc_brw_fini_request(req, rc);
1728
1729         ptlrpc_req_finished(req);
1730         /* When server return -EINPROGRESS, client should always retry
1731          * regardless of the number of times the bulk was resent already.*/
1732         if (osc_recoverable_error(rc)) {
1733                 resends++;
1734                 if (rc != -EINPROGRESS &&
1735                     !client_should_resend(resends, &exp->exp_obd->u.cli)) {
1736                         CERROR("%s: too many resend retries for object: "
1737                                ""LPU64":"LPU64", rc = %d.\n",
1738                                exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1739                         goto out;
1740                 }
1741                 if (generation !=
1742                     exp->exp_obd->u.cli.cl_import->imp_generation) {
1743                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
1744                                ""LPU64":"LPU64", rc = %d.\n",
1745                                exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1746                         goto out;
1747                 }
1748
1749                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1750                                        NULL);
1751                 l_wait_event(waitq, 0, &lwi);
1752
1753                 goto restart_bulk;
1754         }
1755 out:
1756         if (rc == -EAGAIN || rc == -EINPROGRESS)
1757                 rc = -EIO;
1758         RETURN (rc);
1759 }
1760
1761 int osc_brw_redo_request(struct ptlrpc_request *request,
1762                          struct osc_brw_async_args *aa)
1763 {
1764         struct ptlrpc_request *new_req;
1765         struct ptlrpc_request_set *set = request->rq_set;
1766         struct osc_brw_async_args *new_aa;
1767         struct osc_async_page *oap;
1768         int rc = 0;
1769         ENTRY;
1770
1771         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1772
1773         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1774                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1775                                   aa->aa_cli, aa->aa_oa,
1776                                   NULL /* lsm unused by osc currently */,
1777                                   aa->aa_page_count, aa->aa_ppga,
1778                                   &new_req, aa->aa_ocapa, 0, 1);
1779         if (rc)
1780                 RETURN(rc);
1781
1782         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1783
1784         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1785                 if (oap->oap_request != NULL) {
1786                         LASSERTF(request == oap->oap_request,
1787                                  "request %p != oap_request %p\n",
1788                                  request, oap->oap_request);
1789                         if (oap->oap_interrupted) {
1790                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1791                                 ptlrpc_req_finished(new_req);
1792                                 RETURN(-EINTR);
1793                         }
1794                 }
1795         }
1796         /* New request takes over pga and oaps from old request.
1797          * Note that copying a list_head doesn't work, need to move it... */
1798         aa->aa_resends++;
1799         new_req->rq_interpret_reply = request->rq_interpret_reply;
1800         new_req->rq_async_args = request->rq_async_args;
1801         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1802         new_req->rq_generation_set = 1;
1803         new_req->rq_import_generation = request->rq_import_generation;
1804
1805         new_aa = ptlrpc_req_async_args(new_req);
1806
1807         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1808         cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1809         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1810
1811         cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1812                 if (oap->oap_request) {
1813                         ptlrpc_req_finished(oap->oap_request);
1814                         oap->oap_request = ptlrpc_request_addref(new_req);
1815                 }
1816         }
1817
1818         new_aa->aa_ocapa = aa->aa_ocapa;
1819         aa->aa_ocapa = NULL;
1820
1821         /* use ptlrpc_set_add_req is safe because interpret functions work
1822          * in check_set context. only one way exist with access to request
1823          * from different thread got -EINTR - this way protected with
1824          * cl_loi_list_lock */
1825         ptlrpc_set_add_req(set, new_req);
1826
1827         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1828
1829         DEBUG_REQ(D_INFO, new_req, "new request");
1830         RETURN(0);
1831 }
1832
1833 /*
1834  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1835  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1836  * fine for our small page arrays and doesn't require allocation.  its an
1837  * insertion sort that swaps elements that are strides apart, shrinking the
1838  * stride down until its '1' and the array is sorted.
1839  */
1840 static void sort_brw_pages(struct brw_page **array, int num)
1841 {
1842         int stride, i, j;
1843         struct brw_page *tmp;
1844
1845         if (num == 1)
1846                 return;
1847         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1848                 ;
1849
1850         do {
1851                 stride /= 3;
1852                 for (i = stride ; i < num ; i++) {
1853                         tmp = array[i];
1854                         j = i;
1855                         while (j >= stride && array[j - stride]->off > tmp->off) {
1856                                 array[j] = array[j - stride];
1857                                 j -= stride;
1858                         }
1859                         array[j] = tmp;
1860                 }
1861         } while (stride > 1);
1862 }
1863
1864 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1865 {
1866         int count = 1;
1867         int offset;
1868         int i = 0;
1869
1870         LASSERT (pages > 0);
1871         offset = pg[i]->off & ~CFS_PAGE_MASK;
1872
1873         for (;;) {
1874                 pages--;
1875                 if (pages == 0)         /* that's all */
1876                         return count;
1877
1878                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1879                         return count;   /* doesn't end on page boundary */
1880
1881                 i++;
1882                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1883                 if (offset != 0)        /* doesn't start on page boundary */
1884                         return count;
1885
1886                 count++;
1887         }
1888 }
1889
1890 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1891 {
1892         struct brw_page **ppga;
1893         int i;
1894
1895         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1896         if (ppga == NULL)
1897                 return NULL;
1898
1899         for (i = 0; i < count; i++)
1900                 ppga[i] = pga + i;
1901         return ppga;
1902 }
1903
1904 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1905 {
1906         LASSERT(ppga != NULL);
1907         OBD_FREE(ppga, sizeof(*ppga) * count);
1908 }
1909
1910 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1911                    obd_count page_count, struct brw_page *pga,
1912                    struct obd_trans_info *oti)
1913 {
1914         struct obdo *saved_oa = NULL;
1915         struct brw_page **ppga, **orig;
1916         struct obd_import *imp = class_exp2cliimp(exp);
1917         struct client_obd *cli;
1918         int rc, page_count_orig;
1919         ENTRY;
1920
1921         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1922         cli = &imp->imp_obd->u.cli;
1923
1924         if (cmd & OBD_BRW_CHECK) {
1925                 /* The caller just wants to know if there's a chance that this
1926                  * I/O can succeed */
1927
1928                 if (imp->imp_invalid)
1929                         RETURN(-EIO);
1930                 RETURN(0);
1931         }
1932
1933         /* test_brw with a failed create can trip this, maybe others. */
1934         LASSERT(cli->cl_max_pages_per_rpc);
1935
1936         rc = 0;
1937
1938         orig = ppga = osc_build_ppga(pga, page_count);
1939         if (ppga == NULL)
1940                 RETURN(-ENOMEM);
1941         page_count_orig = page_count;
1942
1943         sort_brw_pages(ppga, page_count);
1944         while (page_count) {
1945                 obd_count pages_per_brw;
1946
1947                 if (page_count > cli->cl_max_pages_per_rpc)
1948                         pages_per_brw = cli->cl_max_pages_per_rpc;
1949                 else
1950                         pages_per_brw = page_count;
1951
1952                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1953
1954                 if (saved_oa != NULL) {
1955                         /* restore previously saved oa */
1956                         *oinfo->oi_oa = *saved_oa;
1957                 } else if (page_count > pages_per_brw) {
1958                         /* save a copy of oa (brw will clobber it) */
1959                         OBDO_ALLOC(saved_oa);
1960                         if (saved_oa == NULL)
1961                                 GOTO(out, rc = -ENOMEM);
1962                         *saved_oa = *oinfo->oi_oa;
1963                 }
1964
1965                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1966                                       pages_per_brw, ppga, oinfo->oi_capa);
1967
1968                 if (rc != 0)
1969                         break;
1970
1971                 page_count -= pages_per_brw;
1972                 ppga += pages_per_brw;
1973         }
1974
1975 out:
1976         osc_release_ppga(orig, page_count_orig);
1977
1978         if (saved_oa != NULL)
1979                 OBDO_FREE(saved_oa);
1980
1981         RETURN(rc);
1982 }
1983
1984 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1985  * the dirty accounting.  Writeback completes or truncate happens before
1986  * writing starts.  Must be called with the loi lock held. */
1987 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1988                            int sent)
1989 {
1990         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1991 }
1992
1993
1994 /* This maintains the lists of pending pages to read/write for a given object
1995  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1996  * to quickly find objects that are ready to send an RPC. */
1997 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1998                          int cmd)
1999 {
2000         ENTRY;
2001
2002         if (lop->lop_num_pending == 0)
2003                 RETURN(0);
2004
2005         /* if we have an invalid import we want to drain the queued pages
2006          * by forcing them through rpcs that immediately fail and complete
2007          * the pages.  recovery relies on this to empty the queued pages
2008          * before canceling the locks and evicting down the llite pages */
2009         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2010                 RETURN(1);
2011
2012         /* stream rpcs in queue order as long as as there is an urgent page
2013          * queued.  this is our cheap solution for good batching in the case
2014          * where writepage marks some random page in the middle of the file
2015          * as urgent because of, say, memory pressure */
2016         if (!cfs_list_empty(&lop->lop_urgent)) {
2017                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
2018                 RETURN(1);
2019         }
2020
2021         if (cmd & OBD_BRW_WRITE) {
2022                 /* trigger a write rpc stream as long as there are dirtiers
2023                  * waiting for space.  as they're waiting, they're not going to
2024                  * create more pages to coalesce with what's waiting.. */
2025                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
2026                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
2027                         RETURN(1);
2028                 }
2029         }
2030         if (lop->lop_num_pending >= cli->cl_max_pages_per_rpc)
2031                 RETURN(1);
2032
2033         RETURN(0);
2034 }
2035
2036 static int lop_makes_hprpc(struct loi_oap_pages *lop)
2037 {
2038         struct osc_async_page *oap;
2039         ENTRY;
2040
2041         if (cfs_list_empty(&lop->lop_urgent))
2042                 RETURN(0);
2043
2044         oap = cfs_list_entry(lop->lop_urgent.next,
2045                          struct osc_async_page, oap_urgent_item);
2046
2047         if (oap->oap_async_flags & ASYNC_HP) {
2048                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
2049                 RETURN(1);
2050         }
2051
2052         RETURN(0);
2053 }
2054
2055 static void on_list(cfs_list_t *item, cfs_list_t *list,
2056                     int should_be_on)
2057 {
2058         if (cfs_list_empty(item) && should_be_on)
2059                 cfs_list_add_tail(item, list);
2060         else if (!cfs_list_empty(item) && !should_be_on)
2061                 cfs_list_del_init(item);
2062 }
2063
2064 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
2065  * can find pages to build into rpcs quickly */
2066 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
2067 {
2068         if (lop_makes_hprpc(&loi->loi_write_lop) ||
2069             lop_makes_hprpc(&loi->loi_read_lop)) {
2070                 /* HP rpc */
2071                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2072                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2073         } else {
2074                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2075                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2076                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2077                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2078         }
2079
2080         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2081                 loi->loi_write_lop.lop_num_pending);
2082
2083         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2084                 loi->loi_read_lop.lop_num_pending);
2085 }
2086
2087 static void lop_update_pending(struct client_obd *cli,
2088                                struct loi_oap_pages *lop, int cmd, int delta)
2089 {
2090         lop->lop_num_pending += delta;
2091         if (cmd & OBD_BRW_WRITE)
2092                 cli->cl_pending_w_pages += delta;
2093         else
2094                 cli->cl_pending_r_pages += delta;
2095 }
2096
2097 /**
2098  * this is called when a sync waiter receives an interruption.  Its job is to
2099  * get the caller woken as soon as possible.  If its page hasn't been put in an
2100  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2101  * desiring interruption which will forcefully complete the rpc once the rpc
2102  * has timed out.
2103  */
2104 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2105 {
2106         struct loi_oap_pages *lop;
2107         struct lov_oinfo *loi;
2108         int rc = -EBUSY;
2109         ENTRY;
2110
2111         LASSERT(!oap->oap_interrupted);
2112         oap->oap_interrupted = 1;
2113
2114         /* ok, it's been put in an rpc. only one oap gets a request reference */
2115         if (oap->oap_request != NULL) {
2116                 ptlrpc_mark_interrupted(oap->oap_request);
2117                 ptlrpcd_wake(oap->oap_request);
2118                 ptlrpc_req_finished(oap->oap_request);
2119                 oap->oap_request = NULL;
2120         }
2121
2122         /*
2123          * page completion may be called only if ->cpo_prep() method was
2124          * executed by osc_io_submit(), that also adds page the to pending list
2125          */
2126         if (!cfs_list_empty(&oap->oap_pending_item)) {
2127                 cfs_list_del_init(&oap->oap_pending_item);
2128                 cfs_list_del_init(&oap->oap_urgent_item);
2129
2130                 loi = oap->oap_loi;
2131                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2132                         &loi->loi_write_lop : &loi->loi_read_lop;
2133                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2134                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2135                 rc = oap->oap_caller_ops->ap_completion(env,
2136                                           oap->oap_caller_data,
2137                                           oap->oap_cmd, NULL, -EINTR);
2138         }
2139
2140         RETURN(rc);
2141 }
2142
2143 /* this is trying to propogate async writeback errors back up to the
2144  * application.  As an async write fails we record the error code for later if
2145  * the app does an fsync.  As long as errors persist we force future rpcs to be
2146  * sync so that the app can get a sync error and break the cycle of queueing
2147  * pages for which writeback will fail. */
2148 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2149                            int rc)
2150 {
2151         if (rc) {
2152                 if (!ar->ar_rc)
2153                         ar->ar_rc = rc;
2154
2155                 ar->ar_force_sync = 1;
2156                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2157                 return;
2158
2159         }
2160
2161         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2162                 ar->ar_force_sync = 0;
2163 }
2164
2165 void osc_oap_to_pending(struct osc_async_page *oap)
2166 {
2167         struct loi_oap_pages *lop;
2168
2169         if (oap->oap_cmd & OBD_BRW_WRITE)
2170                 lop = &oap->oap_loi->loi_write_lop;
2171         else
2172                 lop = &oap->oap_loi->loi_read_lop;
2173
2174         if (oap->oap_async_flags & ASYNC_HP)
2175                 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2176         else if (oap->oap_async_flags & ASYNC_URGENT)
2177                 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2178         cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2179         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2180 }
2181
2182 /* this must be called holding the loi list lock to give coverage to exit_cache,
2183  * async_flag maintenance, and oap_request */
2184 static void osc_ap_completion(const struct lu_env *env,
2185                               struct client_obd *cli, struct obdo *oa,
2186                               struct osc_async_page *oap, int sent, int rc)
2187 {
2188         __u64 xid = 0;
2189
2190         ENTRY;
2191         if (oap->oap_request != NULL) {
2192                 xid = ptlrpc_req_xid(oap->oap_request);
2193                 ptlrpc_req_finished(oap->oap_request);
2194                 oap->oap_request = NULL;
2195         }
2196
2197         cfs_spin_lock(&oap->oap_lock);
2198         oap->oap_async_flags = 0;
2199         cfs_spin_unlock(&oap->oap_lock);
2200         oap->oap_interrupted = 0;
2201
2202         if (oap->oap_cmd & OBD_BRW_WRITE) {
2203                 osc_process_ar(&cli->cl_ar, xid, rc);
2204                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2205         }
2206
2207         if (rc == 0 && oa != NULL) {
2208                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2209                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2210                 if (oa->o_valid & OBD_MD_FLMTIME)
2211                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2212                 if (oa->o_valid & OBD_MD_FLATIME)
2213                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2214                 if (oa->o_valid & OBD_MD_FLCTIME)
2215                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2216         }
2217
2218         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2219                                                 oap->oap_cmd, oa, rc);
2220
2221         /* cl_page_completion() drops PG_locked. so, a new I/O on the page could
2222          * start, but OSC calls it under lock and thus we can add oap back to
2223          * pending safely */
2224         if (rc)
2225                 /* upper layer wants to leave the page on pending queue */
2226                 osc_oap_to_pending(oap);
2227         else
2228                 osc_exit_cache(cli, oap, sent);
2229         EXIT;
2230 }
2231
2232 static int brw_queue_work(const struct lu_env *env, void *data)
2233 {
2234         struct client_obd *cli = data;
2235
2236         CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
2237
2238         client_obd_list_lock(&cli->cl_loi_list_lock);
2239         osc_check_rpcs0(env, cli, 1);
2240         client_obd_list_unlock(&cli->cl_loi_list_lock);
2241         RETURN(0);
2242 }
2243
2244 static int brw_interpret(const struct lu_env *env,
2245                          struct ptlrpc_request *req, void *data, int rc)
2246 {
2247         struct osc_brw_async_args *aa = data;
2248         struct client_obd *cli;
2249         int async;
2250         ENTRY;
2251
2252         rc = osc_brw_fini_request(req, rc);
2253         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2254         /* When server return -EINPROGRESS, client should always retry
2255          * regardless of the number of times the bulk was resent already. */
2256         if (osc_recoverable_error(rc)) {
2257                 if (req->rq_import_generation !=
2258                     req->rq_import->imp_generation) {
2259                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
2260                                ""LPU64":"LPU64", rc = %d.\n",
2261                                req->rq_import->imp_obd->obd_name,
2262                                aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
2263                 } else if (rc == -EINPROGRESS ||
2264                     client_should_resend(aa->aa_resends, aa->aa_cli)) {
2265                         rc = osc_brw_redo_request(req, aa);
2266                 } else {
2267                         CERROR("%s: too many resent retries for object: "
2268                                ""LPU64":"LPU64", rc = %d.\n",
2269                                req->rq_import->imp_obd->obd_name,
2270                                aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
2271                 }
2272
2273                 if (rc == 0)
2274                         RETURN(0);
2275                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2276                         rc = -EIO;
2277         }
2278
2279         if (aa->aa_ocapa) {
2280                 capa_put(aa->aa_ocapa);
2281                 aa->aa_ocapa = NULL;
2282         }
2283
2284         cli = aa->aa_cli;
2285         client_obd_list_lock(&cli->cl_loi_list_lock);
2286
2287         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2288          * is called so we know whether to go to sync BRWs or wait for more
2289          * RPCs to complete */
2290         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2291                 cli->cl_w_in_flight--;
2292         else
2293                 cli->cl_r_in_flight--;
2294
2295         async = cfs_list_empty(&aa->aa_oaps);
2296         if (!async) { /* from osc_send_oap_rpc() */
2297                 struct osc_async_page *oap, *tmp;
2298                 /* the caller may re-use the oap after the completion call so
2299                  * we need to clean it up a little */
2300                 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2301                                              oap_rpc_item) {
2302                         cfs_list_del_init(&oap->oap_rpc_item);
2303                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2304                 }
2305                 OBDO_FREE(aa->aa_oa);
2306         } else { /* from async_internal() */
2307                 obd_count i;
2308                 for (i = 0; i < aa->aa_page_count; i++)
2309                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2310         }
2311         osc_wake_cache_waiters(cli);
2312         osc_check_rpcs0(env, cli, 1);
2313         client_obd_list_unlock(&cli->cl_loi_list_lock);
2314
2315         if (!async)
2316                 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2317                                   req->rq_bulk->bd_nob_transferred);
2318         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2319         ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
2320
2321         RETURN(rc);
2322 }
2323
2324 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2325                                             struct client_obd *cli,
2326                                             cfs_list_t *rpc_list,
2327                                             int page_count, int cmd)
2328 {
2329         struct ptlrpc_request *req;
2330         struct brw_page **pga = NULL;
2331         struct osc_brw_async_args *aa;
2332         struct obdo *oa = NULL;
2333         const struct obd_async_page_ops *ops = NULL;
2334         struct osc_async_page *oap;
2335         struct osc_async_page *tmp;
2336         struct cl_req *clerq = NULL;
2337         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2338         struct ldlm_lock *lock = NULL;
2339         struct cl_req_attr crattr;
2340         int i, rc, mpflag = 0;
2341
2342         ENTRY;
2343         LASSERT(!cfs_list_empty(rpc_list));
2344
2345         if (cmd & OBD_BRW_MEMALLOC)
2346                 mpflag = cfs_memory_pressure_get_and_set();
2347
2348         memset(&crattr, 0, sizeof crattr);
2349         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2350         if (pga == NULL)
2351                 GOTO(out, req = ERR_PTR(-ENOMEM));
2352
2353         OBDO_ALLOC(oa);
2354         if (oa == NULL)
2355                 GOTO(out, req = ERR_PTR(-ENOMEM));
2356
2357         i = 0;
2358         cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2359                 struct cl_page *page = osc_oap2cl_page(oap);
2360                 if (ops == NULL) {
2361                         ops = oap->oap_caller_ops;
2362
2363                         clerq = cl_req_alloc(env, page, crt,
2364                                              1 /* only 1-object rpcs for
2365                                                 * now */);
2366                         if (IS_ERR(clerq))
2367                                 GOTO(out, req = (void *)clerq);
2368                         lock = oap->oap_ldlm_lock;
2369                 }
2370                 pga[i] = &oap->oap_brw_page;
2371                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2372                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2373                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2374                 i++;
2375                 cl_req_page_add(env, clerq, page);
2376         }
2377
2378         /* always get the data for the obdo for the rpc */
2379         LASSERT(ops != NULL);
2380         crattr.cra_oa = oa;
2381         crattr.cra_capa = NULL;
2382         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2383         if (lock) {
2384                 oa->o_handle = lock->l_remote_handle;
2385                 oa->o_valid |= OBD_MD_FLHANDLE;
2386         }
2387
2388         rc = cl_req_prep(env, clerq);
2389         if (rc != 0) {
2390                 CERROR("cl_req_prep failed: %d\n", rc);
2391                 GOTO(out, req = ERR_PTR(rc));
2392         }
2393
2394         sort_brw_pages(pga, page_count);
2395         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2396                                   pga, &req, crattr.cra_capa, 1, 0);
2397         if (rc != 0) {
2398                 CERROR("prep_req failed: %d\n", rc);
2399                 GOTO(out, req = ERR_PTR(rc));
2400         }
2401
2402         if (cmd & OBD_BRW_MEMALLOC)
2403                 req->rq_memalloc = 1;
2404
2405         /* Need to update the timestamps after the request is built in case
2406          * we race with setattr (locally or in queue at OST).  If OST gets
2407          * later setattr before earlier BRW (as determined by the request xid),
2408          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2409          * way to do this in a single call.  bug 10150 */
2410         cl_req_attr_set(env, clerq, &crattr,
2411                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2412
2413         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2414         aa = ptlrpc_req_async_args(req);
2415         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2416         cfs_list_splice(rpc_list, &aa->aa_oaps);
2417         CFS_INIT_LIST_HEAD(rpc_list);
2418         aa->aa_clerq = clerq;
2419 out:
2420         if (cmd & OBD_BRW_MEMALLOC)
2421                 cfs_memory_pressure_restore(mpflag);
2422
2423         capa_put(crattr.cra_capa);
2424         if (IS_ERR(req)) {
2425                 if (oa)
2426                         OBDO_FREE(oa);
2427                 if (pga)
2428                         OBD_FREE(pga, sizeof(*pga) * page_count);
2429                 /* this should happen rarely and is pretty bad, it makes the
2430                  * pending list not follow the dirty order */
2431                 client_obd_list_lock(&cli->cl_loi_list_lock);
2432                 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2433                         cfs_list_del_init(&oap->oap_rpc_item);
2434
2435                         /* queued sync pages can be torn down while the pages
2436                          * were between the pending list and the rpc */
2437                         if (oap->oap_interrupted) {
2438                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2439                                 osc_ap_completion(env, cli, NULL, oap, 0,
2440                                                   oap->oap_count);
2441                                 continue;
2442                         }
2443                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2444                 }
2445                 if (clerq && !IS_ERR(clerq))
2446                         cl_req_completion(env, clerq, PTR_ERR(req));
2447         }
2448         RETURN(req);
2449 }
2450
2451 /**
2452  * prepare pages for ASYNC io and put pages in send queue.
2453  *
2454  * \param cmd OBD_BRW_* macroses
2455  * \param lop pending pages
2456  *
2457  * \return zero if no page added to send queue.
2458  * \return 1 if pages successfully added to send queue.
2459  * \return negative on errors.
2460  */
2461 static int
2462 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2463                  struct lov_oinfo *loi, int cmd,
2464                  struct loi_oap_pages *lop, pdl_policy_t pol)
2465 {
2466         struct ptlrpc_request *req;
2467         obd_count page_count = 0;
2468         struct osc_async_page *oap = NULL, *tmp;
2469         struct osc_brw_async_args *aa;
2470         const struct obd_async_page_ops *ops;
2471         CFS_LIST_HEAD(rpc_list);
2472         int srvlock = 0, mem_tight = 0;
2473         struct cl_object *clob = NULL;
2474         obd_off starting_offset = OBD_OBJECT_EOF;
2475         unsigned int ending_offset;
2476         int starting_page_off = 0;
2477         ENTRY;
2478
2479         /* ASYNC_HP pages first. At present, when the lock the pages is
2480          * to be canceled, the pages covered by the lock will be sent out
2481          * with ASYNC_HP. We have to send out them as soon as possible. */
2482         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2483                 if (oap->oap_async_flags & ASYNC_HP)
2484                         cfs_list_move(&oap->oap_pending_item, &rpc_list);
2485                 else if (!(oap->oap_brw_flags & OBD_BRW_SYNC))
2486                         /* only do this for writeback pages. */
2487                         cfs_list_move_tail(&oap->oap_pending_item, &rpc_list);
2488                 if (++page_count >= cli->cl_max_pages_per_rpc)
2489                         break;
2490         }
2491         cfs_list_splice_init(&rpc_list, &lop->lop_pending);
2492         page_count = 0;
2493
2494         /* first we find the pages we're allowed to work with */
2495         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2496                                      oap_pending_item) {
2497                 ops = oap->oap_caller_ops;
2498
2499                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2500                          "magic 0x%x\n", oap, oap->oap_magic);
2501
2502                 if (clob == NULL) {
2503                         /* pin object in memory, so that completion call-backs
2504                          * can be safely called under client_obd_list lock. */
2505                         clob = osc_oap2cl_page(oap)->cp_obj;
2506                         cl_object_get(clob);
2507                 }
2508
2509                 if (page_count != 0 &&
2510                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2511                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2512                                " oap %p, page %p, srvlock %u\n",
2513                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2514                         break;
2515                 }
2516
2517                 /* If there is a gap at the start of this page, it can't merge
2518                  * with any previous page, so we'll hand the network a
2519                  * "fragmented" page array that it can't transfer in 1 RDMA */
2520                 if (oap->oap_obj_off < starting_offset) {
2521                         if (starting_page_off != 0)
2522                                 break;
2523
2524                         starting_page_off = oap->oap_page_off;
2525                         starting_offset = oap->oap_obj_off + starting_page_off;
2526                 } else if (oap->oap_page_off != 0)
2527                         break;
2528
2529                 /* in llite being 'ready' equates to the page being locked
2530                  * until completion unlocks it.  commit_write submits a page
2531                  * as not ready because its unlock will happen unconditionally
2532                  * as the call returns.  if we race with commit_write giving
2533                  * us that page we don't want to create a hole in the page
2534                  * stream, so we stop and leave the rpc to be fired by
2535                  * another dirtier or kupdated interval (the not ready page
2536                  * will still be on the dirty list).  we could call in
2537                  * at the end of ll_file_write to process the queue again. */
2538                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2539                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2540                                                     cmd);
2541                         if (rc < 0)
2542                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2543                                                 "instead of ready\n", oap,
2544                                                 oap->oap_page, rc);
2545                         switch (rc) {
2546                         case -EAGAIN:
2547                                 /* llite is telling us that the page is still
2548                                  * in commit_write and that we should try
2549                                  * and put it in an rpc again later.  we
2550                                  * break out of the loop so we don't create
2551                                  * a hole in the sequence of pages in the rpc
2552                                  * stream.*/
2553                                 oap = NULL;
2554                                 break;
2555                         case -EINTR:
2556                                 /* the io isn't needed.. tell the checks
2557                                  * below to complete the rpc with EINTR */
2558                                 cfs_spin_lock(&oap->oap_lock);
2559                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2560                                 cfs_spin_unlock(&oap->oap_lock);
2561                                 oap->oap_count = -EINTR;
2562                                 break;
2563                         case 0:
2564                                 cfs_spin_lock(&oap->oap_lock);
2565                                 oap->oap_async_flags |= ASYNC_READY;
2566                                 cfs_spin_unlock(&oap->oap_lock);
2567                                 break;
2568                         default:
2569                                 LASSERTF(0, "oap %p page %p returned %d "
2570                                             "from make_ready\n", oap,
2571                                             oap->oap_page, rc);
2572                                 break;
2573                         }
2574                 }
2575                 if (oap == NULL)
2576                         break;
2577
2578                 /* take the page out of our book-keeping */
2579                 cfs_list_del_init(&oap->oap_pending_item);
2580                 lop_update_pending(cli, lop, cmd, -1);
2581                 cfs_list_del_init(&oap->oap_urgent_item);
2582
2583                 /* ask the caller for the size of the io as the rpc leaves. */
2584                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2585                         oap->oap_count =
2586                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2587                                                       cmd);
2588                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2589                 }
2590                 if (oap->oap_count <= 0) {
2591                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2592                                oap->oap_count);
2593                         osc_ap_completion(env, cli, NULL,
2594                                           oap, 0, oap->oap_count);
2595                         continue;
2596                 }
2597
2598                 /* now put the page back in our accounting */
2599                 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2600                 if (page_count++ == 0)
2601                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2602
2603                 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2604                         mem_tight = 1;
2605
2606                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2607                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2608                  * have the same alignment as the initial writes that allocated
2609                  * extents on the server. */
2610                 ending_offset = oap->oap_obj_off + oap->oap_page_off +
2611                                 oap->oap_count;
2612                 if (!(ending_offset & (PTLRPC_MAX_BRW_SIZE - 1)))
2613                         break;
2614
2615                 if (page_count >= cli->cl_max_pages_per_rpc)
2616                         break;
2617
2618                 /* If there is a gap at the end of this page, it can't merge
2619                  * with any subsequent pages, so we'll hand the network a
2620                  * "fragmented" page array that it can't transfer in 1 RDMA */
2621                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2622                         break;
2623         }
2624
2625         loi_list_maint(cli, loi);
2626
2627         client_obd_list_unlock(&cli->cl_loi_list_lock);
2628
2629         if (clob != NULL)
2630                 cl_object_put(env, clob);
2631
2632         if (page_count == 0) {
2633                 client_obd_list_lock(&cli->cl_loi_list_lock);
2634                 RETURN(0);
2635         }
2636
2637         req = osc_build_req(env, cli, &rpc_list, page_count,
2638                             mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2639         if (IS_ERR(req)) {
2640                 LASSERT(cfs_list_empty(&rpc_list));
2641                 loi_list_maint(cli, loi);
2642                 RETURN(PTR_ERR(req));
2643         }
2644
2645         aa = ptlrpc_req_async_args(req);
2646
2647         starting_offset &= PTLRPC_MAX_BRW_SIZE - 1;
2648         if (cmd == OBD_BRW_READ) {
2649                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2650                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2651                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2652                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2653         } else {
2654                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2655                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2656                                  cli->cl_w_in_flight);
2657                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2658                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2659         }
2660
2661         client_obd_list_lock(&cli->cl_loi_list_lock);
2662
2663         if (cmd == OBD_BRW_READ)
2664                 cli->cl_r_in_flight++;
2665         else
2666                 cli->cl_w_in_flight++;
2667
2668         /* queued sync pages can be torn down while the pages
2669          * were between the pending list and the rpc */
2670         tmp = NULL;
2671         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2672                 /* only one oap gets a request reference */
2673                 if (tmp == NULL)
2674                         tmp = oap;
2675                 if (oap->oap_interrupted && !req->rq_intr) {
2676                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2677                                oap, req);
2678                         ptlrpc_mark_interrupted(req);
2679                 }
2680         }
2681         if (tmp != NULL)
2682                 tmp->oap_request = ptlrpc_request_addref(req);
2683
2684         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2685                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2686
2687         req->rq_interpret_reply = brw_interpret;
2688
2689         /* XXX: Maybe the caller can check the RPC bulk descriptor to see which
2690          *      CPU/NUMA node the majority of pages were allocated on, and try
2691          *      to assign the async RPC to the CPU core (PDL_POLICY_PREFERRED)
2692          *      to reduce cross-CPU memory traffic.
2693          *
2694          *      But on the other hand, we expect that multiple ptlrpcd threads
2695          *      and the initial write sponsor can run in parallel, especially
2696          *      when data checksum is enabled, which is CPU-bound operation and
2697          *      single ptlrpcd thread cannot process in time. So more ptlrpcd
2698          *      threads sharing BRW load (with PDL_POLICY_ROUND) seems better.
2699          */
2700         ptlrpcd_add_req(req, pol, -1);
2701         RETURN(1);
2702 }
2703
2704 #define LOI_DEBUG(LOI, STR, args...)                                     \
2705         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2706                !cfs_list_empty(&(LOI)->loi_ready_item) ||                \
2707                !cfs_list_empty(&(LOI)->loi_hp_ready_item),               \
2708                (LOI)->loi_write_lop.lop_num_pending,                     \
2709                !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent),        \
2710                (LOI)->loi_read_lop.lop_num_pending,                      \
2711                !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent),         \
2712                args)                                                     \
2713
2714 /* This is called by osc_check_rpcs() to find which objects have pages that
2715  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2716 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2717 {
2718         ENTRY;
2719
2720         /* First return objects that have blocked locks so that they
2721          * will be flushed quickly and other clients can get the lock,
2722          * then objects which have pages ready to be stuffed into RPCs */
2723         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2724                 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2725                                       struct lov_oinfo, loi_hp_ready_item));
2726         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2727                 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2728                                       struct lov_oinfo, loi_ready_item));
2729
2730         /* then if we have cache waiters, return all objects with queued
2731          * writes.  This is especially important when many small files
2732          * have filled up the cache and not been fired into rpcs because
2733          * they don't pass the nr_pending/object threshhold */
2734         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2735             !cfs_list_empty(&cli->cl_loi_write_list))
2736                 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2737                                       struct lov_oinfo, loi_write_item));
2738
2739         /* then return all queued objects when we have an invalid import
2740          * so that they get flushed */
2741         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2742                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2743                         RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2744                                               struct lov_oinfo,
2745                                               loi_write_item));
2746                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2747                         RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2748                                               struct lov_oinfo, loi_read_item));
2749         }
2750         RETURN(NULL);
2751 }
2752
2753 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2754 {
2755         struct osc_async_page *oap;
2756         int hprpc = 0;
2757
2758         if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2759                 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2760                                      struct osc_async_page, oap_urgent_item);
2761                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2762         }
2763
2764         if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2765                 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2766                                      struct osc_async_page, oap_urgent_item);
2767                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2768         }
2769
2770         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2771 }
2772
2773 /* called with the loi list lock held */
2774 static void osc_check_rpcs0(const struct lu_env *env, struct client_obd *cli, int ptlrpc)
2775 {
2776         struct lov_oinfo *loi;
2777         int rc = 0, race_counter = 0;
2778         pdl_policy_t pol;
2779         ENTRY;
2780
2781         pol = ptlrpc ? PDL_POLICY_SAME : PDL_POLICY_ROUND;
2782
2783         while ((loi = osc_next_loi(cli)) != NULL) {
2784                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2785
2786                 if (osc_max_rpc_in_flight(cli, loi))
2787                         break;
2788
2789                 /* attempt some read/write balancing by alternating between
2790                  * reads and writes in an object.  The makes_rpc checks here
2791                  * would be redundant if we were getting read/write work items
2792                  * instead of objects.  we don't want send_oap_rpc to drain a
2793                  * partial read pending queue when we're given this object to
2794                  * do io on writes while there are cache waiters */
2795                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2796                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2797                                               &loi->loi_write_lop, pol);
2798                         if (rc < 0) {
2799                                 CERROR("Write request failed with %d\n", rc);
2800
2801                                 /* osc_send_oap_rpc failed, mostly because of
2802                                  * memory pressure.
2803                                  *
2804                                  * It can't break here, because if:
2805                                  *  - a page was submitted by osc_io_submit, so
2806                                  *    page locked;
2807                                  *  - no request in flight
2808                                  *  - no subsequent request
2809                                  * The system will be in live-lock state,
2810                                  * because there is no chance to call
2811                                  * osc_io_unplug() and osc_check_rpcs() any
2812                                  * more. pdflush can't help in this case,
2813                                  * because it might be blocked at grabbing
2814                                  * the page lock as we mentioned.
2815                                  *
2816                                  * Anyway, continue to drain pages. */
2817                                 /* break; */
2818                         }
2819
2820                         if (rc > 0)
2821                                 race_counter = 0;
2822                         else if (rc == 0)
2823                                 race_counter++;
2824                 }
2825                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2826                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2827                                               &loi->loi_read_lop, pol);
2828                         if (rc < 0)
2829                                 CERROR("Read request failed with %d\n", rc);
2830
2831                         if (rc > 0)
2832                                 race_counter = 0;
2833                         else if (rc == 0)
2834                                 race_counter++;
2835                 }
2836
2837                 /* attempt some inter-object balancing by issuing rpcs
2838                  * for each object in turn */
2839                 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2840                         cfs_list_del_init(&loi->loi_hp_ready_item);
2841                 if (!cfs_list_empty(&loi->loi_ready_item))
2842                         cfs_list_del_init(&loi->loi_ready_item);
2843                 if (!cfs_list_empty(&loi->loi_write_item))
2844                         cfs_list_del_init(&loi->loi_write_item);
2845                 if (!cfs_list_empty(&loi->loi_read_item))
2846                         cfs_list_del_init(&loi->loi_read_item);
2847
2848                 loi_list_maint(cli, loi);
2849
2850                 /* send_oap_rpc fails with 0 when make_ready tells it to
2851                  * back off.  llite's make_ready does this when it tries
2852                  * to lock a page queued for write that is already locked.
2853                  * we want to try sending rpcs from many objects, but we
2854                  * don't want to spin failing with 0.  */
2855                 if (race_counter == 10)
2856                         break;
2857         }
2858 }
2859
2860 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2861 {
2862         osc_check_rpcs0(env, cli, 0);
2863 }
2864
2865 /**
2866  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2867  * is available.
2868  */
2869 int osc_enter_cache_try(const struct lu_env *env,
2870                         struct client_obd *cli, struct lov_oinfo *loi,
2871                         struct osc_async_page *oap, int transient)
2872 {
2873         int has_grant;
2874
2875         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2876         if (has_grant) {
2877                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2878                 if (transient) {
2879                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2880                         cfs_atomic_inc(&obd_dirty_transit_pages);
2881                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2882                 }
2883         }
2884         return has_grant;
2885 }
2886
2887 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2888  * grant or cache space. */
2889 static int osc_enter_cache(const struct lu_env *env,
2890                            struct client_obd *cli, struct lov_oinfo *loi,
2891                            struct osc_async_page *oap)
2892 {
2893         struct osc_cache_waiter ocw;
2894         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
2895         int rc = -EDQUOT;
2896         ENTRY;
2897
2898         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2899                "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2900                cli->cl_dirty_max, obd_max_dirty_pages,
2901                cli->cl_lost_grant, cli->cl_avail_grant);
2902
2903         /* force the caller to try sync io.  this can jump the list
2904          * of queued writes and create a discontiguous rpc stream */
2905         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
2906             cli->cl_dirty_max < CFS_PAGE_SIZE     ||
2907             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
2908                 RETURN(-EDQUOT);
2909
2910         /* Hopefully normal case - cache space and write credits available */
2911         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2912             cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2913             osc_enter_cache_try(env, cli, loi, oap, 0))
2914                 RETURN(0);
2915
2916         /* We can get here for two reasons: too many dirty pages in cache, or
2917          * run out of grants. In both cases we should write dirty pages out.
2918          * Adding a cache waiter will trigger urgent write-out no matter what
2919          * RPC size will be.
2920          * The exiting condition is no avail grants and no dirty pages caching,
2921          * that really means there is no space on the OST. */
2922         cfs_waitq_init(&ocw.ocw_waitq);
2923         ocw.ocw_oap = oap;
2924         while (cli->cl_dirty > 0) {
2925                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2926                 ocw.ocw_rc = 0;
2927
2928                 loi_list_maint(cli, loi);
2929                 osc_check_rpcs(env, cli);
2930                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2931
2932                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
2933                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
2934
2935                 rc = l_wait_event(ocw.ocw_waitq, cfs_list_empty(&ocw.ocw_entry), &lwi);
2936
2937                 client_obd_list_lock(&cli->cl_loi_list_lock);
2938                 cfs_list_del_init(&ocw.ocw_entry);
2939                 if (rc < 0)
2940                         break;
2941
2942                 rc = ocw.ocw_rc;
2943                 if (rc != -EDQUOT)
2944                         break;
2945         }
2946
2947         RETURN(rc);
2948 }
2949
2950
2951 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2952                         struct lov_oinfo *loi, cfs_page_t *page,
2953                         obd_off offset, const struct obd_async_page_ops *ops,
2954                         void *data, void **res, int nocache,
2955                         struct lustre_handle *lockh)
2956 {
2957         struct osc_async_page *oap;
2958
2959         ENTRY;
2960
2961         if (!page)
2962                 return cfs_size_round(sizeof(*oap));
2963
2964         oap = *res;
2965         oap->oap_magic = OAP_MAGIC;
2966         oap->oap_cli = &exp->exp_obd->u.cli;
2967         oap->oap_loi = loi;
2968
2969         oap->oap_caller_ops = ops;
2970         oap->oap_caller_data = data;
2971
2972         oap->oap_page = page;
2973         oap->oap_obj_off = offset;
2974         if (!client_is_remote(exp) &&
2975             cfs_capable(CFS_CAP_SYS_RESOURCE))
2976                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2977
2978         LASSERT(!(offset & ~CFS_PAGE_MASK));
2979
2980         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2981         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2982         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2983         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2984
2985         cfs_spin_lock_init(&oap->oap_lock);
2986         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2987         RETURN(0);
2988 }
2989
2990 int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
2991                        struct lov_stripe_md *lsm, struct lov_oinfo *loi,
2992                        struct osc_async_page *oap, int cmd, int off,
2993                        int count, obd_flag brw_flags, enum async_flags async_flags)
2994 {
2995         struct client_obd *cli = &exp->exp_obd->u.cli;
2996         int rc = 0;
2997         ENTRY;
2998
2999         if (oap->oap_magic != OAP_MAGIC)
3000                 RETURN(-EINVAL);
3001
3002         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
3003                 RETURN(-EIO);
3004
3005         if (!cfs_list_empty(&oap->oap_pending_item) ||
3006             !cfs_list_empty(&oap->oap_urgent_item) ||
3007             !cfs_list_empty(&oap->oap_rpc_item))
3008                 RETURN(-EBUSY);
3009
3010         /* check if the file's owner/group is over quota */
3011         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
3012                 struct cl_object *obj;
3013                 struct cl_attr    attr; /* XXX put attr into thread info */
3014                 unsigned int qid[MAXQUOTAS];
3015
3016                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
3017
3018                 cl_object_attr_lock(obj);
3019                 rc = cl_object_attr_get(env, obj, &attr);
3020                 cl_object_attr_unlock(obj);
3021
3022                 qid[USRQUOTA] = attr.cat_uid;
3023                 qid[GRPQUOTA] = attr.cat_gid;
3024                 if (rc == 0 &&
3025                     osc_quota_chkdq(cli, qid) == NO_QUOTA)
3026                         rc = -EDQUOT;
3027                 if (rc)
3028                         RETURN(rc);
3029         }
3030
3031         if (loi == NULL)
3032                 loi = lsm->lsm_oinfo[0];
3033
3034         client_obd_list_lock(&cli->cl_loi_list_lock);
3035
3036         LASSERT(off + count <= CFS_PAGE_SIZE);
3037         oap->oap_cmd = cmd;
3038         oap->oap_page_off = off;
3039         oap->oap_count = count;
3040         oap->oap_brw_flags = brw_flags;
3041         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
3042         if (cfs_memory_pressure_get())
3043                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
3044         cfs_spin_lock(&oap->oap_lock);
3045         oap->oap_async_flags = async_flags;
3046         cfs_spin_unlock(&oap->oap_lock);
3047
3048         if (cmd & OBD_BRW_WRITE) {
3049                 rc = osc_enter_cache(env, cli, loi, oap);
3050                 if (rc) {
3051                         client_obd_list_unlock(&cli->cl_loi_list_lock);
3052                         RETURN(rc);
3053                 }
3054         }
3055
3056         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3057                   cmd);
3058
3059         osc_oap_to_pending(oap);
3060         loi_list_maint(cli, loi);
3061         if (!osc_max_rpc_in_flight(cli, loi) &&
3062             lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
3063                 LASSERT(cli->cl_writeback_work != NULL);
3064                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
3065
3066                 CDEBUG(D_CACHE, "Queued writeback work for client obd %p/%d.\n",
3067                        cli, rc);
3068         }
3069         client_obd_list_unlock(&cli->cl_loi_list_lock);
3070
3071         RETURN(0);
3072 }
3073
3074 /* aka (~was & now & flag), but this is more clear :) */
3075 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3076
3077 int osc_set_async_flags_base(struct client_obd *cli,
3078                              struct lov_oinfo *loi, struct osc_async_page *oap,
3079                              obd_flag async_flags)
3080 {
3081         struct loi_oap_pages *lop;
3082         int flags = 0;
3083         ENTRY;
3084
3085         LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3086
3087         if (oap->oap_cmd & OBD_BRW_WRITE) {
3088                 lop = &loi->loi_write_lop;
3089         } else {
3090                 lop = &loi->loi_read_lop;
3091         }
3092
3093         if ((oap->oap_async_flags & async_flags) == async_flags)
3094                 RETURN(0);
3095
3096         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3097                 flags |= ASYNC_READY;
3098
3099         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3100             cfs_list_empty(&oap->oap_rpc_item)) {
3101                 if (oap->oap_async_flags & ASYNC_HP)
3102                         cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3103                 else
3104                         cfs_list_add_tail(&oap->oap_urgent_item,
3105                                           &lop->lop_urgent);
3106                 flags |= ASYNC_URGENT;
3107                 loi_list_maint(cli, loi);
3108         }
3109         cfs_spin_lock(&oap->oap_lock);
3110         oap->oap_async_flags |= flags;
3111         cfs_spin_unlock(&oap->oap_lock);
3112
3113         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3114                         oap->oap_async_flags);
3115         RETURN(0);
3116 }
3117
3118 int osc_teardown_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
3119                             struct lov_oinfo *loi, struct osc_async_page *oap)
3120 {
3121         struct client_obd *cli = &exp->exp_obd->u.cli;
3122         struct loi_oap_pages *lop;
3123         int rc = 0;
3124         ENTRY;
3125
3126         if (oap->oap_magic != OAP_MAGIC)
3127                 RETURN(-EINVAL);
3128
3129         if (loi == NULL)
3130                 loi = lsm->lsm_oinfo[0];
3131
3132         if (oap->oap_cmd & OBD_BRW_WRITE) {
3133                 lop = &loi->loi_write_lop;
3134         } else {
3135                 lop = &loi->loi_read_lop;
3136         }
3137
3138         client_obd_list_lock(&cli->cl_loi_list_lock);
3139
3140         if (!cfs_list_empty(&oap->oap_rpc_item))
3141                 GOTO(out, rc = -EBUSY);
3142
3143         osc_exit_cache(cli, oap, 0);
3144         osc_wake_cache_waiters(cli);
3145
3146         if (!cfs_list_empty(&oap->oap_urgent_item)) {
3147                 cfs_list_del_init(&oap->oap_urgent_item);
3148                 cfs_spin_lock(&oap->oap_lock);
3149                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3150                 cfs_spin_unlock(&oap->oap_lock);
3151         }
3152         if (!cfs_list_empty(&oap->oap_pending_item)) {
3153                 cfs_list_del_init(&oap->oap_pending_item);
3154                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3155         }
3156         loi_list_maint(cli, loi);
3157         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3158 out:
3159         client_obd_list_unlock(&cli->cl_loi_list_lock);
3160         RETURN(rc);
3161 }
3162
3163 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
3164                                         struct ldlm_enqueue_info *einfo)
3165 {
3166         void *data = einfo->ei_cbdata;
3167         int set = 0;
3168
3169         LASSERT(lock != NULL);
3170         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3171         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3172         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3173         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3174
3175         lock_res_and_lock(lock);
3176         cfs_spin_lock(&osc_ast_guard);
3177
3178         if (lock->l_ast_data == NULL)
3179                 lock->l_ast_data = data;
3180         if (lock->l_ast_data == data)
3181                 set = 1;
3182
3183         cfs_spin_unlock(&osc_ast_guard);
3184         unlock_res_and_lock(lock);
3185
3186         return set;
3187 }
3188
3189 static int osc_set_data_with_check(struct lustre_handle *lockh,
3190                                    struct ldlm_enqueue_info *einfo)
3191 {
3192         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3193         int set = 0;
3194
3195         if (lock != NULL) {
3196                 set = osc_set_lock_data_with_check(lock, einfo);
3197                 LDLM_LOCK_PUT(lock);
3198         } else
3199                 CERROR("lockh %p, data %p - client evicted?\n",
3200                        lockh, einfo->ei_cbdata);
3201         return set;
3202 }
3203
3204 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3205                              ldlm_iterator_t replace, void *data)
3206 {
3207         struct ldlm_res_id res_id;
3208         struct obd_device *obd = class_exp2obd(exp);
3209
3210         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3211         ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3212         return 0;
3213 }
3214
3215 /* find any ldlm lock of the inode in osc
3216  * return 0    not find
3217  *        1    find one
3218  *      < 0    error */
3219 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3220                            ldlm_iterator_t replace, void *data)
3221 {
3222         struct ldlm_res_id res_id;
3223         struct obd_device *obd = class_exp2obd(exp);
3224         int rc = 0;
3225
3226         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3227         rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3228         if (rc == LDLM_ITER_STOP)
3229                 return(1);
3230         if (rc == LDLM_ITER_CONTINUE)
3231                 return(0);
3232         return(rc);
3233 }
3234
3235 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3236                             obd_enqueue_update_f upcall, void *cookie,
3237                             int *flags, int agl, int rc)
3238 {
3239         int intent = *flags & LDLM_FL_HAS_INTENT;
3240         ENTRY;
3241
3242         if (intent) {
3243                 /* The request was created before ldlm_cli_enqueue call. */
3244                 if (rc == ELDLM_LOCK_ABORTED) {
3245                         struct ldlm_reply *rep;
3246                         rep = req_capsule_server_get(&req->rq_pill,
3247                                                      &RMF_DLM_REP);
3248
3249                         LASSERT(rep != NULL);
3250                         if (rep->lock_policy_res1)
3251                                 rc = rep->lock_policy_res1;
3252                 }
3253         }
3254
3255         if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
3256             (rc == 0)) {
3257                 *flags |= LDLM_FL_LVB_READY;
3258                 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3259                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3260         }
3261
3262         /* Call the update callback. */
3263         rc = (*upcall)(cookie, rc);
3264         RETURN(rc);
3265 }
3266
3267 static int osc_enqueue_interpret(const struct lu_env *env,
3268                                  struct ptlrpc_request *req,
3269                                  struct osc_enqueue_args *aa, int rc)
3270 {
3271         struct ldlm_lock *lock;
3272         struct lustre_handle handle;
3273         __u32 mode;
3274         struct ost_lvb *lvb;
3275         __u32 lvb_len;
3276         int *flags = aa->oa_flags;
3277
3278         /* Make a local copy of a lock handle and a mode, because aa->oa_*
3279          * might be freed anytime after lock upcall has been called. */
3280         lustre_handle_copy(&handle, aa->oa_lockh);
3281         mode = aa->oa_ei->ei_mode;
3282
3283         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3284          * be valid. */
3285         lock = ldlm_handle2lock(&handle);
3286
3287         /* Take an additional reference so that a blocking AST that
3288          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3289          * to arrive after an upcall has been executed by
3290          * osc_enqueue_fini(). */
3291         ldlm_lock_addref(&handle, mode);
3292
3293         /* Let CP AST to grant the lock first. */
3294         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
3295
3296         if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
3297                 lvb = NULL;
3298                 lvb_len = 0;
3299         } else {
3300                 lvb = aa->oa_lvb;
3301                 lvb_len = sizeof(*aa->oa_lvb);
3302         }
3303
3304         /* Complete obtaining the lock procedure. */
3305         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3306                                    mode, flags, lvb, lvb_len, &handle, rc);
3307         /* Complete osc stuff. */
3308         rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
3309                               flags, aa->oa_agl, rc);
3310
3311         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3312
3313         /* Release the lock for async request. */
3314         if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3315                 /*
3316                  * Releases a reference taken by ldlm_cli_enqueue(), if it is
3317                  * not already released by
3318                  * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3319                  */
3320                 ldlm_lock_decref(&handle, mode);
3321
3322         LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3323                  aa->oa_lockh, req, aa);
3324         ldlm_lock_decref(&handle, mode);
3325         LDLM_LOCK_PUT(lock);
3326         return rc;
3327 }
3328
3329 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3330                         struct lov_oinfo *loi, int flags,
3331                         struct ost_lvb *lvb, __u32 mode, int rc)
3332 {
3333         struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3334
3335         if (rc == ELDLM_OK) {
3336                 __u64 tmp;
3337
3338                 LASSERT(lock != NULL);
3339                 loi->loi_lvb = *lvb;
3340                 tmp = loi->loi_lvb.lvb_size;
3341                 /* Extend KMS up to the end of this lock and no further
3342                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3343                 if (tmp > lock->l_policy_data.l_extent.end)
3344                         tmp = lock->l_policy_data.l_extent.end + 1;
3345                 if (tmp >= loi->loi_kms) {
3346                         LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3347                                    ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3348                         loi_kms_set(loi, tmp);
3349                 } else {
3350                         LDLM_DEBUG(lock, "lock acquired, setting rss="
3351                                    LPU64"; leaving kms="LPU64", end="LPU64,
3352                                    loi->loi_lvb.lvb_size, loi->loi_kms,
3353                                    lock->l_policy_data.l_extent.end);
3354                 }
3355                 ldlm_lock_allow_match(lock);
3356         } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3357                 LASSERT(lock != NULL);
3358                 loi->loi_lvb = *lvb;
3359                 ldlm_lock_allow_match(lock);
3360                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3361                        " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3362                 rc = ELDLM_OK;
3363         }
3364
3365         if (lock != NULL) {
3366                 if (rc != ELDLM_OK)
3367                         ldlm_lock_fail_match(lock);
3368
3369                 LDLM_LOCK_PUT(lock);
3370         }
3371 }
3372 EXPORT_SYMBOL(osc_update_enqueue);
3373
3374 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3375
3376 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3377  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3378  * other synchronous requests, however keeping some locks and trying to obtain
3379  * others may take a considerable amount of time in a case of ost failure; and
3380  * when other sync requests do not get released lock from a client, the client
3381  * is excluded from the cluster -- such scenarious make the life difficult, so
3382  * release locks just after they are obtained. */
3383 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3384                      int *flags, ldlm_policy_data_t *policy,
3385                      struct ost_lvb *lvb, int kms_valid,
3386                      obd_enqueue_update_f upcall, void *cookie,
3387                      struct ldlm_enqueue_info *einfo,
3388                      struct lustre_handle *lockh,
3389                      struct ptlrpc_request_set *rqset, int async, int agl)
3390 {
3391         struct obd_device *obd = exp->exp_obd;
3392         struct ptlrpc_request *req = NULL;
3393         int intent = *flags & LDLM_FL_HAS_INTENT;
3394         int match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
3395         ldlm_mode_t mode;
3396         int rc;
3397         ENTRY;
3398
3399         /* Filesystem lock extents are extended to page boundaries so that
3400          * dealing with the page cache is a little smoother.  */
3401         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3402         policy->l_extent.end |= ~CFS_PAGE_MASK;
3403
3404         /*
3405          * kms is not valid when either object is completely fresh (so that no
3406          * locks are cached), or object was evicted. In the latter case cached
3407          * lock cannot be used, because it would prime inode state with
3408          * potentially stale LVB.
3409          */
3410         if (!kms_valid)
3411                 goto no_match;
3412
3413         /* Next, search for already existing extent locks that will cover us */
3414         /* If we're trying to read, we also search for an existing PW lock.  The
3415          * VFS and page cache already protect us locally, so lots of readers/
3416          * writers can share a single PW lock.
3417          *
3418          * There are problems with conversion deadlocks, so instead of
3419          * converting a read lock to a write lock, we'll just enqueue a new
3420          * one.
3421          *
3422          * At some point we should cancel the read lock instead of making them
3423          * send us a blocking callback, but there are problems with canceling
3424          * locks out from other users right now, too. */
3425         mode = einfo->ei_mode;
3426         if (einfo->ei_mode == LCK_PR)
3427                 mode |= LCK_PW;
3428         mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
3429                                einfo->ei_type, policy, mode, lockh, 0);
3430         if (mode) {
3431                 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3432
3433                 if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
3434                         /* For AGL, if enqueue RPC is sent but the lock is not
3435                          * granted, then skip to process this strpe.
3436                          * Return -ECANCELED to tell the caller. */
3437                         ldlm_lock_decref(lockh, mode);
3438                         LDLM_LOCK_PUT(matched);
3439                         RETURN(-ECANCELED);
3440                 } else if (osc_set_lock_data_with_check(matched, einfo)) {
3441                         *flags |= LDLM_FL_LVB_READY;
3442                         /* addref the lock only if not async requests and PW
3443                          * lock is matched whereas we asked for PR. */
3444                         if (!rqset && einfo->ei_mode != mode)
3445                                 ldlm_lock_addref(lockh, LCK_PR);
3446                         if (intent) {
3447                                 /* I would like to be able to ASSERT here that
3448                                  * rss <= kms, but I can't, for reasons which
3449                                  * are explained in lov_enqueue() */
3450                         }
3451
3452                         /* We already have a lock, and it's referenced */
3453                         (*upcall)(cookie, ELDLM_OK);
3454
3455                         if (einfo->ei_mode != mode)
3456                                 ldlm_lock_decref(lockh, LCK_PW);
3457                         else if (rqset)
3458                                 /* For async requests, decref the lock. */
3459                                 ldlm_lock_decref(lockh, einfo->ei_mode);
3460                         LDLM_LOCK_PUT(matched);
3461                         RETURN(ELDLM_OK);
3462                 } else {
3463                         ldlm_lock_decref(lockh, mode);
3464                         LDLM_LOCK_PUT(matched);
3465                 }
3466         }
3467
3468  no_match:
3469         if (intent) {
3470                 CFS_LIST_HEAD(cancels);
3471                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3472                                            &RQF_LDLM_ENQUEUE_LVB);
3473                 if (req == NULL)
3474                         RETURN(-ENOMEM);
3475
3476                 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3477                 if (rc) {
3478                         ptlrpc_request_free(req);
3479                         RETURN(rc);
3480                 }
3481
3482                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3483                                      sizeof *lvb);
3484                 ptlrpc_request_set_replen(req);
3485         }
3486
3487         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3488         *flags &= ~LDLM_FL_BLOCK_GRANTED;
3489
3490         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3491                               sizeof(*lvb), lockh, async);
3492         if (rqset) {
3493                 if (!rc) {
3494                         struct osc_enqueue_args *aa;
3495                         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3496                         aa = ptlrpc_req_async_args(req);
3497                         aa->oa_ei = einfo;
3498                         aa->oa_exp = exp;
3499                         aa->oa_flags  = flags;
3500                         aa->oa_upcall = upcall;
3501                         aa->oa_cookie = cookie;
3502                         aa->oa_lvb    = lvb;
3503                         aa->oa_lockh  = lockh;
3504                         aa->oa_agl    = !!agl;
3505
3506                         req->rq_interpret_reply =
3507                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3508                         if (rqset == PTLRPCD_SET)
3509                                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3510                         else
3511                                 ptlrpc_set_add_req(rqset, req);
3512                 } else if (intent) {
3513                         ptlrpc_req_finished(req);
3514                 }
3515                 RETURN(rc);
3516         }
3517
3518         rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
3519         if (intent)
3520                 ptlrpc_req_finished(req);
3521
3522         RETURN(rc);
3523 }
3524
3525 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3526                        struct ldlm_enqueue_info *einfo,
3527                        struct ptlrpc_request_set *rqset)
3528 {
3529         struct ldlm_res_id res_id;
3530         int rc;
3531         ENTRY;
3532
3533         osc_build_res_name(oinfo->oi_md->lsm_object_id,
3534                            oinfo->oi_md->lsm_object_seq, &res_id);
3535
3536         rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3537                               &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3538                               oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3539                               oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3540                               rqset, rqset != NULL, 0);
3541         RETURN(rc);
3542 }
3543
3544 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3545                    __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3546                    int *flags, void *data, struct lustre_handle *lockh,
3547                    int unref)
3548 {
3549         struct obd_device *obd = exp->exp_obd;
3550         int lflags = *flags;
3551         ldlm_mode_t rc;
3552         ENTRY;
3553
3554         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3555                 RETURN(-EIO);
3556
3557         /* Filesystem lock extents are extended to page boundaries so that
3558          * dealing with the page cache is a little smoother */
3559         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3560         policy->l_extent.end |= ~CFS_PAGE_MASK;
3561
3562         /* Next, search for already existing extent locks that will cover us */
3563         /* If we're trying to read, we also search for an existing PW lock.  The
3564          * VFS and page cache already protect us locally, so lots of readers/
3565          * writers can share a single PW lock. */
3566         rc = mode;
3567         if (mode == LCK_PR)
3568                 rc |= LCK_PW;
3569         rc = ldlm_lock_match(obd->obd_namespace, lflags,
3570                              res_id, type, policy, rc, lockh, unref);
3571         if (rc) {
3572                 if (data != NULL) {
3573                         if (!osc_set_data_with_check(lockh, data)) {
3574                                 if (!(lflags & LDLM_FL_TEST_LOCK))
3575                                         ldlm_lock_decref(lockh, rc);
3576                                 RETURN(0);
3577                         }
3578                 }
3579                 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3580                         ldlm_lock_addref(lockh, LCK_PR);
3581                         ldlm_lock_decref(lockh, LCK_PW);
3582                 }
3583                 RETURN(rc);
3584         }
3585         RETURN(rc);
3586 }
3587
3588 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3589 {
3590         ENTRY;
3591
3592         if (unlikely(mode == LCK_GROUP))
3593                 ldlm_lock_decref_and_cancel(lockh, mode);
3594         else
3595                 ldlm_lock_decref(lockh, mode);
3596
3597         RETURN(0);
3598 }
3599
3600 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3601                       __u32 mode, struct lustre_handle *lockh)
3602 {
3603         ENTRY;
3604         RETURN(osc_cancel_base(lockh, mode));
3605 }
3606
3607 static int osc_cancel_unused(struct obd_export *exp,
3608                              struct lov_stripe_md *lsm,
3609                              ldlm_cancel_flags_t flags,
3610                              void *opaque)
3611 {
3612         struct obd_device *obd = class_exp2obd(exp);
3613         struct ldlm_res_id res_id, *resp = NULL;
3614
3615         if (lsm != NULL) {
3616                 resp = osc_build_res_name(lsm->lsm_object_id,
3617                                           lsm->lsm_object_seq, &res_id);
3618         }
3619
3620         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3621 }
3622
3623 static int osc_statfs_interpret(const struct lu_env *env,
3624                                 struct ptlrpc_request *req,
3625                                 struct osc_async_args *aa, int rc)
3626 {
3627         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3628         struct obd_statfs *msfs;
3629         __u64 used;
3630         ENTRY;
3631
3632         if (rc == -EBADR)
3633                 /* The request has in fact never been sent