Whamcloud - gitweb
b=20997 Enable time limit on the request to be queued
[fs/lustre-release.git] / lustre / osc / osc_request.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 #ifndef EXPORT_SYMTAB
38 # define EXPORT_SYMTAB
39 #endif
40 #define DEBUG_SUBSYSTEM S_OSC
41
42 #include <libcfs/libcfs.h>
43
44 #ifndef __KERNEL__
45 # include <liblustre.h>
46 #endif
47
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre/lustre_user.h>
51 #include <obd_cksum.h>
52 #include <obd_ost.h>
53 #include <obd_lov.h>
54
55 #ifdef  __CYGWIN__
56 # include <ctype.h>
57 #endif
58
59 #include <lustre_ha.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_debug.h>
63 #include <lustre_param.h>
64 #include "osc_internal.h"
65
66 static quota_interface_t *quota_interface = NULL;
67 extern quota_interface_t osc_quota_interface;
68
69 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
70 static int brw_interpret(const struct lu_env *env,
71                          struct ptlrpc_request *req, void *data, int rc);
72 int osc_cleanup(struct obd_device *obd);
73
74 /* Pack OSC object metadata for disk storage (LE byte order). */
75 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
76                       struct lov_stripe_md *lsm)
77 {
78         int lmm_size;
79         ENTRY;
80
81         lmm_size = sizeof(**lmmp);
82         if (!lmmp)
83                 RETURN(lmm_size);
84
85         if (*lmmp && !lsm) {
86                 OBD_FREE(*lmmp, lmm_size);
87                 *lmmp = NULL;
88                 RETURN(0);
89         }
90
91         if (!*lmmp) {
92                 OBD_ALLOC(*lmmp, lmm_size);
93                 if (!*lmmp)
94                         RETURN(-ENOMEM);
95         }
96
97         if (lsm) {
98                 LASSERT(lsm->lsm_object_id);
99                 LASSERT_MDS_GROUP(lsm->lsm_object_gr);
100                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101                 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
102         }
103
104         RETURN(lmm_size);
105 }
106
107 /* Unpack OSC object metadata from disk storage (LE byte order). */
108 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
109                         struct lov_mds_md *lmm, int lmm_bytes)
110 {
111         int lsm_size;
112         ENTRY;
113
114         if (lmm != NULL) {
115                 if (lmm_bytes < sizeof (*lmm)) {
116                         CERROR("lov_mds_md too small: %d, need %d\n",
117                                lmm_bytes, (int)sizeof(*lmm));
118                         RETURN(-EINVAL);
119                 }
120                 /* XXX LOV_MAGIC etc check? */
121
122                 if (lmm->lmm_object_id == 0) {
123                         CERROR("lov_mds_md: zero lmm_object_id\n");
124                         RETURN(-EINVAL);
125                 }
126         }
127
128         lsm_size = lov_stripe_md_size(1);
129         if (lsmp == NULL)
130                 RETURN(lsm_size);
131
132         if (*lsmp != NULL && lmm == NULL) {
133                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
134                 OBD_FREE(*lsmp, lsm_size);
135                 *lsmp = NULL;
136                 RETURN(0);
137         }
138
139         if (*lsmp == NULL) {
140                 OBD_ALLOC(*lsmp, lsm_size);
141                 if (*lsmp == NULL)
142                         RETURN(-ENOMEM);
143                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
144                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
145                         OBD_FREE(*lsmp, lsm_size);
146                         RETURN(-ENOMEM);
147                 }
148                 loi_init((*lsmp)->lsm_oinfo[0]);
149         }
150
151         if (lmm != NULL) {
152                 /* XXX zero *lsmp? */
153                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
154                 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
155                 LASSERT((*lsmp)->lsm_object_id);
156                 LASSERT_MDS_GROUP((*lsmp)->lsm_object_gr);
157         }
158
159         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
160
161         RETURN(lsm_size);
162 }
163
164 static inline void osc_pack_capa(struct ptlrpc_request *req,
165                                  struct ost_body *body, void *capa)
166 {
167         struct obd_capa *oc = (struct obd_capa *)capa;
168         struct lustre_capa *c;
169
170         if (!capa)
171                 return;
172
173         c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
174         LASSERT(c);
175         capa_cpy(c, oc);
176         body->oa.o_valid |= OBD_MD_FLOSSCAPA;
177         DEBUG_CAPA(D_SEC, c, "pack");
178 }
179
180 static inline void osc_pack_req_body(struct ptlrpc_request *req,
181                                      struct obd_info *oinfo)
182 {
183         struct ost_body *body;
184
185         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
186         LASSERT(body);
187
188         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
189         osc_pack_capa(req, body, oinfo->oi_capa);
190 }
191
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193                                      const struct req_msg_field *field,
194                                      struct obd_capa *oc)
195 {
196         if (oc == NULL)
197                 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
198         else
199                 /* it is already calculated as sizeof struct obd_capa */
200                 ;
201 }
202
203 static int osc_getattr_interpret(const struct lu_env *env,
204                                  struct ptlrpc_request *req,
205                                  struct osc_async_args *aa, int rc)
206 {
207         struct ost_body *body;
208         ENTRY;
209
210         if (rc != 0)
211                 GOTO(out, rc);
212
213         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
214         if (body) {
215                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
216                 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
217
218                 /* This should really be sent by the OST */
219                 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
220                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
221         } else {
222                 CDEBUG(D_INFO, "can't unpack ost_body\n");
223                 rc = -EPROTO;
224                 aa->aa_oi->oi_oa->o_valid = 0;
225         }
226 out:
227         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
228         RETURN(rc);
229 }
230
231 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
232                              struct ptlrpc_request_set *set)
233 {
234         struct ptlrpc_request *req;
235         struct osc_async_args *aa;
236         int                    rc;
237         ENTRY;
238
239         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
240         if (req == NULL)
241                 RETURN(-ENOMEM);
242
243         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
244         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
245         if (rc) {
246                 ptlrpc_request_free(req);
247                 RETURN(rc);
248         }
249
250         osc_pack_req_body(req, oinfo);
251
252         ptlrpc_request_set_replen(req);
253         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
254
255         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
256         aa = ptlrpc_req_async_args(req);
257         aa->aa_oi = oinfo;
258
259         ptlrpc_set_add_req(set, req);
260         RETURN(0);
261 }
262
263 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
264 {
265         struct ptlrpc_request *req;
266         struct ost_body       *body;
267         int                    rc;
268         ENTRY;
269
270         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
271         if (req == NULL)
272                 RETURN(-ENOMEM);
273
274         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
275         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
276         if (rc) {
277                 ptlrpc_request_free(req);
278                 RETURN(rc);
279         }
280
281         osc_pack_req_body(req, oinfo);
282
283         ptlrpc_request_set_replen(req);
284
285         rc = ptlrpc_queue_wait(req);
286         if (rc)
287                 GOTO(out, rc);
288
289         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
290         if (body == NULL)
291                 GOTO(out, rc = -EPROTO);
292
293         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
294         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
295
296         /* This should really be sent by the OST */
297         oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
298         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
299
300         EXIT;
301  out:
302         ptlrpc_req_finished(req);
303         return rc;
304 }
305
306 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
307                        struct obd_trans_info *oti)
308 {
309         struct ptlrpc_request *req;
310         struct ost_body       *body;
311         int                    rc;
312         ENTRY;
313
314         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
315
316         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
317         if (req == NULL)
318                 RETURN(-ENOMEM);
319
320         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
321         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
322         if (rc) {
323                 ptlrpc_request_free(req);
324                 RETURN(rc);
325         }
326
327         osc_pack_req_body(req, oinfo);
328
329         ptlrpc_request_set_replen(req);
330
331         rc = ptlrpc_queue_wait(req);
332         if (rc)
333                 GOTO(out, rc);
334
335         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
336         if (body == NULL)
337                 GOTO(out, rc = -EPROTO);
338
339         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
340
341         EXIT;
342 out:
343         ptlrpc_req_finished(req);
344         RETURN(rc);
345 }
346
347 static int osc_setattr_interpret(const struct lu_env *env,
348                                  struct ptlrpc_request *req,
349                                  struct osc_setattr_args *sa, int rc)
350 {
351         struct ost_body *body;
352         ENTRY;
353
354         if (rc != 0)
355                 GOTO(out, rc);
356
357         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
358         if (body == NULL)
359                 GOTO(out, rc = -EPROTO);
360
361         lustre_get_wire_obdo(sa->sa_oa, &body->oa);
362 out:
363         rc = sa->sa_upcall(sa->sa_cookie, rc);
364         RETURN(rc);
365 }
366
367 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
368                            struct obd_trans_info *oti,
369                            obd_enqueue_update_f upcall, void *cookie,
370                            struct ptlrpc_request_set *rqset)
371 {
372         struct ptlrpc_request   *req;
373         struct osc_setattr_args *sa;
374         int                      rc;
375         ENTRY;
376
377         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
378         if (req == NULL)
379                 RETURN(-ENOMEM);
380
381         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
382         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
383         if (rc) {
384                 ptlrpc_request_free(req);
385                 RETURN(rc);
386         }
387
388         if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
389                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
390
391         osc_pack_req_body(req, oinfo);
392
393         ptlrpc_request_set_replen(req);
394
395         /* do mds to ost setattr asynchronously */
396         if (!rqset) {
397                 /* Do not wait for response. */
398                 ptlrpcd_add_req(req, PSCOPE_OTHER);
399         } else {
400                 req->rq_interpret_reply =
401                         (ptlrpc_interpterer_t)osc_setattr_interpret;
402
403                 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
404                 sa = ptlrpc_req_async_args(req);
405                 sa->sa_oa = oinfo->oi_oa;
406                 sa->sa_upcall = upcall;
407                 sa->sa_cookie = cookie;
408
409                 if (rqset == PTLRPCD_SET)
410                         ptlrpcd_add_req(req, PSCOPE_OTHER);
411                 else
412                         ptlrpc_set_add_req(rqset, req);
413         }
414
415         RETURN(0);
416 }
417
418 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
419                              struct obd_trans_info *oti,
420                              struct ptlrpc_request_set *rqset)
421 {
422         return osc_setattr_async_base(exp, oinfo, oti,
423                                       oinfo->oi_cb_up, oinfo, rqset);
424 }
425
426 int osc_real_create(struct obd_export *exp, struct obdo *oa,
427                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
428 {
429         struct ptlrpc_request *req;
430         struct ost_body       *body;
431         struct lov_stripe_md  *lsm;
432         int                    rc;
433         ENTRY;
434
435         LASSERT(oa);
436         LASSERT(ea);
437
438         lsm = *ea;
439         if (!lsm) {
440                 rc = obd_alloc_memmd(exp, &lsm);
441                 if (rc < 0)
442                         RETURN(rc);
443         }
444
445         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
446         if (req == NULL)
447                 GOTO(out, rc = -ENOMEM);
448
449         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
450         if (rc) {
451                 ptlrpc_request_free(req);
452                 GOTO(out, rc);
453         }
454
455         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
456         LASSERT(body);
457         lustre_set_wire_obdo(&body->oa, oa);
458
459         ptlrpc_request_set_replen(req);
460
461         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
462             oa->o_flags == OBD_FL_DELORPHAN) {
463                 DEBUG_REQ(D_HA, req,
464                           "delorphan from OST integration");
465                 /* Don't resend the delorphan req */
466                 req->rq_no_resend = req->rq_no_delay = 1;
467         }
468
469         rc = ptlrpc_queue_wait(req);
470         if (rc)
471                 GOTO(out_req, rc);
472
473         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
474         if (body == NULL)
475                 GOTO(out_req, rc = -EPROTO);
476
477         lustre_get_wire_obdo(oa, &body->oa);
478
479         /* This should really be sent by the OST */
480         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
481         oa->o_valid |= OBD_MD_FLBLKSZ;
482
483         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
484          * have valid lsm_oinfo data structs, so don't go touching that.
485          * This needs to be fixed in a big way.
486          */
487         lsm->lsm_object_id = oa->o_id;
488         lsm->lsm_object_gr = oa->o_gr;
489         *ea = lsm;
490
491         if (oti != NULL) {
492                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
493
494                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
495                         if (!oti->oti_logcookies)
496                                 oti_alloc_cookies(oti, 1);
497                         *oti->oti_logcookies = oa->o_lcookie;
498                 }
499         }
500
501         CDEBUG(D_HA, "transno: "LPD64"\n",
502                lustre_msg_get_transno(req->rq_repmsg));
503 out_req:
504         ptlrpc_req_finished(req);
505 out:
506         if (rc && !*ea)
507                 obd_free_memmd(exp, &lsm);
508         RETURN(rc);
509 }
510
511 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
512                    obd_enqueue_update_f upcall, void *cookie,
513                    struct ptlrpc_request_set *rqset)
514 {
515         struct ptlrpc_request   *req;
516         struct osc_setattr_args *sa;
517         struct ost_body         *body;
518         int                      rc;
519         ENTRY;
520
521         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
522         if (req == NULL)
523                 RETURN(-ENOMEM);
524
525         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
526         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
527         if (rc) {
528                 ptlrpc_request_free(req);
529                 RETURN(rc);
530         }
531         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
532         ptlrpc_at_set_req_timeout(req);
533
534         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
535         LASSERT(body);
536         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
537         osc_pack_capa(req, body, oinfo->oi_capa);
538
539         ptlrpc_request_set_replen(req);
540
541
542         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
543         CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
544         sa = ptlrpc_req_async_args(req);
545         sa->sa_oa     = oinfo->oi_oa;
546         sa->sa_upcall = upcall;
547         sa->sa_cookie = cookie;
548         if (rqset == PTLRPCD_SET)
549                 ptlrpcd_add_req(req, PSCOPE_OTHER);
550         else
551                 ptlrpc_set_add_req(rqset, req);
552
553         RETURN(0);
554 }
555
556 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
557                      struct obd_trans_info *oti,
558                      struct ptlrpc_request_set *rqset)
559 {
560         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
561         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
562         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
563         return osc_punch_base(exp, oinfo,
564                               oinfo->oi_cb_up, oinfo, rqset);
565 }
566
567 static int osc_sync(struct obd_export *exp, struct obdo *oa,
568                     struct lov_stripe_md *md, obd_size start, obd_size end,
569                     void *capa)
570 {
571         struct ptlrpc_request *req;
572         struct ost_body       *body;
573         int                    rc;
574         ENTRY;
575
576         if (!oa) {
577                 CDEBUG(D_INFO, "oa NULL\n");
578                 RETURN(-EINVAL);
579         }
580
581         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
582         if (req == NULL)
583                 RETURN(-ENOMEM);
584
585         osc_set_capa_size(req, &RMF_CAPA1, capa);
586         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
587         if (rc) {
588                 ptlrpc_request_free(req);
589                 RETURN(rc);
590         }
591
592         /* overload the size and blocks fields in the oa with start/end */
593         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
594         LASSERT(body);
595         lustre_set_wire_obdo(&body->oa, oa);
596         body->oa.o_size = start;
597         body->oa.o_blocks = end;
598         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
599         osc_pack_capa(req, body, capa);
600
601         ptlrpc_request_set_replen(req);
602
603         rc = ptlrpc_queue_wait(req);
604         if (rc)
605                 GOTO(out, rc);
606
607         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
608         if (body == NULL)
609                 GOTO(out, rc = -EPROTO);
610
611         lustre_get_wire_obdo(oa, &body->oa);
612
613         EXIT;
614  out:
615         ptlrpc_req_finished(req);
616         return rc;
617 }
618
619 /* Find and cancel locally locks matched by @mode in the resource found by
620  * @objid. Found locks are added into @cancel list. Returns the amount of
621  * locks added to @cancels list. */
622 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
623                                    cfs_list_t *cancels,
624                                    ldlm_mode_t mode, int lock_flags)
625 {
626         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
627         struct ldlm_res_id res_id;
628         struct ldlm_resource *res;
629         int count;
630         ENTRY;
631
632         osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
633         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
634         if (res == NULL)
635                 RETURN(0);
636
637         LDLM_RESOURCE_ADDREF(res);
638         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
639                                            lock_flags, 0, NULL);
640         LDLM_RESOURCE_DELREF(res);
641         ldlm_resource_putref(res);
642         RETURN(count);
643 }
644
645 static int osc_destroy_interpret(const struct lu_env *env,
646                                  struct ptlrpc_request *req, void *data,
647                                  int rc)
648 {
649         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
650
651         cfs_atomic_dec(&cli->cl_destroy_in_flight);
652         cfs_waitq_signal(&cli->cl_destroy_waitq);
653         return 0;
654 }
655
656 static int osc_can_send_destroy(struct client_obd *cli)
657 {
658         if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
659             cli->cl_max_rpcs_in_flight) {
660                 /* The destroy request can be sent */
661                 return 1;
662         }
663         if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
664             cli->cl_max_rpcs_in_flight) {
665                 /*
666                  * The counter has been modified between the two atomic
667                  * operations.
668                  */
669                 cfs_waitq_signal(&cli->cl_destroy_waitq);
670         }
671         return 0;
672 }
673
674 /* Destroy requests can be async always on the client, and we don't even really
675  * care about the return code since the client cannot do anything at all about
676  * a destroy failure.
677  * When the MDS is unlinking a filename, it saves the file objects into a
678  * recovery llog, and these object records are cancelled when the OST reports
679  * they were destroyed and sync'd to disk (i.e. transaction committed).
680  * If the client dies, or the OST is down when the object should be destroyed,
681  * the records are not cancelled, and when the OST reconnects to the MDS next,
682  * it will retrieve the llog unlink logs and then sends the log cancellation
683  * cookies to the MDS after committing destroy transactions. */
684 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
685                        struct lov_stripe_md *ea, struct obd_trans_info *oti,
686                        struct obd_export *md_export, void *capa)
687 {
688         struct client_obd     *cli = &exp->exp_obd->u.cli;
689         struct ptlrpc_request *req;
690         struct ost_body       *body;
691         CFS_LIST_HEAD(cancels);
692         int rc, count;
693         ENTRY;
694
695         if (!oa) {
696                 CDEBUG(D_INFO, "oa NULL\n");
697                 RETURN(-EINVAL);
698         }
699
700         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
701                                         LDLM_FL_DISCARD_DATA);
702
703         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
704         if (req == NULL) {
705                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
706                 RETURN(-ENOMEM);
707         }
708
709         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
710         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
711                                0, &cancels, count);
712         if (rc) {
713                 ptlrpc_request_free(req);
714                 RETURN(rc);
715         }
716
717         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
718         ptlrpc_at_set_req_timeout(req);
719
720         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
721                 oa->o_lcookie = *oti->oti_logcookies;
722         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
723         LASSERT(body);
724         lustre_set_wire_obdo(&body->oa, oa);
725
726         osc_pack_capa(req, body, (struct obd_capa *)capa);
727         ptlrpc_request_set_replen(req);
728
729         /* don't throttle destroy RPCs for the MDT */
730         if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
731                 req->rq_interpret_reply = osc_destroy_interpret;
732                 if (!osc_can_send_destroy(cli)) {
733                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
734                                                           NULL);
735
736                         /*
737                          * Wait until the number of on-going destroy RPCs drops
738                          * under max_rpc_in_flight
739                          */
740                         l_wait_event_exclusive(cli->cl_destroy_waitq,
741                                                osc_can_send_destroy(cli), &lwi);
742                 }
743         }
744
745         /* Do not wait for response */
746         ptlrpcd_add_req(req, PSCOPE_OTHER);
747         RETURN(0);
748 }
749
750 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
751                                 long writing_bytes)
752 {
753         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
754
755         LASSERT(!(oa->o_valid & bits));
756
757         oa->o_valid |= bits;
758         client_obd_list_lock(&cli->cl_loi_list_lock);
759         oa->o_dirty = cli->cl_dirty;
760         if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
761                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
762                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
763                 oa->o_undirty = 0;
764         } else if (cfs_atomic_read(&obd_dirty_pages) -
765                    cfs_atomic_read(&obd_dirty_transit_pages) >
766                    obd_max_dirty_pages + 1){
767                 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
768                  * not covered by a lock thus they may safely race and trip
769                  * this CERROR() unless we add in a small fudge factor (+1). */
770                 CERROR("dirty %d - %d > system dirty_max %d\n",
771                        cfs_atomic_read(&obd_dirty_pages),
772                        cfs_atomic_read(&obd_dirty_transit_pages),
773                        obd_max_dirty_pages);
774                 oa->o_undirty = 0;
775         } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
776                 CERROR("dirty %lu - dirty_max %lu too big???\n",
777                        cli->cl_dirty, cli->cl_dirty_max);
778                 oa->o_undirty = 0;
779         } else {
780                 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
781                                 (cli->cl_max_rpcs_in_flight + 1);
782                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
783         }
784         oa->o_grant = cli->cl_avail_grant;
785         oa->o_dropped = cli->cl_lost_grant;
786         cli->cl_lost_grant = 0;
787         client_obd_list_unlock(&cli->cl_loi_list_lock);
788         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
789                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
790
791 }
792
793 static void osc_update_next_shrink(struct client_obd *cli)
794 {
795         cli->cl_next_shrink_grant =
796                 cfs_time_shift(cli->cl_grant_shrink_interval);
797         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
798                cli->cl_next_shrink_grant);
799 }
800
801 /* caller must hold loi_list_lock */
802 static void osc_consume_write_grant(struct client_obd *cli,
803                                     struct brw_page *pga)
804 {
805         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
806         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
807         cfs_atomic_inc(&obd_dirty_pages);
808         cli->cl_dirty += CFS_PAGE_SIZE;
809         cli->cl_avail_grant -= CFS_PAGE_SIZE;
810         pga->flag |= OBD_BRW_FROM_GRANT;
811         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
812                CFS_PAGE_SIZE, pga, pga->pg);
813         LASSERT(cli->cl_avail_grant >= 0);
814         osc_update_next_shrink(cli);
815 }
816
817 /* the companion to osc_consume_write_grant, called when a brw has completed.
818  * must be called with the loi lock held. */
819 static void osc_release_write_grant(struct client_obd *cli,
820                                     struct brw_page *pga, int sent)
821 {
822         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
823         ENTRY;
824
825         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
826         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
827                 EXIT;
828                 return;
829         }
830
831         pga->flag &= ~OBD_BRW_FROM_GRANT;
832         cfs_atomic_dec(&obd_dirty_pages);
833         cli->cl_dirty -= CFS_PAGE_SIZE;
834         if (pga->flag & OBD_BRW_NOCACHE) {
835                 pga->flag &= ~OBD_BRW_NOCACHE;
836                 cfs_atomic_dec(&obd_dirty_transit_pages);
837                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
838         }
839         if (!sent) {
840                 cli->cl_lost_grant += CFS_PAGE_SIZE;
841                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
842                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
843         } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
844                 /* For short writes we shouldn't count parts of pages that
845                  * span a whole block on the OST side, or our accounting goes
846                  * wrong.  Should match the code in filter_grant_check. */
847                 int offset = pga->off & ~CFS_PAGE_MASK;
848                 int count = pga->count + (offset & (blocksize - 1));
849                 int end = (offset + pga->count) & (blocksize - 1);
850                 if (end)
851                         count += blocksize - end;
852
853                 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
854                 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
855                        CFS_PAGE_SIZE - count, cli->cl_lost_grant,
856                        cli->cl_avail_grant, cli->cl_dirty);
857         }
858
859         EXIT;
860 }
861
862 static unsigned long rpcs_in_flight(struct client_obd *cli)
863 {
864         return cli->cl_r_in_flight + cli->cl_w_in_flight;
865 }
866
867 /* caller must hold loi_list_lock */
868 void osc_wake_cache_waiters(struct client_obd *cli)
869 {
870         cfs_list_t *l, *tmp;
871         struct osc_cache_waiter *ocw;
872
873         ENTRY;
874         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
875                 /* if we can't dirty more, we must wait until some is written */
876                 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
877                    (cfs_atomic_read(&obd_dirty_pages) + 1 >
878                     obd_max_dirty_pages)) {
879                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
880                                "osc max %ld, sys max %d\n", cli->cl_dirty,
881                                cli->cl_dirty_max, obd_max_dirty_pages);
882                         return;
883                 }
884
885                 /* if still dirty cache but no grant wait for pending RPCs that
886                  * may yet return us some grant before doing sync writes */
887                 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
888                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
889                                cli->cl_w_in_flight);
890                         return;
891                 }
892
893                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
894                 cfs_list_del_init(&ocw->ocw_entry);
895                 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
896                         /* no more RPCs in flight to return grant, do sync IO */
897                         ocw->ocw_rc = -EDQUOT;
898                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
899                 } else {
900                         osc_consume_write_grant(cli,
901                                                 &ocw->ocw_oap->oap_brw_page);
902                 }
903
904                 cfs_waitq_signal(&ocw->ocw_waitq);
905         }
906
907         EXIT;
908 }
909
910 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
911 {
912         client_obd_list_lock(&cli->cl_loi_list_lock);
913         cli->cl_avail_grant += grant;
914         client_obd_list_unlock(&cli->cl_loi_list_lock);
915 }
916
917 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
918 {
919         if (body->oa.o_valid & OBD_MD_FLGRANT) {
920                 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
921                 __osc_update_grant(cli, body->oa.o_grant);
922         }
923 }
924
925 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
926                               void *key, obd_count vallen, void *val,
927                               struct ptlrpc_request_set *set);
928
929 static int osc_shrink_grant_interpret(const struct lu_env *env,
930                                       struct ptlrpc_request *req,
931                                       void *aa, int rc)
932 {
933         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
934         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
935         struct ost_body *body;
936
937         if (rc != 0) {
938                 __osc_update_grant(cli, oa->o_grant);
939                 GOTO(out, rc);
940         }
941
942         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
943         LASSERT(body);
944         osc_update_grant(cli, body);
945 out:
946         OBD_FREE_PTR(oa);
947         return rc;
948 }
949
950 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
951 {
952         client_obd_list_lock(&cli->cl_loi_list_lock);
953         oa->o_grant = cli->cl_avail_grant / 4;
954         cli->cl_avail_grant -= oa->o_grant;
955         client_obd_list_unlock(&cli->cl_loi_list_lock);
956         oa->o_flags |= OBD_FL_SHRINK_GRANT;
957         osc_update_next_shrink(cli);
958 }
959
960 /* Shrink the current grant, either from some large amount to enough for a
961  * full set of in-flight RPCs, or if we have already shrunk to that limit
962  * then to enough for a single RPC.  This avoids keeping more grant than
963  * needed, and avoids shrinking the grant piecemeal. */
964 static int osc_shrink_grant(struct client_obd *cli)
965 {
966         long target = (cli->cl_max_rpcs_in_flight + 1) *
967                       cli->cl_max_pages_per_rpc;
968
969         client_obd_list_lock(&cli->cl_loi_list_lock);
970         if (cli->cl_avail_grant <= target)
971                 target = cli->cl_max_pages_per_rpc;
972         client_obd_list_unlock(&cli->cl_loi_list_lock);
973
974         return osc_shrink_grant_to_target(cli, target);
975 }
976
977 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
978 {
979         int    rc = 0;
980         struct ost_body     *body;
981         ENTRY;
982
983         client_obd_list_lock(&cli->cl_loi_list_lock);
984         /* Don't shrink if we are already above or below the desired limit
985          * We don't want to shrink below a single RPC, as that will negatively
986          * impact block allocation and long-term performance. */
987         if (target < cli->cl_max_pages_per_rpc)
988                 target = cli->cl_max_pages_per_rpc;
989
990         if (target >= cli->cl_avail_grant) {
991                 client_obd_list_unlock(&cli->cl_loi_list_lock);
992                 RETURN(0);
993         }
994         client_obd_list_unlock(&cli->cl_loi_list_lock);
995
996         OBD_ALLOC_PTR(body);
997         if (!body)
998                 RETURN(-ENOMEM);
999
1000         osc_announce_cached(cli, &body->oa, 0);
1001
1002         client_obd_list_lock(&cli->cl_loi_list_lock);
1003         body->oa.o_grant = cli->cl_avail_grant - target;
1004         cli->cl_avail_grant = target;
1005         client_obd_list_unlock(&cli->cl_loi_list_lock);
1006         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1007         osc_update_next_shrink(cli);
1008
1009         rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1010                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1011                                 sizeof(*body), body, NULL);
1012         if (rc != 0)
1013                 __osc_update_grant(cli, body->oa.o_grant);
1014         OBD_FREE_PTR(body);
1015         RETURN(rc);
1016 }
1017
1018 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1019 static int osc_should_shrink_grant(struct client_obd *client)
1020 {
1021         cfs_time_t time = cfs_time_current();
1022         cfs_time_t next_shrink = client->cl_next_shrink_grant;
1023         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1024                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1025                     client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1026                         return 1;
1027                 else
1028                         osc_update_next_shrink(client);
1029         }
1030         return 0;
1031 }
1032
1033 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1034 {
1035         struct client_obd *client;
1036
1037         cfs_list_for_each_entry(client, &item->ti_obd_list,
1038                                 cl_grant_shrink_list) {
1039                 if (osc_should_shrink_grant(client))
1040                         osc_shrink_grant(client);
1041         }
1042         return 0;
1043 }
1044
1045 static int osc_add_shrink_grant(struct client_obd *client)
1046 {
1047         int rc;
1048
1049         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1050                                        TIMEOUT_GRANT,
1051                                        osc_grant_shrink_grant_cb, NULL,
1052                                        &client->cl_grant_shrink_list);
1053         if (rc) {
1054                 CERROR("add grant client %s error %d\n",
1055                         client->cl_import->imp_obd->obd_name, rc);
1056                 return rc;
1057         }
1058         CDEBUG(D_CACHE, "add grant client %s \n",
1059                client->cl_import->imp_obd->obd_name);
1060         osc_update_next_shrink(client);
1061         return 0;
1062 }
1063
1064 static int osc_del_shrink_grant(struct client_obd *client)
1065 {
1066         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1067                                          TIMEOUT_GRANT);
1068 }
1069
1070 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1071 {
1072         /*
1073          * ocd_grant is the total grant amount we're expect to hold: if we've
1074          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1075          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1076          *
1077          * race is tolerable here: if we're evicted, but imp_state already
1078          * left EVICTED state, then cl_dirty must be 0 already.
1079          */
1080         client_obd_list_lock(&cli->cl_loi_list_lock);
1081         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1082                 cli->cl_avail_grant = ocd->ocd_grant;
1083         else
1084                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1085         client_obd_list_unlock(&cli->cl_loi_list_lock);
1086
1087         CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1088                cli->cl_avail_grant, cli->cl_lost_grant);
1089         LASSERT(cli->cl_avail_grant >= 0);
1090
1091         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1092             cfs_list_empty(&cli->cl_grant_shrink_list))
1093                 osc_add_shrink_grant(cli);
1094 }
1095
1096 /* We assume that the reason this OSC got a short read is because it read
1097  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1098  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1099  * this stripe never got written at or beyond this stripe offset yet. */
1100 static void handle_short_read(int nob_read, obd_count page_count,
1101                               struct brw_page **pga)
1102 {
1103         char *ptr;
1104         int i = 0;
1105
1106         /* skip bytes read OK */
1107         while (nob_read > 0) {
1108                 LASSERT (page_count > 0);
1109
1110                 if (pga[i]->count > nob_read) {
1111                         /* EOF inside this page */
1112                         ptr = cfs_kmap(pga[i]->pg) +
1113                                 (pga[i]->off & ~CFS_PAGE_MASK);
1114                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1115                         cfs_kunmap(pga[i]->pg);
1116                         page_count--;
1117                         i++;
1118                         break;
1119                 }
1120
1121                 nob_read -= pga[i]->count;
1122                 page_count--;
1123                 i++;
1124         }
1125
1126         /* zero remaining pages */
1127         while (page_count-- > 0) {
1128                 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1129                 memset(ptr, 0, pga[i]->count);
1130                 cfs_kunmap(pga[i]->pg);
1131                 i++;
1132         }
1133 }
1134
1135 static int check_write_rcs(struct ptlrpc_request *req,
1136                            int requested_nob, int niocount,
1137                            obd_count page_count, struct brw_page **pga)
1138 {
1139         int     i;
1140         __u32   *remote_rcs;
1141
1142         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1143                                                   sizeof(*remote_rcs) *
1144                                                   niocount);
1145         if (remote_rcs == NULL) {
1146                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1147                 return(-EPROTO);
1148         }
1149
1150         /* return error if any niobuf was in error */
1151         for (i = 0; i < niocount; i++) {
1152                 if (remote_rcs[i] < 0)
1153                         return(remote_rcs[i]);
1154
1155                 if (remote_rcs[i] != 0) {
1156                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1157                                 i, remote_rcs[i], req);
1158                         return(-EPROTO);
1159                 }
1160         }
1161
1162         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1163                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1164                        req->rq_bulk->bd_nob_transferred, requested_nob);
1165                 return(-EPROTO);
1166         }
1167
1168         return (0);
1169 }
1170
1171 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1172 {
1173         if (p1->flag != p2->flag) {
1174                 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1175                                   OBD_BRW_NOCACHE|OBD_BRW_SYNC);
1176
1177                 /* warn if we try to combine flags that we don't know to be
1178                  * safe to combine */
1179                 if ((p1->flag & mask) != (p2->flag & mask))
1180                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
1181                                "same brw?\n", p1->flag, p2->flag);
1182                 return 0;
1183         }
1184
1185         return (p1->off + p1->count == p2->off);
1186 }
1187
1188 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1189                                    struct brw_page **pga, int opc,
1190                                    cksum_type_t cksum_type)
1191 {
1192         __u32 cksum;
1193         int i = 0;
1194
1195         LASSERT (pg_count > 0);
1196         cksum = init_checksum(cksum_type);
1197         while (nob > 0 && pg_count > 0) {
1198                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1199                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1200                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1201
1202                 /* corrupt the data before we compute the checksum, to
1203                  * simulate an OST->client data error */
1204                 if (i == 0 && opc == OST_READ &&
1205                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1206                         memcpy(ptr + off, "bad1", min(4, nob));
1207                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1208                 cfs_kunmap(pga[i]->pg);
1209                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1210                                off, cksum);
1211
1212                 nob -= pga[i]->count;
1213                 pg_count--;
1214                 i++;
1215         }
1216         /* For sending we only compute the wrong checksum instead
1217          * of corrupting the data so it is still correct on a redo */
1218         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1219                 cksum++;
1220
1221         return cksum;
1222 }
1223
1224 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1225                                 struct lov_stripe_md *lsm, obd_count page_count,
1226                                 struct brw_page **pga,
1227                                 struct ptlrpc_request **reqp,
1228                                 struct obd_capa *ocapa, int reserve)
1229 {
1230         struct ptlrpc_request   *req;
1231         struct ptlrpc_bulk_desc *desc;
1232         struct ost_body         *body;
1233         struct obd_ioobj        *ioobj;
1234         struct niobuf_remote    *niobuf;
1235         int niocount, i, requested_nob, opc, rc;
1236         struct osc_brw_async_args *aa;
1237         struct req_capsule      *pill;
1238         struct brw_page *pg_prev;
1239
1240         ENTRY;
1241         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1242                 RETURN(-ENOMEM); /* Recoverable */
1243         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1244                 RETURN(-EINVAL); /* Fatal */
1245
1246         if ((cmd & OBD_BRW_WRITE) != 0) {
1247                 opc = OST_WRITE;
1248                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1249                                                 cli->cl_import->imp_rq_pool,
1250                                                 &RQF_OST_BRW_WRITE);
1251         } else {
1252                 opc = OST_READ;
1253                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1254         }
1255         if (req == NULL)
1256                 RETURN(-ENOMEM);
1257
1258         for (niocount = i = 1; i < page_count; i++) {
1259                 if (!can_merge_pages(pga[i - 1], pga[i]))
1260                         niocount++;
1261         }
1262
1263         pill = &req->rq_pill;
1264         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1265                              sizeof(*ioobj));
1266         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1267                              niocount * sizeof(*niobuf));
1268         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1269
1270         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1271         if (rc) {
1272                 ptlrpc_request_free(req);
1273                 RETURN(rc);
1274         }
1275         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1276         ptlrpc_at_set_req_timeout(req);
1277
1278         if (opc == OST_WRITE)
1279                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1280                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1281         else
1282                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1283                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1284
1285         if (desc == NULL)
1286                 GOTO(out, rc = -ENOMEM);
1287         /* NB request now owns desc and will free it when it gets freed */
1288
1289         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1290         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1291         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1292         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1293
1294         lustre_set_wire_obdo(&body->oa, oa);
1295
1296         obdo_to_ioobj(oa, ioobj);
1297         ioobj->ioo_bufcnt = niocount;
1298         osc_pack_capa(req, body, ocapa);
1299         LASSERT (page_count > 0);
1300         pg_prev = pga[0];
1301         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1302                 struct brw_page *pg = pga[i];
1303
1304                 LASSERT(pg->count > 0);
1305                 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1306                          "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1307                          pg->off, pg->count);
1308 #ifdef __linux__
1309                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1310                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1311                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1312                          i, page_count,
1313                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1314                          pg_prev->pg, page_private(pg_prev->pg),
1315                          pg_prev->pg->index, pg_prev->off);
1316 #else
1317                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1318                          "i %d p_c %u\n", i, page_count);
1319 #endif
1320                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1321                         (pg->flag & OBD_BRW_SRVLOCK));
1322
1323                 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1324                                       pg->count);
1325                 requested_nob += pg->count;
1326
1327                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1328                         niobuf--;
1329                         niobuf->len += pg->count;
1330                 } else {
1331                         niobuf->offset = pg->off;
1332                         niobuf->len    = pg->count;
1333                         niobuf->flags  = pg->flag;
1334                 }
1335                 pg_prev = pg;
1336         }
1337
1338         LASSERTF((void *)(niobuf - niocount) ==
1339                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1340                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1341                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1342
1343         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1344         if (osc_should_shrink_grant(cli))
1345                 osc_shrink_grant_local(cli, &body->oa);
1346
1347         /* size[REQ_REC_OFF] still sizeof (*body) */
1348         if (opc == OST_WRITE) {
1349                 if (unlikely(cli->cl_checksum) &&
1350                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1351                         /* store cl_cksum_type in a local variable since
1352                          * it can be changed via lprocfs */
1353                         cksum_type_t cksum_type = cli->cl_cksum_type;
1354
1355                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1356                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1357                                 body->oa.o_flags = 0;
1358                         }
1359                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1360                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1361                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1362                                                              page_count, pga,
1363                                                              OST_WRITE,
1364                                                              cksum_type);
1365                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1366                                body->oa.o_cksum);
1367                         /* save this in 'oa', too, for later checking */
1368                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1369                         oa->o_flags |= cksum_type_pack(cksum_type);
1370                 } else {
1371                         /* clear out the checksum flag, in case this is a
1372                          * resend but cl_checksum is no longer set. b=11238 */
1373                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1374                 }
1375                 oa->o_cksum = body->oa.o_cksum;
1376                 /* 1 RC per niobuf */
1377                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1378                                      sizeof(__u32) * niocount);
1379         } else {
1380                 if (unlikely(cli->cl_checksum) &&
1381                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1382                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1383                                 body->oa.o_flags = 0;
1384                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1385                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1386                 }
1387         }
1388         ptlrpc_request_set_replen(req);
1389
1390         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1391         aa = ptlrpc_req_async_args(req);
1392         aa->aa_oa = oa;
1393         aa->aa_requested_nob = requested_nob;
1394         aa->aa_nio_count = niocount;
1395         aa->aa_page_count = page_count;
1396         aa->aa_resends = 0;
1397         aa->aa_ppga = pga;
1398         aa->aa_cli = cli;
1399         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1400         if (ocapa && reserve)
1401                 aa->aa_ocapa = capa_get(ocapa);
1402
1403         *reqp = req;
1404         RETURN(0);
1405
1406  out:
1407         ptlrpc_req_finished(req);
1408         RETURN(rc);
1409 }
1410
1411 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1412                                 __u32 client_cksum, __u32 server_cksum, int nob,
1413                                 obd_count page_count, struct brw_page **pga,
1414                                 cksum_type_t client_cksum_type)
1415 {
1416         __u32 new_cksum;
1417         char *msg;
1418         cksum_type_t cksum_type;
1419
1420         if (server_cksum == client_cksum) {
1421                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1422                 return 0;
1423         }
1424
1425         if (oa->o_valid & OBD_MD_FLFLAGS)
1426                 cksum_type = cksum_type_unpack(oa->o_flags);
1427         else
1428                 cksum_type = OBD_CKSUM_CRC32;
1429
1430         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1431                                       cksum_type);
1432
1433         if (cksum_type != client_cksum_type)
1434                 msg = "the server did not use the checksum type specified in "
1435                       "the original request - likely a protocol problem";
1436         else if (new_cksum == server_cksum)
1437                 msg = "changed on the client after we checksummed it - "
1438                       "likely false positive due to mmap IO (bug 11742)";
1439         else if (new_cksum == client_cksum)
1440                 msg = "changed in transit before arrival at OST";
1441         else
1442                 msg = "changed in transit AND doesn't match the original - "
1443                       "likely false positive due to mmap IO (bug 11742)";
1444
1445         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1446                            LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1447                            "["LPU64"-"LPU64"]\n",
1448                            msg, libcfs_nid2str(peer->nid),
1449                            oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1450                            oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1451                                                         (__u64)0,
1452                            oa->o_id,
1453                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1454                            pga[0]->off,
1455                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1456         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1457                "client csum now %x\n", client_cksum, client_cksum_type,
1458                server_cksum, cksum_type, new_cksum);
1459         return 1;
1460 }
1461
1462 /* Note rc enters this function as number of bytes transferred */
1463 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1464 {
1465         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1466         const lnet_process_id_t *peer =
1467                         &req->rq_import->imp_connection->c_peer;
1468         struct client_obd *cli = aa->aa_cli;
1469         struct ost_body *body;
1470         __u32 client_cksum = 0;
1471         ENTRY;
1472
1473         if (rc < 0 && rc != -EDQUOT) {
1474                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1475                 RETURN(rc);
1476         }
1477
1478         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1479         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1480         if (body == NULL) {
1481                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1482                 RETURN(-EPROTO);
1483         }
1484
1485 #ifdef HAVE_QUOTA_SUPPORT
1486         /* set/clear over quota flag for a uid/gid */
1487         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1488             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1489                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1490
1491                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1492                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1493                        body->oa.o_flags);
1494                 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1495                              body->oa.o_flags);
1496         }
1497 #endif
1498
1499         if (rc < 0)
1500                 RETURN(rc);
1501
1502         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1503                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1504
1505         osc_update_grant(cli, body);
1506
1507         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1508                 if (rc > 0) {
1509                         CERROR("Unexpected +ve rc %d\n", rc);
1510                         RETURN(-EPROTO);
1511                 }
1512                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1513
1514                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1515                         RETURN(-EAGAIN);
1516
1517                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1518                     check_write_checksum(&body->oa, peer, client_cksum,
1519                                          body->oa.o_cksum, aa->aa_requested_nob,
1520                                          aa->aa_page_count, aa->aa_ppga,
1521                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1522                         RETURN(-EAGAIN);
1523
1524                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1525                                      aa->aa_page_count, aa->aa_ppga);
1526                 GOTO(out, rc);
1527         }
1528
1529         /* The rest of this function executes only for OST_READs */
1530
1531         /* if unwrap_bulk failed, return -EAGAIN to retry */
1532         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1533         if (rc < 0)
1534                 GOTO(out, rc = -EAGAIN);
1535
1536         if (rc > aa->aa_requested_nob) {
1537                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1538                        aa->aa_requested_nob);
1539                 RETURN(-EPROTO);
1540         }
1541
1542         if (rc != req->rq_bulk->bd_nob_transferred) {
1543                 CERROR ("Unexpected rc %d (%d transferred)\n",
1544                         rc, req->rq_bulk->bd_nob_transferred);
1545                 return (-EPROTO);
1546         }
1547
1548         if (rc < aa->aa_requested_nob)
1549                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1550
1551         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1552                 static int cksum_counter;
1553                 __u32      server_cksum = body->oa.o_cksum;
1554                 char      *via;
1555                 char      *router;
1556                 cksum_type_t cksum_type;
1557
1558                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1559                         cksum_type = cksum_type_unpack(body->oa.o_flags);
1560                 else
1561                         cksum_type = OBD_CKSUM_CRC32;
1562                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1563                                                  aa->aa_ppga, OST_READ,
1564                                                  cksum_type);
1565
1566                 if (peer->nid == req->rq_bulk->bd_sender) {
1567                         via = router = "";
1568                 } else {
1569                         via = " via ";
1570                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1571                 }
1572
1573                 if (server_cksum == ~0 && rc > 0) {
1574                         CERROR("Protocol error: server %s set the 'checksum' "
1575                                "bit, but didn't send a checksum.  Not fatal, "
1576                                "but please notify on http://bugzilla.lustre.org/\n",
1577                                libcfs_nid2str(peer->nid));
1578                 } else if (server_cksum != client_cksum) {
1579                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1580                                            "%s%s%s inum "LPU64"/"LPU64" object "
1581                                            LPU64"/"LPU64" extent "
1582                                            "["LPU64"-"LPU64"]\n",
1583                                            req->rq_import->imp_obd->obd_name,
1584                                            libcfs_nid2str(peer->nid),
1585                                            via, router,
1586                                            body->oa.o_valid & OBD_MD_FLFID ?
1587                                                 body->oa.o_fid : (__u64)0,
1588                                            body->oa.o_valid & OBD_MD_FLFID ?
1589                                                 body->oa.o_generation :(__u64)0,
1590                                            body->oa.o_id,
1591                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1592                                                 body->oa.o_gr : (__u64)0,
1593                                            aa->aa_ppga[0]->off,
1594                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1595                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1596                                                                         1);
1597                         CERROR("client %x, server %x, cksum_type %x\n",
1598                                client_cksum, server_cksum, cksum_type);
1599                         cksum_counter = 0;
1600                         aa->aa_oa->o_cksum = client_cksum;
1601                         rc = -EAGAIN;
1602                 } else {
1603                         cksum_counter++;
1604                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1605                         rc = 0;
1606                 }
1607         } else if (unlikely(client_cksum)) {
1608                 static int cksum_missed;
1609
1610                 cksum_missed++;
1611                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1612                         CERROR("Checksum %u requested from %s but not sent\n",
1613                                cksum_missed, libcfs_nid2str(peer->nid));
1614         } else {
1615                 rc = 0;
1616         }
1617 out:
1618         if (rc >= 0)
1619                 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1620
1621         RETURN(rc);
1622 }
1623
1624 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1625                             struct lov_stripe_md *lsm,
1626                             obd_count page_count, struct brw_page **pga,
1627                             struct obd_capa *ocapa)
1628 {
1629         struct ptlrpc_request *req;
1630         int                    rc;
1631         cfs_waitq_t            waitq;
1632         int                    resends = 0;
1633         struct l_wait_info     lwi;
1634
1635         ENTRY;
1636
1637         cfs_waitq_init(&waitq);
1638
1639 restart_bulk:
1640         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1641                                   page_count, pga, &req, ocapa, 0);
1642         if (rc != 0)
1643                 return (rc);
1644
1645         rc = ptlrpc_queue_wait(req);
1646
1647         if (rc == -ETIMEDOUT && req->rq_resend) {
1648                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1649                 ptlrpc_req_finished(req);
1650                 goto restart_bulk;
1651         }
1652
1653         rc = osc_brw_fini_request(req, rc);
1654
1655         ptlrpc_req_finished(req);
1656         if (osc_recoverable_error(rc)) {
1657                 resends++;
1658                 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1659                         CERROR("too many resend retries, returning error\n");
1660                         RETURN(-EIO);
1661                 }
1662
1663                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1664                 l_wait_event(waitq, 0, &lwi);
1665
1666                 goto restart_bulk;
1667         }
1668
1669         RETURN (rc);
1670 }
1671
1672 int osc_brw_redo_request(struct ptlrpc_request *request,
1673                          struct osc_brw_async_args *aa)
1674 {
1675         struct ptlrpc_request *new_req;
1676         struct ptlrpc_request_set *set = request->rq_set;
1677         struct osc_brw_async_args *new_aa;
1678         struct osc_async_page *oap;
1679         int rc = 0;
1680         ENTRY;
1681
1682         if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1683                 CERROR("too many resend retries, returning error\n");
1684                 RETURN(-EIO);
1685         }
1686
1687         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1688
1689         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1690                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1691                                   aa->aa_cli, aa->aa_oa,
1692                                   NULL /* lsm unused by osc currently */,
1693                                   aa->aa_page_count, aa->aa_ppga,
1694                                   &new_req, aa->aa_ocapa, 0);
1695         if (rc)
1696                 RETURN(rc);
1697
1698         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1699
1700         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1701                 if (oap->oap_request != NULL) {
1702                         LASSERTF(request == oap->oap_request,
1703                                  "request %p != oap_request %p\n",
1704                                  request, oap->oap_request);
1705                         if (oap->oap_interrupted) {
1706                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1707                                 ptlrpc_req_finished(new_req);
1708                                 RETURN(-EINTR);
1709                         }
1710                 }
1711         }
1712         /* New request takes over pga and oaps from old request.
1713          * Note that copying a list_head doesn't work, need to move it... */
1714         aa->aa_resends++;
1715         new_req->rq_interpret_reply = request->rq_interpret_reply;
1716         new_req->rq_async_args = request->rq_async_args;
1717         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1718
1719         new_aa = ptlrpc_req_async_args(new_req);
1720
1721         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1722         cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1723         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1724
1725         cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1726                 if (oap->oap_request) {
1727                         ptlrpc_req_finished(oap->oap_request);
1728                         oap->oap_request = ptlrpc_request_addref(new_req);
1729                 }
1730         }
1731
1732         new_aa->aa_ocapa = aa->aa_ocapa;
1733         aa->aa_ocapa = NULL;
1734
1735         /* use ptlrpc_set_add_req is safe because interpret functions work
1736          * in check_set context. only one way exist with access to request
1737          * from different thread got -EINTR - this way protected with
1738          * cl_loi_list_lock */
1739         ptlrpc_set_add_req(set, new_req);
1740
1741         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1742
1743         DEBUG_REQ(D_INFO, new_req, "new request");
1744         RETURN(0);
1745 }
1746
1747 /*
1748  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1749  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1750  * fine for our small page arrays and doesn't require allocation.  its an
1751  * insertion sort that swaps elements that are strides apart, shrinking the
1752  * stride down until its '1' and the array is sorted.
1753  */
1754 static void sort_brw_pages(struct brw_page **array, int num)
1755 {
1756         int stride, i, j;
1757         struct brw_page *tmp;
1758
1759         if (num == 1)
1760                 return;
1761         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1762                 ;
1763
1764         do {
1765                 stride /= 3;
1766                 for (i = stride ; i < num ; i++) {
1767                         tmp = array[i];
1768                         j = i;
1769                         while (j >= stride && array[j - stride]->off > tmp->off) {
1770                                 array[j] = array[j - stride];
1771                                 j -= stride;
1772                         }
1773                         array[j] = tmp;
1774                 }
1775         } while (stride > 1);
1776 }
1777
1778 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1779 {
1780         int count = 1;
1781         int offset;
1782         int i = 0;
1783
1784         LASSERT (pages > 0);
1785         offset = pg[i]->off & ~CFS_PAGE_MASK;
1786
1787         for (;;) {
1788                 pages--;
1789                 if (pages == 0)         /* that's all */
1790                         return count;
1791
1792                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1793                         return count;   /* doesn't end on page boundary */
1794
1795                 i++;
1796                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1797                 if (offset != 0)        /* doesn't start on page boundary */
1798                         return count;
1799
1800                 count++;
1801         }
1802 }
1803
1804 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1805 {
1806         struct brw_page **ppga;
1807         int i;
1808
1809         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1810         if (ppga == NULL)
1811                 return NULL;
1812
1813         for (i = 0; i < count; i++)
1814                 ppga[i] = pga + i;
1815         return ppga;
1816 }
1817
1818 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1819 {
1820         LASSERT(ppga != NULL);
1821         OBD_FREE(ppga, sizeof(*ppga) * count);
1822 }
1823
1824 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1825                    obd_count page_count, struct brw_page *pga,
1826                    struct obd_trans_info *oti)
1827 {
1828         struct obdo *saved_oa = NULL;
1829         struct brw_page **ppga, **orig;
1830         struct obd_import *imp = class_exp2cliimp(exp);
1831         struct client_obd *cli;
1832         int rc, page_count_orig;
1833         ENTRY;
1834
1835         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1836         cli = &imp->imp_obd->u.cli;
1837
1838         if (cmd & OBD_BRW_CHECK) {
1839                 /* The caller just wants to know if there's a chance that this
1840                  * I/O can succeed */
1841
1842                 if (imp->imp_invalid)
1843                         RETURN(-EIO);
1844                 RETURN(0);
1845         }
1846
1847         /* test_brw with a failed create can trip this, maybe others. */
1848         LASSERT(cli->cl_max_pages_per_rpc);
1849
1850         rc = 0;
1851
1852         orig = ppga = osc_build_ppga(pga, page_count);
1853         if (ppga == NULL)
1854                 RETURN(-ENOMEM);
1855         page_count_orig = page_count;
1856
1857         sort_brw_pages(ppga, page_count);
1858         while (page_count) {
1859                 obd_count pages_per_brw;
1860
1861                 if (page_count > cli->cl_max_pages_per_rpc)
1862                         pages_per_brw = cli->cl_max_pages_per_rpc;
1863                 else
1864                         pages_per_brw = page_count;
1865
1866                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1867
1868                 if (saved_oa != NULL) {
1869                         /* restore previously saved oa */
1870                         *oinfo->oi_oa = *saved_oa;
1871                 } else if (page_count > pages_per_brw) {
1872                         /* save a copy of oa (brw will clobber it) */
1873                         OBDO_ALLOC(saved_oa);
1874                         if (saved_oa == NULL)
1875                                 GOTO(out, rc = -ENOMEM);
1876                         *saved_oa = *oinfo->oi_oa;
1877                 }
1878
1879                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1880                                       pages_per_brw, ppga, oinfo->oi_capa);
1881
1882                 if (rc != 0)
1883                         break;
1884
1885                 page_count -= pages_per_brw;
1886                 ppga += pages_per_brw;
1887         }
1888
1889 out:
1890         osc_release_ppga(orig, page_count_orig);
1891
1892         if (saved_oa != NULL)
1893                 OBDO_FREE(saved_oa);
1894
1895         RETURN(rc);
1896 }
1897
1898 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1899  * the dirty accounting.  Writeback completes or truncate happens before
1900  * writing starts.  Must be called with the loi lock held. */
1901 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1902                            int sent)
1903 {
1904         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1905 }
1906
1907
1908 /* This maintains the lists of pending pages to read/write for a given object
1909  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1910  * to quickly find objects that are ready to send an RPC. */
1911 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1912                          int cmd)
1913 {
1914         int optimal;
1915         ENTRY;
1916
1917         if (lop->lop_num_pending == 0)
1918                 RETURN(0);
1919
1920         /* if we have an invalid import we want to drain the queued pages
1921          * by forcing them through rpcs that immediately fail and complete
1922          * the pages.  recovery relies on this to empty the queued pages
1923          * before canceling the locks and evicting down the llite pages */
1924         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1925                 RETURN(1);
1926
1927         /* stream rpcs in queue order as long as as there is an urgent page
1928          * queued.  this is our cheap solution for good batching in the case
1929          * where writepage marks some random page in the middle of the file
1930          * as urgent because of, say, memory pressure */
1931         if (!cfs_list_empty(&lop->lop_urgent)) {
1932                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1933                 RETURN(1);
1934         }
1935         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1936         optimal = cli->cl_max_pages_per_rpc;
1937         if (cmd & OBD_BRW_WRITE) {
1938                 /* trigger a write rpc stream as long as there are dirtiers
1939                  * waiting for space.  as they're waiting, they're not going to
1940                  * create more pages to coallesce with what's waiting.. */
1941                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
1942                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1943                         RETURN(1);
1944                 }
1945                 /* +16 to avoid triggering rpcs that would want to include pages
1946                  * that are being queued but which can't be made ready until
1947                  * the queuer finishes with the page. this is a wart for
1948                  * llite::commit_write() */
1949                 optimal += 16;
1950         }
1951         if (lop->lop_num_pending >= optimal)
1952                 RETURN(1);
1953
1954         RETURN(0);
1955 }
1956
1957 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1958 {
1959         struct osc_async_page *oap;
1960         ENTRY;
1961
1962         if (cfs_list_empty(&lop->lop_urgent))
1963                 RETURN(0);
1964
1965         oap = cfs_list_entry(lop->lop_urgent.next,
1966                          struct osc_async_page, oap_urgent_item);
1967
1968         if (oap->oap_async_flags & ASYNC_HP) {
1969                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1970                 RETURN(1);
1971         }
1972
1973         RETURN(0);
1974 }
1975
1976 static void on_list(cfs_list_t *item, cfs_list_t *list,
1977                     int should_be_on)
1978 {
1979         if (cfs_list_empty(item) && should_be_on)
1980                 cfs_list_add_tail(item, list);
1981         else if (!cfs_list_empty(item) && !should_be_on)
1982                 cfs_list_del_init(item);
1983 }
1984
1985 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1986  * can find pages to build into rpcs quickly */
1987 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1988 {
1989         if (lop_makes_hprpc(&loi->loi_write_lop) ||
1990             lop_makes_hprpc(&loi->loi_read_lop)) {
1991                 /* HP rpc */
1992                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1993                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1994         } else {
1995                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1996                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
1997                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
1998                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1999         }
2000
2001         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2002                 loi->loi_write_lop.lop_num_pending);
2003
2004         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2005                 loi->loi_read_lop.lop_num_pending);
2006 }
2007
2008 static void lop_update_pending(struct client_obd *cli,
2009                                struct loi_oap_pages *lop, int cmd, int delta)
2010 {
2011         lop->lop_num_pending += delta;
2012         if (cmd & OBD_BRW_WRITE)
2013                 cli->cl_pending_w_pages += delta;
2014         else
2015                 cli->cl_pending_r_pages += delta;
2016 }
2017
2018 /**
2019  * this is called when a sync waiter receives an interruption.  Its job is to
2020  * get the caller woken as soon as possible.  If its page hasn't been put in an
2021  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2022  * desiring interruption which will forcefully complete the rpc once the rpc
2023  * has timed out.
2024  */
2025 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2026 {
2027         struct loi_oap_pages *lop;
2028         struct lov_oinfo *loi;
2029         int rc = -EBUSY;
2030         ENTRY;
2031
2032         LASSERT(!oap->oap_interrupted);
2033         oap->oap_interrupted = 1;
2034
2035         /* ok, it's been put in an rpc. only one oap gets a request reference */
2036         if (oap->oap_request != NULL) {
2037                 ptlrpc_mark_interrupted(oap->oap_request);
2038                 ptlrpcd_wake(oap->oap_request);
2039                 ptlrpc_req_finished(oap->oap_request);
2040                 oap->oap_request = NULL;
2041         }
2042
2043         /*
2044          * page completion may be called only if ->cpo_prep() method was
2045          * executed by osc_io_submit(), that also adds page the to pending list
2046          */
2047         if (!cfs_list_empty(&oap->oap_pending_item)) {
2048                 cfs_list_del_init(&oap->oap_pending_item);
2049                 cfs_list_del_init(&oap->oap_urgent_item);
2050
2051                 loi = oap->oap_loi;
2052                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2053                         &loi->loi_write_lop : &loi->loi_read_lop;
2054                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2055                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2056                 rc = oap->oap_caller_ops->ap_completion(env,
2057                                           oap->oap_caller_data,
2058                                           oap->oap_cmd, NULL, -EINTR);
2059         }
2060
2061         RETURN(rc);
2062 }
2063
2064 /* this is trying to propogate async writeback errors back up to the
2065  * application.  As an async write fails we record the error code for later if
2066  * the app does an fsync.  As long as errors persist we force future rpcs to be
2067  * sync so that the app can get a sync error and break the cycle of queueing
2068  * pages for which writeback will fail. */
2069 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2070                            int rc)
2071 {
2072         if (rc) {
2073                 if (!ar->ar_rc)
2074                         ar->ar_rc = rc;
2075
2076                 ar->ar_force_sync = 1;
2077                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2078                 return;
2079
2080         }
2081
2082         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2083                 ar->ar_force_sync = 0;
2084 }
2085
2086 void osc_oap_to_pending(struct osc_async_page *oap)
2087 {
2088         struct loi_oap_pages *lop;
2089
2090         if (oap->oap_cmd & OBD_BRW_WRITE)
2091                 lop = &oap->oap_loi->loi_write_lop;
2092         else
2093                 lop = &oap->oap_loi->loi_read_lop;
2094
2095         if (oap->oap_async_flags & ASYNC_HP)
2096                 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2097         else if (oap->oap_async_flags & ASYNC_URGENT)
2098                 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2099         cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2100         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2101 }
2102
2103 /* this must be called holding the loi list lock to give coverage to exit_cache,
2104  * async_flag maintenance, and oap_request */
2105 static void osc_ap_completion(const struct lu_env *env,
2106                               struct client_obd *cli, struct obdo *oa,
2107                               struct osc_async_page *oap, int sent, int rc)
2108 {
2109         __u64 xid = 0;
2110
2111         ENTRY;
2112         if (oap->oap_request != NULL) {
2113                 xid = ptlrpc_req_xid(oap->oap_request);
2114                 ptlrpc_req_finished(oap->oap_request);
2115                 oap->oap_request = NULL;
2116         }
2117
2118         cfs_spin_lock(&oap->oap_lock);
2119         oap->oap_async_flags = 0;
2120         cfs_spin_unlock(&oap->oap_lock);
2121         oap->oap_interrupted = 0;
2122
2123         if (oap->oap_cmd & OBD_BRW_WRITE) {
2124                 osc_process_ar(&cli->cl_ar, xid, rc);
2125                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2126         }
2127
2128         if (rc == 0 && oa != NULL) {
2129                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2130                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2131                 if (oa->o_valid & OBD_MD_FLMTIME)
2132                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2133                 if (oa->o_valid & OBD_MD_FLATIME)
2134                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2135                 if (oa->o_valid & OBD_MD_FLCTIME)
2136                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2137         }
2138
2139         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2140                                                 oap->oap_cmd, oa, rc);
2141
2142         /* ll_ap_completion (from llite) drops PG_locked. so, a new
2143          * I/O on the page could start, but OSC calls it under lock
2144          * and thus we can add oap back to pending safely */
2145         if (rc)
2146                 /* upper layer wants to leave the page on pending queue */
2147                 osc_oap_to_pending(oap);
2148         else
2149                 osc_exit_cache(cli, oap, sent);
2150         EXIT;
2151 }
2152
2153 static int brw_interpret(const struct lu_env *env,
2154                          struct ptlrpc_request *req, void *data, int rc)
2155 {
2156         struct osc_brw_async_args *aa = data;
2157         struct client_obd *cli;
2158         int async;
2159         ENTRY;
2160
2161         rc = osc_brw_fini_request(req, rc);
2162         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2163         if (osc_recoverable_error(rc)) {
2164                 rc = osc_brw_redo_request(req, aa);
2165                 if (rc == 0)
2166                         RETURN(0);
2167         }
2168
2169         if (aa->aa_ocapa) {
2170                 capa_put(aa->aa_ocapa);
2171                 aa->aa_ocapa = NULL;
2172         }
2173
2174         cli = aa->aa_cli;
2175
2176         client_obd_list_lock(&cli->cl_loi_list_lock);
2177
2178         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2179          * is called so we know whether to go to sync BRWs or wait for more
2180          * RPCs to complete */
2181         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2182                 cli->cl_w_in_flight--;
2183         else
2184                 cli->cl_r_in_flight--;
2185
2186         async = cfs_list_empty(&aa->aa_oaps);
2187         if (!async) { /* from osc_send_oap_rpc() */
2188                 struct osc_async_page *oap, *tmp;
2189                 /* the caller may re-use the oap after the completion call so
2190                  * we need to clean it up a little */
2191                 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2192                                              oap_rpc_item) {
2193                         cfs_list_del_init(&oap->oap_rpc_item);
2194                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2195                 }
2196                 OBDO_FREE(aa->aa_oa);
2197         } else { /* from async_internal() */
2198                 int i;
2199                 for (i = 0; i < aa->aa_page_count; i++)
2200                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2201
2202                 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2203                         OBDO_FREE(aa->aa_oa);
2204         }
2205         osc_wake_cache_waiters(cli);
2206         osc_check_rpcs(env, cli);
2207         client_obd_list_unlock(&cli->cl_loi_list_lock);
2208         if (!async)
2209                 cl_req_completion(env, aa->aa_clerq, rc);
2210         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2211         RETURN(rc);
2212 }
2213
2214 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2215                                             struct client_obd *cli,
2216                                             cfs_list_t *rpc_list,
2217                                             int page_count, int cmd)
2218 {
2219         struct ptlrpc_request *req;
2220         struct brw_page **pga = NULL;
2221         struct osc_brw_async_args *aa;
2222         struct obdo *oa = NULL;
2223         const struct obd_async_page_ops *ops = NULL;
2224         void *caller_data = NULL;
2225         struct osc_async_page *oap;
2226         struct osc_async_page *tmp;
2227         struct ost_body *body;
2228         struct cl_req *clerq = NULL;
2229         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2230         struct ldlm_lock *lock = NULL;
2231         struct cl_req_attr crattr;
2232         int i, rc;
2233
2234         ENTRY;
2235         LASSERT(!cfs_list_empty(rpc_list));
2236
2237         memset(&crattr, 0, sizeof crattr);
2238         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2239         if (pga == NULL)
2240                 GOTO(out, req = ERR_PTR(-ENOMEM));
2241
2242         OBDO_ALLOC(oa);
2243         if (oa == NULL)
2244                 GOTO(out, req = ERR_PTR(-ENOMEM));
2245
2246         i = 0;
2247         cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2248                 struct cl_page *page = osc_oap2cl_page(oap);
2249                 if (ops == NULL) {
2250                         ops = oap->oap_caller_ops;
2251                         caller_data = oap->oap_caller_data;
2252
2253                         clerq = cl_req_alloc(env, page, crt,
2254                                              1 /* only 1-object rpcs for
2255                                                 * now */);
2256                         if (IS_ERR(clerq))
2257                                 GOTO(out, req = (void *)clerq);
2258                         lock = oap->oap_ldlm_lock;
2259                 }
2260                 pga[i] = &oap->oap_brw_page;
2261                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2262                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2263                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2264                 i++;
2265                 cl_req_page_add(env, clerq, page);
2266         }
2267
2268         /* always get the data for the obdo for the rpc */
2269         LASSERT(ops != NULL);
2270         crattr.cra_oa = oa;
2271         crattr.cra_capa = NULL;
2272         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2273         if (lock) {
2274                 oa->o_handle = lock->l_remote_handle;
2275                 oa->o_valid |= OBD_MD_FLHANDLE;
2276         }
2277
2278         rc = cl_req_prep(env, clerq);
2279         if (rc != 0) {
2280                 CERROR("cl_req_prep failed: %d\n", rc);
2281                 GOTO(out, req = ERR_PTR(rc));
2282         }
2283
2284         sort_brw_pages(pga, page_count);
2285         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2286                                   pga, &req, crattr.cra_capa, 1);
2287         if (rc != 0) {
2288                 CERROR("prep_req failed: %d\n", rc);
2289                 GOTO(out, req = ERR_PTR(rc));
2290         }
2291
2292         /* Need to update the timestamps after the request is built in case
2293          * we race with setattr (locally or in queue at OST).  If OST gets
2294          * later setattr before earlier BRW (as determined by the request xid),
2295          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2296          * way to do this in a single call.  bug 10150 */
2297         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2298         cl_req_attr_set(env, clerq, &crattr,
2299                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2300
2301         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2302         aa = ptlrpc_req_async_args(req);
2303         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2304         cfs_list_splice(rpc_list, &aa->aa_oaps);
2305         CFS_INIT_LIST_HEAD(rpc_list);
2306         aa->aa_clerq = clerq;
2307 out:
2308         capa_put(crattr.cra_capa);
2309         if (IS_ERR(req)) {
2310                 if (oa)
2311                         OBDO_FREE(oa);
2312                 if (pga)
2313                         OBD_FREE(pga, sizeof(*pga) * page_count);
2314                 /* this should happen rarely and is pretty bad, it makes the
2315                  * pending list not follow the dirty order */
2316                 client_obd_list_lock(&cli->cl_loi_list_lock);
2317                 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2318                         cfs_list_del_init(&oap->oap_rpc_item);
2319
2320                         /* queued sync pages can be torn down while the pages
2321                          * were between the pending list and the rpc */
2322                         if (oap->oap_interrupted) {
2323                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2324                                 osc_ap_completion(env, cli, NULL, oap, 0,
2325                                                   oap->oap_count);
2326                                 continue;
2327                         }
2328                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2329                 }
2330                 if (clerq && !IS_ERR(clerq))
2331                         cl_req_completion(env, clerq, PTR_ERR(req));
2332         }
2333         RETURN(req);
2334 }
2335
2336 /**
2337  * prepare pages for ASYNC io and put pages in send queue.
2338  *
2339  * \param cmd OBD_BRW_* macroses
2340  * \param lop pending pages
2341  *
2342  * \return zero if pages successfully add to send queue.
2343  * \return not zere if error occurring.
2344  */
2345 static int
2346 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2347                  struct lov_oinfo *loi,
2348                  int cmd, struct loi_oap_pages *lop)
2349 {
2350         struct ptlrpc_request *req;
2351         obd_count page_count = 0;
2352         struct osc_async_page *oap = NULL, *tmp;
2353         struct osc_brw_async_args *aa;
2354         const struct obd_async_page_ops *ops;
2355         CFS_LIST_HEAD(rpc_list);
2356         CFS_LIST_HEAD(tmp_list);
2357         unsigned int ending_offset;
2358         unsigned  starting_offset = 0;
2359         int srvlock = 0;
2360         struct cl_object *clob = NULL;
2361         ENTRY;
2362
2363         /* ASYNC_HP pages first. At present, when the lock the pages is
2364          * to be canceled, the pages covered by the lock will be sent out
2365          * with ASYNC_HP. We have to send out them as soon as possible. */
2366         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2367                 if (oap->oap_async_flags & ASYNC_HP) 
2368                         cfs_list_move(&oap->oap_pending_item, &tmp_list);
2369                 else
2370                         cfs_list_move_tail(&oap->oap_pending_item, &tmp_list);
2371                 if (++page_count >= cli->cl_max_pages_per_rpc)
2372                         break;
2373         }
2374
2375         cfs_list_splice(&tmp_list, &lop->lop_pending);
2376         page_count = 0;
2377
2378         /* first we find the pages we're allowed to work with */
2379         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2380                                      oap_pending_item) {
2381                 ops = oap->oap_caller_ops;
2382
2383                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2384                          "magic 0x%x\n", oap, oap->oap_magic);
2385
2386                 if (clob == NULL) {
2387                         /* pin object in memory, so that completion call-backs
2388                          * can be safely called under client_obd_list lock. */
2389                         clob = osc_oap2cl_page(oap)->cp_obj;
2390                         cl_object_get(clob);
2391                 }
2392
2393                 if (page_count != 0 &&
2394                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2395                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2396                                " oap %p, page %p, srvlock %u\n",
2397                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2398                         break;
2399                 }
2400
2401                 /* If there is a gap at the start of this page, it can't merge
2402                  * with any previous page, so we'll hand the network a
2403                  * "fragmented" page array that it can't transfer in 1 RDMA */
2404                 if (page_count != 0 && oap->oap_page_off != 0)
2405                         break;
2406
2407                 /* in llite being 'ready' equates to the page being locked
2408                  * until completion unlocks it.  commit_write submits a page
2409                  * as not ready because its unlock will happen unconditionally
2410                  * as the call returns.  if we race with commit_write giving
2411                  * us that page we dont' want to create a hole in the page
2412                  * stream, so we stop and leave the rpc to be fired by
2413                  * another dirtier or kupdated interval (the not ready page
2414                  * will still be on the dirty list).  we could call in
2415                  * at the end of ll_file_write to process the queue again. */
2416                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2417                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2418                                                     cmd);
2419                         if (rc < 0)
2420                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2421                                                 "instead of ready\n", oap,
2422                                                 oap->oap_page, rc);
2423                         switch (rc) {
2424                         case -EAGAIN:
2425                                 /* llite is telling us that the page is still
2426                                  * in commit_write and that we should try
2427                                  * and put it in an rpc again later.  we
2428                                  * break out of the loop so we don't create
2429                                  * a hole in the sequence of pages in the rpc
2430                                  * stream.*/
2431                                 oap = NULL;
2432                                 break;
2433                         case -EINTR:
2434                                 /* the io isn't needed.. tell the checks
2435                                  * below to complete the rpc with EINTR */
2436                                 cfs_spin_lock(&oap->oap_lock);
2437                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2438                                 cfs_spin_unlock(&oap->oap_lock);
2439                                 oap->oap_count = -EINTR;
2440                                 break;
2441                         case 0:
2442                                 cfs_spin_lock(&oap->oap_lock);
2443                                 oap->oap_async_flags |= ASYNC_READY;
2444                                 cfs_spin_unlock(&oap->oap_lock);
2445                                 break;
2446                         default:
2447                                 LASSERTF(0, "oap %p page %p returned %d "
2448                                             "from make_ready\n", oap,
2449                                             oap->oap_page, rc);
2450                                 break;
2451                         }
2452                 }
2453                 if (oap == NULL)
2454                         break;
2455                 /*
2456                  * Page submitted for IO has to be locked. Either by
2457                  * ->ap_make_ready() or by higher layers.
2458                  */
2459 #if defined(__KERNEL__) && defined(__linux__)
2460                 {
2461                         struct cl_page *page;
2462
2463                         page = osc_oap2cl_page(oap);
2464
2465                         if (page->cp_type == CPT_CACHEABLE &&
2466                             !(PageLocked(oap->oap_page) &&
2467                               (CheckWriteback(oap->oap_page, cmd)))) {
2468                                 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2469                                        oap->oap_page,
2470                                        (long)oap->oap_page->flags,
2471                                        oap->oap_async_flags);
2472                                 LBUG();
2473                         }
2474                 }
2475 #endif
2476
2477                 /* take the page out of our book-keeping */
2478                 cfs_list_del_init(&oap->oap_pending_item);
2479                 lop_update_pending(cli, lop, cmd, -1);
2480                 cfs_list_del_init(&oap->oap_urgent_item);
2481
2482                 if (page_count == 0)
2483                         starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2484                                           (PTLRPC_MAX_BRW_SIZE - 1);
2485
2486                 /* ask the caller for the size of the io as the rpc leaves. */
2487                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2488                         oap->oap_count =
2489                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2490                                                       cmd);
2491                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2492                 }
2493                 if (oap->oap_count <= 0) {
2494                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2495                                oap->oap_count);
2496                         osc_ap_completion(env, cli, NULL,
2497                                           oap, 0, oap->oap_count);
2498                         continue;
2499                 }
2500
2501                 /* now put the page back in our accounting */
2502                 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2503                 if (page_count == 0)
2504                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2505                 if (++page_count >= cli->cl_max_pages_per_rpc)
2506                         break;
2507
2508                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2509                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2510                  * have the same alignment as the initial writes that allocated
2511                  * extents on the server. */
2512                 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2513                                  oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2514                 if (ending_offset == 0)
2515                         break;
2516
2517                 /* If there is a gap at the end of this page, it can't merge
2518                  * with any subsequent pages, so we'll hand the network a
2519                  * "fragmented" page array that it can't transfer in 1 RDMA */
2520                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2521                         break;
2522         }
2523
2524         osc_wake_cache_waiters(cli);
2525
2526         loi_list_maint(cli, loi);
2527
2528         client_obd_list_unlock(&cli->cl_loi_list_lock);
2529
2530         if (clob != NULL)
2531                 cl_object_put(env, clob);
2532
2533         if (page_count == 0) {
2534                 client_obd_list_lock(&cli->cl_loi_list_lock);
2535                 RETURN(0);
2536         }
2537
2538         req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2539         if (IS_ERR(req)) {
2540                 LASSERT(cfs_list_empty(&rpc_list));
2541                 loi_list_maint(cli, loi);
2542                 RETURN(PTR_ERR(req));
2543         }
2544
2545         aa = ptlrpc_req_async_args(req);
2546
2547         if (cmd == OBD_BRW_READ) {
2548                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2549                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2550                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2551                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2552         } else {
2553                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2554                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2555                                  cli->cl_w_in_flight);
2556                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2557                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2558         }
2559         ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2560
2561         client_obd_list_lock(&cli->cl_loi_list_lock);
2562
2563         if (cmd == OBD_BRW_READ)
2564                 cli->cl_r_in_flight++;
2565         else
2566                 cli->cl_w_in_flight++;
2567
2568         /* queued sync pages can be torn down while the pages
2569          * were between the pending list and the rpc */
2570         tmp = NULL;
2571         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2572                 /* only one oap gets a request reference */
2573                 if (tmp == NULL)
2574                         tmp = oap;
2575                 if (oap->oap_interrupted && !req->rq_intr) {
2576                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2577                                oap, req);
2578                         ptlrpc_mark_interrupted(req);
2579                 }
2580         }
2581         if (tmp != NULL)
2582                 tmp->oap_request = ptlrpc_request_addref(req);
2583
2584         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2585                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2586
2587         req->rq_interpret_reply = brw_interpret;
2588         ptlrpcd_add_req(req, PSCOPE_BRW);
2589         RETURN(1);
2590 }
2591
2592 #define LOI_DEBUG(LOI, STR, args...)                                     \
2593         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2594                !cfs_list_empty(&(LOI)->loi_ready_item) ||                \
2595                !cfs_list_empty(&(LOI)->loi_hp_ready_item),               \
2596                (LOI)->loi_write_lop.lop_num_pending,                     \
2597                !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent),        \
2598                (LOI)->loi_read_lop.lop_num_pending,                      \
2599                !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent),         \
2600                args)                                                     \
2601
2602 /* This is called by osc_check_rpcs() to find which objects have pages that
2603  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2604 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2605 {
2606         ENTRY;
2607
2608         /* First return objects that have blocked locks so that they
2609          * will be flushed quickly and other clients can get the lock,
2610          * then objects which have pages ready to be stuffed into RPCs */
2611         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2612                 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2613                                       struct lov_oinfo, loi_hp_ready_item));
2614         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2615                 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2616                                       struct lov_oinfo, loi_ready_item));
2617
2618         /* then if we have cache waiters, return all objects with queued
2619          * writes.  This is especially important when many small files
2620          * have filled up the cache and not been fired into rpcs because
2621          * they don't pass the nr_pending/object threshhold */
2622         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2623             !cfs_list_empty(&cli->cl_loi_write_list))
2624                 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2625                                       struct lov_oinfo, loi_write_item));
2626
2627         /* then return all queued objects when we have an invalid import
2628          * so that they get flushed */
2629         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2630                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2631                         RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2632                                               struct lov_oinfo,
2633                                               loi_write_item));
2634                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2635                         RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2636                                               struct lov_oinfo, loi_read_item));
2637         }
2638         RETURN(NULL);
2639 }
2640
2641 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2642 {
2643         struct osc_async_page *oap;
2644         int hprpc = 0;
2645
2646         if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2647                 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2648                                      struct osc_async_page, oap_urgent_item);
2649                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2650         }
2651
2652         if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2653                 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2654                                      struct osc_async_page, oap_urgent_item);
2655                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2656         }
2657
2658         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2659 }
2660
2661 /* called with the loi list lock held */
2662 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2663 {
2664         struct lov_oinfo *loi;
2665         int rc = 0, race_counter = 0;
2666         ENTRY;
2667
2668         while ((loi = osc_next_loi(cli)) != NULL) {
2669                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2670
2671                 if (osc_max_rpc_in_flight(cli, loi))
2672                         break;
2673
2674                 /* attempt some read/write balancing by alternating between
2675                  * reads and writes in an object.  The makes_rpc checks here
2676                  * would be redundant if we were getting read/write work items
2677                  * instead of objects.  we don't want send_oap_rpc to drain a
2678                  * partial read pending queue when we're given this object to
2679                  * do io on writes while there are cache waiters */
2680                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2681                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2682                                               &loi->loi_write_lop);
2683                         if (rc < 0) {
2684                                 CERROR("Write request failed with %d\n", rc);
2685
2686                                 /* osc_send_oap_rpc failed, mostly because of
2687                                  * memory pressure.
2688                                  *
2689                                  * It can't break here, because if:
2690                                  *  - a page was submitted by osc_io_submit, so
2691                                  *    page locked;
2692                                  *  - no request in flight
2693                                  *  - no subsequent request
2694                                  * The system will be in live-lock state,
2695                                  * because there is no chance to call
2696                                  * osc_io_unplug() and osc_check_rpcs() any
2697                                  * more. pdflush can't help in this case,
2698                                  * because it might be blocked at grabbing
2699                                  * the page lock as we mentioned.
2700                                  *
2701                                  * Anyway, continue to drain pages. */
2702                                 /* break; */
2703                         }
2704
2705                         if (rc > 0)
2706                                 race_counter = 0;
2707                         else
2708                                 race_counter++;
2709                 }
2710                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2711                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2712                                               &loi->loi_read_lop);
2713                         if (rc < 0)
2714                                 CERROR("Read request failed with %d\n", rc);
2715
2716                         if (rc > 0)
2717                                 race_counter = 0;
2718                         else
2719                                 race_counter++;
2720                 }
2721
2722                 /* attempt some inter-object balancing by issueing rpcs
2723                  * for each object in turn */
2724                 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2725                         cfs_list_del_init(&loi->loi_hp_ready_item);
2726                 if (!cfs_list_empty(&loi->loi_ready_item))
2727                         cfs_list_del_init(&loi->loi_ready_item);
2728                 if (!cfs_list_empty(&loi->loi_write_item))
2729                         cfs_list_del_init(&loi->loi_write_item);
2730                 if (!cfs_list_empty(&loi->loi_read_item))
2731                         cfs_list_del_init(&loi->loi_read_item);
2732
2733                 loi_list_maint(cli, loi);
2734
2735                 /* send_oap_rpc fails with 0 when make_ready tells it to
2736                  * back off.  llite's make_ready does this when it tries
2737                  * to lock a page queued for write that is already locked.
2738                  * we want to try sending rpcs from many objects, but we
2739                  * don't want to spin failing with 0.  */
2740                 if (race_counter == 10)
2741                         break;
2742         }
2743         EXIT;
2744 }
2745
2746 /* we're trying to queue a page in the osc so we're subject to the
2747  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2748  * If the osc's queued pages are already at that limit, then we want to sleep
2749  * until there is space in the osc's queue for us.  We also may be waiting for
2750  * write credits from the OST if there are RPCs in flight that may return some
2751  * before we fall back to sync writes.
2752  *
2753  * We need this know our allocation was granted in the presence of signals */
2754 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2755 {
2756         int rc;
2757         ENTRY;
2758         client_obd_list_lock(&cli->cl_loi_list_lock);
2759         rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2760         client_obd_list_unlock(&cli->cl_loi_list_lock);
2761         RETURN(rc);
2762 };
2763
2764 /**
2765  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2766  * is available.
2767  */
2768 int osc_enter_cache_try(const struct lu_env *env,
2769                         struct client_obd *cli, struct lov_oinfo *loi,
2770                         struct osc_async_page *oap, int transient)
2771 {
2772         int has_grant;
2773
2774         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2775         if (has_grant) {
2776                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2777                 if (transient) {
2778                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2779                         cfs_atomic_inc(&obd_dirty_transit_pages);
2780                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2781                 }
2782         }
2783         return has_grant;
2784 }
2785
2786 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2787  * grant or cache space. */
2788 static int osc_enter_cache(const struct lu_env *env,
2789                            struct client_obd *cli, struct lov_oinfo *loi,
2790                            struct osc_async_page *oap)
2791 {
2792         struct osc_cache_waiter ocw;
2793         struct l_wait_info lwi = { 0 };
2794
2795         ENTRY;
2796
2797         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2798                "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2799                cli->cl_dirty_max, obd_max_dirty_pages,
2800                cli->cl_lost_grant, cli->cl_avail_grant);
2801
2802         /* force the caller to try sync io.  this can jump the list
2803          * of queued writes and create a discontiguous rpc stream */
2804         if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2805             loi->loi_ar.ar_force_sync)
2806                 RETURN(-EDQUOT);
2807
2808         /* Hopefully normal case - cache space and write credits available */
2809         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2810             cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2811             osc_enter_cache_try(env, cli, loi, oap, 0))
2812                 RETURN(0);
2813
2814         /* It is safe to block as a cache waiter as long as there is grant
2815          * space available or the hope of additional grant being returned
2816          * when an in flight write completes.  Using the write back cache
2817          * if possible is preferable to sending the data synchronously
2818          * because write pages can then be merged in to large requests.
2819          * The addition of this cache waiter will causing pending write
2820          * pages to be sent immediately. */
2821         if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2822                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2823                 cfs_waitq_init(&ocw.ocw_waitq);
2824                 ocw.ocw_oap = oap;
2825                 ocw.ocw_rc = 0;
2826
2827                 loi_list_maint(cli, loi);
2828                 osc_check_rpcs(env, cli);
2829                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2830
2831                 CDEBUG(D_CACHE, "sleeping for cache space\n");
2832                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2833
2834                 client_obd_list_lock(&cli->cl_loi_list_lock);
2835                 if (!cfs_list_empty(&ocw.ocw_entry)) {
2836                         cfs_list_del(&ocw.ocw_entry);
2837                         RETURN(-EINTR);
2838                 }
2839                 RETURN(ocw.ocw_rc);
2840         }
2841
2842         RETURN(-EDQUOT);
2843 }
2844
2845
2846 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2847                         struct lov_oinfo *loi, cfs_page_t *page,
2848                         obd_off offset, const struct obd_async_page_ops *ops,
2849                         void *data, void **res, int nocache,
2850                         struct lustre_handle *lockh)
2851 {
2852         struct osc_async_page *oap;
2853
2854         ENTRY;
2855
2856         if (!page)
2857                 return cfs_size_round(sizeof(*oap));
2858
2859         oap = *res;
2860         oap->oap_magic = OAP_MAGIC;
2861         oap->oap_cli = &exp->exp_obd->u.cli;
2862         oap->oap_loi = loi;
2863
2864         oap->oap_caller_ops = ops;
2865         oap->oap_caller_data = data;
2866
2867         oap->oap_page = page;
2868         oap->oap_obj_off = offset;
2869         if (!client_is_remote(exp) &&
2870             cfs_capable(CFS_CAP_SYS_RESOURCE))
2871                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2872
2873         LASSERT(!(offset & ~CFS_PAGE_MASK));
2874
2875         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2876         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2877         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2878         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2879
2880         cfs_spin_lock_init(&oap->oap_lock);
2881         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2882         RETURN(0);
2883 }
2884
2885 struct osc_async_page *oap_from_cookie(void *cookie)
2886 {
2887         struct osc_async_page *oap = cookie;
2888         if (oap->oap_magic != OAP_MAGIC)
2889                 return ERR_PTR(-EINVAL);
2890         return oap;
2891 };
2892
2893 int osc_queue_async_io(const struct lu_env *env,
2894                        struct obd_export *exp, struct lov_stripe_md *lsm,
2895                        struct lov_oinfo *loi, void *cookie,
2896                        int cmd, obd_off off, int count,
2897                        obd_flag brw_flags, enum async_flags async_flags)
2898 {
2899         struct client_obd *cli = &exp->exp_obd->u.cli;
2900         struct osc_async_page *oap;
2901         int rc = 0;
2902         ENTRY;
2903
2904         oap = oap_from_cookie(cookie);
2905         if (IS_ERR(oap))
2906                 RETURN(PTR_ERR(oap));
2907
2908         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2909                 RETURN(-EIO);
2910
2911         if (!cfs_list_empty(&oap->oap_pending_item) ||
2912             !cfs_list_empty(&oap->oap_urgent_item) ||
2913             !cfs_list_empty(&oap->oap_rpc_item))
2914                 RETURN(-EBUSY);
2915
2916         /* check if the file's owner/group is over quota */
2917         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2918                 struct cl_object *obj;
2919                 struct cl_attr    attr; /* XXX put attr into thread info */
2920                 unsigned int qid[MAXQUOTAS];
2921
2922                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2923
2924                 cl_object_attr_lock(obj);
2925                 rc = cl_object_attr_get(env, obj, &attr);
2926                 cl_object_attr_unlock(obj);
2927
2928                 qid[USRQUOTA] = attr.cat_uid;
2929                 qid[GRPQUOTA] = attr.cat_gid;
2930                 if (rc == 0 &&
2931                     lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2932                         rc = -EDQUOT;
2933                 if (rc)
2934                         RETURN(rc);
2935         }
2936
2937         if (loi == NULL)
2938                 loi = lsm->lsm_oinfo[0];
2939
2940         client_obd_list_lock(&cli->cl_loi_list_lock);
2941
2942         LASSERT(off + count <= CFS_PAGE_SIZE);
2943         oap->oap_cmd = cmd;
2944         oap->oap_page_off = off;
2945         oap->oap_count = count;
2946         oap->oap_brw_flags = brw_flags;
2947         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2948         if (libcfs_memory_pressure_get())
2949                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2950         cfs_spin_lock(&oap->oap_lock);
2951         oap->oap_async_flags = async_flags;
2952         cfs_spin_unlock(&oap->oap_lock);
2953
2954         if (cmd & OBD_BRW_WRITE) {
2955                 rc = osc_enter_cache(env, cli, loi, oap);
2956                 if (rc) {
2957                         client_obd_list_unlock(&cli->cl_loi_list_lock);
2958                         RETURN(rc);
2959                 }
2960         }
2961
2962         osc_oap_to_pending(oap);
2963         loi_list_maint(cli, loi);
2964
2965         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2966                   cmd);
2967
2968         osc_check_rpcs(env, cli);
2969         client_obd_list_unlock(&cli->cl_loi_list_lock);
2970
2971         RETURN(0);
2972 }
2973
2974 /* aka (~was & now & flag), but this is more clear :) */
2975 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2976
2977 int osc_set_async_flags_base(struct client_obd *cli,
2978                              struct lov_oinfo *loi, struct osc_async_page *oap,
2979                              obd_flag async_flags)
2980 {
2981         struct loi_oap_pages *lop;
2982         int flags = 0;
2983         ENTRY;
2984
2985         LASSERT(!cfs_list_empty(&oap->oap_pending_item));
2986
2987         if (oap->oap_cmd & OBD_BRW_WRITE) {
2988                 lop = &loi->loi_write_lop;
2989         } else {
2990                 lop = &loi->loi_read_lop;
2991         }
2992
2993         if ((oap->oap_async_flags & async_flags) == async_flags)
2994                 RETURN(0);
2995
2996         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2997                 flags |= ASYNC_READY;
2998
2999         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3000             cfs_list_empty(&oap->oap_rpc_item)) {
3001                 if (oap->oap_async_flags & ASYNC_HP)
3002                         cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3003                 else
3004                         cfs_list_add_tail(&oap->oap_urgent_item,
3005                                           &lop->lop_urgent);
3006                 flags |= ASYNC_URGENT;
3007                 loi_list_maint(cli, loi);
3008         }
3009         cfs_spin_lock(&oap->oap_lock);
3010         oap->oap_async_flags |= flags;
3011         cfs_spin_unlock(&oap->oap_lock);
3012
3013         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3014                         oap->oap_async_flags);
3015         RETURN(0);
3016 }
3017
3018 int osc_teardown_async_page(struct obd_export *exp,
3019                             struct lov_stripe_md *lsm,
3020                             struct lov_oinfo *loi, void *cookie)
3021 {
3022         struct client_obd *cli = &exp->exp_obd->u.cli;
3023         struct loi_oap_pages *lop;
3024         struct osc_async_page *oap;
3025         int rc = 0;
3026         ENTRY;
3027
3028         oap = oap_from_cookie(cookie);
3029         if (IS_ERR(oap))
3030                 RETURN(PTR_ERR(oap));
3031
3032         if (loi == NULL)
3033                 loi = lsm->lsm_oinfo[0];
3034
3035         if (oap->oap_cmd & OBD_BRW_WRITE) {
3036                 lop = &loi->loi_write_lop;
3037         } else {
3038                 lop = &loi->loi_read_lop;
3039         }
3040
3041         client_obd_list_lock(&cli->cl_loi_list_lock);
3042
3043         if (!cfs_list_empty(&oap->oap_rpc_item))
3044                 GOTO(out, rc = -EBUSY);
3045
3046         osc_exit_cache(cli, oap, 0);
3047         osc_wake_cache_waiters(cli);
3048
3049         if (!cfs_list_empty(&oap->oap_urgent_item)) {
3050                 cfs_list_del_init(&oap->oap_urgent_item);
3051                 cfs_spin_lock(&oap->oap_lock);
3052                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3053                 cfs_spin_unlock(&oap->oap_lock);
3054         }
3055         if (!cfs_list_empty(&oap->oap_pending_item)) {
3056                 cfs_list_del_init(&oap->oap_pending_item);
3057                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3058         }
3059         loi_list_maint(cli, loi);
3060         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3061 out:
3062         client_obd_list_unlock(&cli->cl_loi_list_lock);
3063         RETURN(rc);
3064 }
3065
3066 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3067                                          struct ldlm_enqueue_info *einfo,
3068                                          int flags)
3069 {
3070         void *data = einfo->ei_cbdata;
3071
3072         LASSERT(lock != NULL);
3073         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3074         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3075         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3076         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3077
3078         lock_res_and_lock(lock);
3079         cfs_spin_lock(&osc_ast_guard);
3080         LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3081         lock->l_ast_data = data;
3082         cfs_spin_unlock(&osc_ast_guard);
3083         unlock_res_and_lock(lock);
3084 }
3085
3086 static void osc_set_data_with_check(struct lustre_handle *lockh,
3087                                     struct ldlm_enqueue_info *einfo,
3088                                     int flags)
3089 {
3090         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3091
3092         if (lock != NULL) {
3093                 osc_set_lock_data_with_check(lock, einfo, flags);
3094                 LDLM_LOCK_PUT(lock);
3095         } else
3096                 CERROR("lockh %p, data %p - client evicted?\n",
3097                        lockh, einfo->ei_cbdata);
3098 }
3099
3100 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3101                              ldlm_iterator_t replace, void *data)
3102 {
3103         struct ldlm_res_id res_id;
3104         struct obd_device *obd = class_exp2obd(exp);
3105
3106         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3107         ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3108         return 0;
3109 }
3110
3111 /* find any ldlm lock of the inode in osc
3112  * return 0    not find
3113  *        1    find one
3114  *      < 0    error */
3115 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3116                            ldlm_iterator_t replace, void *data)
3117 {
3118         struct ldlm_res_id res_id;
3119         struct obd_device *obd = class_exp2obd(exp);
3120         int rc = 0;
3121
3122         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3123         rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3124         if (rc == LDLM_ITER_STOP)
3125                 return(1);
3126         if (rc == LDLM_ITER_CONTINUE)
3127                 return(0);
3128         return(rc);
3129 }
3130
3131 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3132                             obd_enqueue_update_f upcall, void *cookie,
3133                             int *flags, int rc)
3134 {
3135         int intent = *flags & LDLM_FL_HAS_INTENT;
3136         ENTRY;
3137
3138         if (intent) {
3139                 /* The request was created before ldlm_cli_enqueue call. */
3140                 if (rc == ELDLM_LOCK_ABORTED) {
3141                         struct ldlm_reply *rep;
3142                         rep = req_capsule_server_get(&req->rq_pill,
3143                                                      &RMF_DLM_REP);
3144
3145                         LASSERT(rep != NULL);
3146                         if (rep->lock_policy_res1)
3147                                 rc = rep->lock_policy_res1;
3148                 }
3149         }
3150
3151         if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3152                 *flags |= LDLM_FL_LVB_READY;
3153                 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3154                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3155         }
3156
3157         /* Call the update callback. */
3158         rc = (*upcall)(cookie, rc);
3159         RETURN(rc);
3160 }
3161
3162 static int osc_enqueue_interpret(const struct lu_env *env,
3163                                  struct ptlrpc_request *req,
3164                                  struct osc_enqueue_args *aa, int rc)
3165 {
3166         struct ldlm_lock *lock;
3167         struct lustre_handle handle;
3168         __u32 mode;
3169
3170         /* Make a local copy of a lock handle and a mode, because aa->oa_*
3171          * might be freed anytime after lock upcall has been called. */