Whamcloud - gitweb
LU-904 ptlrpc: redo io on -EINPROGRESS
[fs/lustre-release.git] / lustre / osc / osc_request.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  *
32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
33  */
34 /*
35  * This file is part of Lustre, http://www.lustre.org/
36  * Lustre is a trademark of Sun Microsystems, Inc.
37  */
38
39 #ifndef EXPORT_SYMTAB
40 # define EXPORT_SYMTAB
41 #endif
42 #define DEBUG_SUBSYSTEM S_OSC
43
44 #include <libcfs/libcfs.h>
45
46 #ifndef __KERNEL__
47 # include <liblustre.h>
48 #endif
49
50 #include <lustre_dlm.h>
51 #include <lustre_net.h>
52 #include <lustre/lustre_user.h>
53 #include <obd_cksum.h>
54 #include <obd_ost.h>
55 #include <obd_lov.h>
56
57 #ifdef  __CYGWIN__
58 # include <ctype.h>
59 #endif
60
61 #include <lustre_ha.h>
62 #include <lprocfs_status.h>
63 #include <lustre_log.h>
64 #include <lustre_debug.h>
65 #include <lustre_param.h>
66 #include "osc_internal.h"
67
68 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
69 static int brw_interpret(const struct lu_env *env,
70                          struct ptlrpc_request *req, void *data, int rc);
71 static void osc_check_rpcs0(const struct lu_env *env, struct client_obd *cli,
72                             int ptlrpc);
73 int osc_cleanup(struct obd_device *obd);
74
75 /* Pack OSC object metadata for disk storage (LE byte order). */
76 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
77                       struct lov_stripe_md *lsm)
78 {
79         int lmm_size;
80         ENTRY;
81
82         lmm_size = sizeof(**lmmp);
83         if (!lmmp)
84                 RETURN(lmm_size);
85
86         if (*lmmp && !lsm) {
87                 OBD_FREE(*lmmp, lmm_size);
88                 *lmmp = NULL;
89                 RETURN(0);
90         }
91
92         if (!*lmmp) {
93                 OBD_ALLOC(*lmmp, lmm_size);
94                 if (!*lmmp)
95                         RETURN(-ENOMEM);
96         }
97
98         if (lsm) {
99                 LASSERT(lsm->lsm_object_id);
100                 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
101                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
102                 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
103         }
104
105         RETURN(lmm_size);
106 }
107
108 /* Unpack OSC object metadata from disk storage (LE byte order). */
109 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
110                         struct lov_mds_md *lmm, int lmm_bytes)
111 {
112         int lsm_size;
113         struct obd_import *imp = class_exp2cliimp(exp);
114         ENTRY;
115
116         if (lmm != NULL) {
117                 if (lmm_bytes < sizeof (*lmm)) {
118                         CERROR("lov_mds_md too small: %d, need %d\n",
119                                lmm_bytes, (int)sizeof(*lmm));
120                         RETURN(-EINVAL);
121                 }
122                 /* XXX LOV_MAGIC etc check? */
123
124                 if (lmm->lmm_object_id == 0) {
125                         CERROR("lov_mds_md: zero lmm_object_id\n");
126                         RETURN(-EINVAL);
127                 }
128         }
129
130         lsm_size = lov_stripe_md_size(1);
131         if (lsmp == NULL)
132                 RETURN(lsm_size);
133
134         if (*lsmp != NULL && lmm == NULL) {
135                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
136                 OBD_FREE(*lsmp, lsm_size);
137                 *lsmp = NULL;
138                 RETURN(0);
139         }
140
141         if (*lsmp == NULL) {
142                 OBD_ALLOC(*lsmp, lsm_size);
143                 if (*lsmp == NULL)
144                         RETURN(-ENOMEM);
145                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
146                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
147                         OBD_FREE(*lsmp, lsm_size);
148                         RETURN(-ENOMEM);
149                 }
150                 loi_init((*lsmp)->lsm_oinfo[0]);
151         }
152
153         if (lmm != NULL) {
154                 /* XXX zero *lsmp? */
155                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
156                 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
157                 LASSERT((*lsmp)->lsm_object_id);
158                 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
159         }
160
161         if (imp != NULL &&
162             (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
163                 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
164         else
165                 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
166
167         RETURN(lsm_size);
168 }
169
170 static inline void osc_pack_capa(struct ptlrpc_request *req,
171                                  struct ost_body *body, void *capa)
172 {
173         struct obd_capa *oc = (struct obd_capa *)capa;
174         struct lustre_capa *c;
175
176         if (!capa)
177                 return;
178
179         c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
180         LASSERT(c);
181         capa_cpy(c, oc);
182         body->oa.o_valid |= OBD_MD_FLOSSCAPA;
183         DEBUG_CAPA(D_SEC, c, "pack");
184 }
185
186 static inline void osc_pack_req_body(struct ptlrpc_request *req,
187                                      struct obd_info *oinfo)
188 {
189         struct ost_body *body;
190
191         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
192         LASSERT(body);
193
194         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
195         osc_pack_capa(req, body, oinfo->oi_capa);
196 }
197
198 static inline void osc_set_capa_size(struct ptlrpc_request *req,
199                                      const struct req_msg_field *field,
200                                      struct obd_capa *oc)
201 {
202         if (oc == NULL)
203                 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
204         else
205                 /* it is already calculated as sizeof struct obd_capa */
206                 ;
207 }
208
209 static int osc_getattr_interpret(const struct lu_env *env,
210                                  struct ptlrpc_request *req,
211                                  struct osc_async_args *aa, int rc)
212 {
213         struct ost_body *body;
214         ENTRY;
215
216         if (rc != 0)
217                 GOTO(out, rc);
218
219         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
220         if (body) {
221                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
222                 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
223
224                 /* This should really be sent by the OST */
225                 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
226                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
227         } else {
228                 CDEBUG(D_INFO, "can't unpack ost_body\n");
229                 rc = -EPROTO;
230                 aa->aa_oi->oi_oa->o_valid = 0;
231         }
232 out:
233         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
234         RETURN(rc);
235 }
236
237 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
238                              struct ptlrpc_request_set *set)
239 {
240         struct ptlrpc_request *req;
241         struct osc_async_args *aa;
242         int                    rc;
243         ENTRY;
244
245         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
246         if (req == NULL)
247                 RETURN(-ENOMEM);
248
249         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
250         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
251         if (rc) {
252                 ptlrpc_request_free(req);
253                 RETURN(rc);
254         }
255
256         osc_pack_req_body(req, oinfo);
257
258         ptlrpc_request_set_replen(req);
259         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
260
261         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
262         aa = ptlrpc_req_async_args(req);
263         aa->aa_oi = oinfo;
264
265         ptlrpc_set_add_req(set, req);
266         RETURN(0);
267 }
268
269 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
270 {
271         struct ptlrpc_request *req;
272         struct ost_body       *body;
273         int                    rc;
274         ENTRY;
275
276         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
277         if (req == NULL)
278                 RETURN(-ENOMEM);
279
280         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
281         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
282         if (rc) {
283                 ptlrpc_request_free(req);
284                 RETURN(rc);
285         }
286
287         osc_pack_req_body(req, oinfo);
288
289         ptlrpc_request_set_replen(req);
290
291         rc = ptlrpc_queue_wait(req);
292         if (rc)
293                 GOTO(out, rc);
294
295         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
296         if (body == NULL)
297                 GOTO(out, rc = -EPROTO);
298
299         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
300         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
301
302         /* This should really be sent by the OST */
303         oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
304         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
305
306         EXIT;
307  out:
308         ptlrpc_req_finished(req);
309         return rc;
310 }
311
312 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
313                        struct obd_trans_info *oti)
314 {
315         struct ptlrpc_request *req;
316         struct ost_body       *body;
317         int                    rc;
318         ENTRY;
319
320         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
321
322         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
323         if (req == NULL)
324                 RETURN(-ENOMEM);
325
326         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
327         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
328         if (rc) {
329                 ptlrpc_request_free(req);
330                 RETURN(rc);
331         }
332
333         osc_pack_req_body(req, oinfo);
334
335         ptlrpc_request_set_replen(req);
336
337         rc = ptlrpc_queue_wait(req);
338         if (rc)
339                 GOTO(out, rc);
340
341         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
342         if (body == NULL)
343                 GOTO(out, rc = -EPROTO);
344
345         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
346
347         EXIT;
348 out:
349         ptlrpc_req_finished(req);
350         RETURN(rc);
351 }
352
353 static int osc_setattr_interpret(const struct lu_env *env,
354                                  struct ptlrpc_request *req,
355                                  struct osc_setattr_args *sa, int rc)
356 {
357         struct ost_body *body;
358         ENTRY;
359
360         if (rc != 0)
361                 GOTO(out, rc);
362
363         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
364         if (body == NULL)
365                 GOTO(out, rc = -EPROTO);
366
367         lustre_get_wire_obdo(sa->sa_oa, &body->oa);
368 out:
369         rc = sa->sa_upcall(sa->sa_cookie, rc);
370         RETURN(rc);
371 }
372
373 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
374                            struct obd_trans_info *oti,
375                            obd_enqueue_update_f upcall, void *cookie,
376                            struct ptlrpc_request_set *rqset)
377 {
378         struct ptlrpc_request   *req;
379         struct osc_setattr_args *sa;
380         int                      rc;
381         ENTRY;
382
383         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
384         if (req == NULL)
385                 RETURN(-ENOMEM);
386
387         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
388         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
389         if (rc) {
390                 ptlrpc_request_free(req);
391                 RETURN(rc);
392         }
393
394         if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
395                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
396
397         osc_pack_req_body(req, oinfo);
398
399         ptlrpc_request_set_replen(req);
400
401         /* do mds to ost setattr asynchronously */
402         if (!rqset) {
403                 /* Do not wait for response. */
404                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
405         } else {
406                 req->rq_interpret_reply =
407                         (ptlrpc_interpterer_t)osc_setattr_interpret;
408
409                 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
410                 sa = ptlrpc_req_async_args(req);
411                 sa->sa_oa = oinfo->oi_oa;
412                 sa->sa_upcall = upcall;
413                 sa->sa_cookie = cookie;
414
415                 if (rqset == PTLRPCD_SET)
416                         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
417                 else
418                         ptlrpc_set_add_req(rqset, req);
419         }
420
421         RETURN(0);
422 }
423
424 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
425                              struct obd_trans_info *oti,
426                              struct ptlrpc_request_set *rqset)
427 {
428         return osc_setattr_async_base(exp, oinfo, oti,
429                                       oinfo->oi_cb_up, oinfo, rqset);
430 }
431
432 int osc_real_create(struct obd_export *exp, struct obdo *oa,
433                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
434 {
435         struct ptlrpc_request *req;
436         struct ost_body       *body;
437         struct lov_stripe_md  *lsm;
438         int                    rc;
439         ENTRY;
440
441         LASSERT(oa);
442         LASSERT(ea);
443
444         lsm = *ea;
445         if (!lsm) {
446                 rc = obd_alloc_memmd(exp, &lsm);
447                 if (rc < 0)
448                         RETURN(rc);
449         }
450
451         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
452         if (req == NULL)
453                 GOTO(out, rc = -ENOMEM);
454
455         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
456         if (rc) {
457                 ptlrpc_request_free(req);
458                 GOTO(out, rc);
459         }
460
461         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
462         LASSERT(body);
463         lustre_set_wire_obdo(&body->oa, oa);
464
465         ptlrpc_request_set_replen(req);
466
467         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
468             oa->o_flags == OBD_FL_DELORPHAN) {
469                 DEBUG_REQ(D_HA, req,
470                           "delorphan from OST integration");
471                 /* Don't resend the delorphan req */
472                 req->rq_no_resend = req->rq_no_delay = 1;
473         }
474
475         rc = ptlrpc_queue_wait(req);
476         if (rc)
477                 GOTO(out_req, rc);
478
479         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
480         if (body == NULL)
481                 GOTO(out_req, rc = -EPROTO);
482
483         lustre_get_wire_obdo(oa, &body->oa);
484
485         /* This should really be sent by the OST */
486         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
487         oa->o_valid |= OBD_MD_FLBLKSZ;
488
489         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
490          * have valid lsm_oinfo data structs, so don't go touching that.
491          * This needs to be fixed in a big way.
492          */
493         lsm->lsm_object_id = oa->o_id;
494         lsm->lsm_object_seq = oa->o_seq;
495         *ea = lsm;
496
497         if (oti != NULL) {
498                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
499
500                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
501                         if (!oti->oti_logcookies)
502                                 oti_alloc_cookies(oti, 1);
503                         *oti->oti_logcookies = oa->o_lcookie;
504                 }
505         }
506
507         CDEBUG(D_HA, "transno: "LPD64"\n",
508                lustre_msg_get_transno(req->rq_repmsg));
509 out_req:
510         ptlrpc_req_finished(req);
511 out:
512         if (rc && !*ea)
513                 obd_free_memmd(exp, &lsm);
514         RETURN(rc);
515 }
516
517 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
518                    obd_enqueue_update_f upcall, void *cookie,
519                    struct ptlrpc_request_set *rqset)
520 {
521         struct ptlrpc_request   *req;
522         struct osc_setattr_args *sa;
523         struct ost_body         *body;
524         int                      rc;
525         ENTRY;
526
527         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
528         if (req == NULL)
529                 RETURN(-ENOMEM);
530
531         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
532         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
533         if (rc) {
534                 ptlrpc_request_free(req);
535                 RETURN(rc);
536         }
537         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
538         ptlrpc_at_set_req_timeout(req);
539
540         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
541         LASSERT(body);
542         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
543         osc_pack_capa(req, body, oinfo->oi_capa);
544
545         ptlrpc_request_set_replen(req);
546
547
548         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
549         CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
550         sa = ptlrpc_req_async_args(req);
551         sa->sa_oa     = oinfo->oi_oa;
552         sa->sa_upcall = upcall;
553         sa->sa_cookie = cookie;
554         if (rqset == PTLRPCD_SET)
555                 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
556         else
557                 ptlrpc_set_add_req(rqset, req);
558
559         RETURN(0);
560 }
561
562 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
563                      struct obd_trans_info *oti,
564                      struct ptlrpc_request_set *rqset)
565 {
566         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
567         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
568         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
569         return osc_punch_base(exp, oinfo,
570                               oinfo->oi_cb_up, oinfo, rqset);
571 }
572
573 static int osc_sync_interpret(const struct lu_env *env,
574                               struct ptlrpc_request *req,
575                               void *arg, int rc)
576 {
577         struct osc_async_args *aa = arg;
578         struct ost_body *body;
579         ENTRY;
580
581         if (rc)
582                 GOTO(out, rc);
583
584         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
585         if (body == NULL) {
586                 CERROR ("can't unpack ost_body\n");
587                 GOTO(out, rc = -EPROTO);
588         }
589
590         *aa->aa_oi->oi_oa = body->oa;
591 out:
592         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
593         RETURN(rc);
594 }
595
596 static int osc_sync(struct obd_export *exp, struct obd_info *oinfo,
597                     obd_size start, obd_size end,
598                     struct ptlrpc_request_set *set)
599 {
600         struct ptlrpc_request *req;
601         struct ost_body       *body;
602         struct osc_async_args *aa;
603         int                    rc;
604         ENTRY;
605
606         if (!oinfo->oi_oa) {
607                 CDEBUG(D_INFO, "oa NULL\n");
608                 RETURN(-EINVAL);
609         }
610
611         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
612         if (req == NULL)
613                 RETURN(-ENOMEM);
614
615         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
616         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
617         if (rc) {
618                 ptlrpc_request_free(req);
619                 RETURN(rc);
620         }
621
622         /* overload the size and blocks fields in the oa with start/end */
623         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
624         LASSERT(body);
625         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
626         body->oa.o_size = start;
627         body->oa.o_blocks = end;
628         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
629         osc_pack_capa(req, body, oinfo->oi_capa);
630
631         ptlrpc_request_set_replen(req);
632         req->rq_interpret_reply = osc_sync_interpret;
633
634         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
635         aa = ptlrpc_req_async_args(req);
636         aa->aa_oi = oinfo;
637
638         ptlrpc_set_add_req(set, req);
639         RETURN (0);
640 }
641
642 /* Find and cancel locally locks matched by @mode in the resource found by
643  * @objid. Found locks are added into @cancel list. Returns the amount of
644  * locks added to @cancels list. */
645 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
646                                    cfs_list_t *cancels,
647                                    ldlm_mode_t mode, int lock_flags)
648 {
649         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
650         struct ldlm_res_id res_id;
651         struct ldlm_resource *res;
652         int count;
653         ENTRY;
654
655         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
656         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
657         if (res == NULL)
658                 RETURN(0);
659
660         LDLM_RESOURCE_ADDREF(res);
661         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
662                                            lock_flags, 0, NULL);
663         LDLM_RESOURCE_DELREF(res);
664         ldlm_resource_putref(res);
665         RETURN(count);
666 }
667
668 static int osc_destroy_interpret(const struct lu_env *env,
669                                  struct ptlrpc_request *req, void *data,
670                                  int rc)
671 {
672         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
673
674         cfs_atomic_dec(&cli->cl_destroy_in_flight);
675         cfs_waitq_signal(&cli->cl_destroy_waitq);
676         return 0;
677 }
678
679 static int osc_can_send_destroy(struct client_obd *cli)
680 {
681         if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
682             cli->cl_max_rpcs_in_flight) {
683                 /* The destroy request can be sent */
684                 return 1;
685         }
686         if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
687             cli->cl_max_rpcs_in_flight) {
688                 /*
689                  * The counter has been modified between the two atomic
690                  * operations.
691                  */
692                 cfs_waitq_signal(&cli->cl_destroy_waitq);
693         }
694         return 0;
695 }
696
697 /* Destroy requests can be async always on the client, and we don't even really
698  * care about the return code since the client cannot do anything at all about
699  * a destroy failure.
700  * When the MDS is unlinking a filename, it saves the file objects into a
701  * recovery llog, and these object records are cancelled when the OST reports
702  * they were destroyed and sync'd to disk (i.e. transaction committed).
703  * If the client dies, or the OST is down when the object should be destroyed,
704  * the records are not cancelled, and when the OST reconnects to the MDS next,
705  * it will retrieve the llog unlink logs and then sends the log cancellation
706  * cookies to the MDS after committing destroy transactions. */
707 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
708                        struct lov_stripe_md *ea, struct obd_trans_info *oti,
709                        struct obd_export *md_export, void *capa)
710 {
711         struct client_obd     *cli = &exp->exp_obd->u.cli;
712         struct ptlrpc_request *req;
713         struct ost_body       *body;
714         CFS_LIST_HEAD(cancels);
715         int rc, count;
716         ENTRY;
717
718         if (!oa) {
719                 CDEBUG(D_INFO, "oa NULL\n");
720                 RETURN(-EINVAL);
721         }
722
723         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
724                                         LDLM_FL_DISCARD_DATA);
725
726         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
727         if (req == NULL) {
728                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
729                 RETURN(-ENOMEM);
730         }
731
732         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
733         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
734                                0, &cancels, count);
735         if (rc) {
736                 ptlrpc_request_free(req);
737                 RETURN(rc);
738         }
739
740         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
741         ptlrpc_at_set_req_timeout(req);
742
743         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
744                 oa->o_lcookie = *oti->oti_logcookies;
745         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
746         LASSERT(body);
747         lustre_set_wire_obdo(&body->oa, oa);
748
749         osc_pack_capa(req, body, (struct obd_capa *)capa);
750         ptlrpc_request_set_replen(req);
751
752         /* don't throttle destroy RPCs for the MDT */
753         if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
754                 req->rq_interpret_reply = osc_destroy_interpret;
755                 if (!osc_can_send_destroy(cli)) {
756                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
757                                                           NULL);
758
759                         /*
760                          * Wait until the number of on-going destroy RPCs drops
761                          * under max_rpc_in_flight
762                          */
763                         l_wait_event_exclusive(cli->cl_destroy_waitq,
764                                                osc_can_send_destroy(cli), &lwi);
765                 }
766         }
767
768         /* Do not wait for response */
769         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
770         RETURN(0);
771 }
772
773 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
774                                 long writing_bytes)
775 {
776         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
777
778         LASSERT(!(oa->o_valid & bits));
779
780         oa->o_valid |= bits;
781         client_obd_list_lock(&cli->cl_loi_list_lock);
782         oa->o_dirty = cli->cl_dirty;
783         if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
784                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
785                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
786                 oa->o_undirty = 0;
787         } else if (cfs_atomic_read(&obd_dirty_pages) -
788                    cfs_atomic_read(&obd_dirty_transit_pages) >
789                    obd_max_dirty_pages + 1){
790                 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
791                  * not covered by a lock thus they may safely race and trip
792                  * this CERROR() unless we add in a small fudge factor (+1). */
793                 CERROR("dirty %d - %d > system dirty_max %d\n",
794                        cfs_atomic_read(&obd_dirty_pages),
795                        cfs_atomic_read(&obd_dirty_transit_pages),
796                        obd_max_dirty_pages);
797                 oa->o_undirty = 0;
798         } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
799                 CERROR("dirty %lu - dirty_max %lu too big???\n",
800                        cli->cl_dirty, cli->cl_dirty_max);
801                 oa->o_undirty = 0;
802         } else {
803                 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
804                                 (cli->cl_max_rpcs_in_flight + 1);
805                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
806         }
807         oa->o_grant = cli->cl_avail_grant;
808         oa->o_dropped = cli->cl_lost_grant;
809         cli->cl_lost_grant = 0;
810         client_obd_list_unlock(&cli->cl_loi_list_lock);
811         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
812                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
813
814 }
815
816 static void osc_update_next_shrink(struct client_obd *cli)
817 {
818         cli->cl_next_shrink_grant =
819                 cfs_time_shift(cli->cl_grant_shrink_interval);
820         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
821                cli->cl_next_shrink_grant);
822 }
823
824 /* caller must hold loi_list_lock */
825 static void osc_consume_write_grant(struct client_obd *cli,
826                                     struct brw_page *pga)
827 {
828         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
829         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
830         cfs_atomic_inc(&obd_dirty_pages);
831         cli->cl_dirty += CFS_PAGE_SIZE;
832         cli->cl_avail_grant -= CFS_PAGE_SIZE;
833         pga->flag |= OBD_BRW_FROM_GRANT;
834         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
835                CFS_PAGE_SIZE, pga, pga->pg);
836         LASSERT(cli->cl_avail_grant >= 0);
837         osc_update_next_shrink(cli);
838 }
839
840 /* the companion to osc_consume_write_grant, called when a brw has completed.
841  * must be called with the loi lock held. */
842 static void osc_release_write_grant(struct client_obd *cli,
843                                     struct brw_page *pga, int sent)
844 {
845         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
846         ENTRY;
847
848         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
849         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
850                 EXIT;
851                 return;
852         }
853
854         pga->flag &= ~OBD_BRW_FROM_GRANT;
855         cfs_atomic_dec(&obd_dirty_pages);
856         cli->cl_dirty -= CFS_PAGE_SIZE;
857         if (pga->flag & OBD_BRW_NOCACHE) {
858                 pga->flag &= ~OBD_BRW_NOCACHE;
859                 cfs_atomic_dec(&obd_dirty_transit_pages);
860                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
861         }
862         if (!sent) {
863                 /* Reclaim grant from truncated pages. This is used to solve
864                  * write-truncate and grant all gone(to lost_grant) problem.
865                  * For a vfs write this problem can be easily solved by a sync
866                  * write, however, this is not an option for page_mkwrite()
867                  * because grant has to be allocated before a page becomes
868                  * dirty. */
869                 if (cli->cl_avail_grant < PTLRPC_MAX_BRW_SIZE)
870                         cli->cl_avail_grant += CFS_PAGE_SIZE;
871                 else
872                         cli->cl_lost_grant += CFS_PAGE_SIZE;
873                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
874                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
875         } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
876                 /* For short writes we shouldn't count parts of pages that
877                  * span a whole block on the OST side, or our accounting goes
878                  * wrong.  Should match the code in filter_grant_check. */
879                 int offset = pga->off & ~CFS_PAGE_MASK;
880                 int count = pga->count + (offset & (blocksize - 1));
881                 int end = (offset + pga->count) & (blocksize - 1);
882                 if (end)
883                         count += blocksize - end;
884
885                 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
886                 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
887                        CFS_PAGE_SIZE - count, cli->cl_lost_grant,
888                        cli->cl_avail_grant, cli->cl_dirty);
889         }
890
891         EXIT;
892 }
893
894 static unsigned long rpcs_in_flight(struct client_obd *cli)
895 {
896         return cli->cl_r_in_flight + cli->cl_w_in_flight;
897 }
898
899 /* caller must hold loi_list_lock */
900 void osc_wake_cache_waiters(struct client_obd *cli)
901 {
902         cfs_list_t *l, *tmp;
903         struct osc_cache_waiter *ocw;
904
905         ENTRY;
906         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
907                 /* if we can't dirty more, we must wait until some is written */
908                 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
909                    (cfs_atomic_read(&obd_dirty_pages) + 1 >
910                     obd_max_dirty_pages)) {
911                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
912                                "osc max %ld, sys max %d\n", cli->cl_dirty,
913                                cli->cl_dirty_max, obd_max_dirty_pages);
914                         return;
915                 }
916
917                 /* if still dirty cache but no grant wait for pending RPCs that
918                  * may yet return us some grant before doing sync writes */
919                 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
920                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
921                                cli->cl_w_in_flight);
922                         return;
923                 }
924
925                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
926                 cfs_list_del_init(&ocw->ocw_entry);
927                 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
928                         /* no more RPCs in flight to return grant, do sync IO */
929                         ocw->ocw_rc = -EDQUOT;
930                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
931                 } else {
932                         osc_consume_write_grant(cli,
933                                                 &ocw->ocw_oap->oap_brw_page);
934                 }
935
936                 CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld\n",
937                        ocw, ocw->ocw_oap, cli->cl_avail_grant);
938
939                 cfs_waitq_signal(&ocw->ocw_waitq);
940         }
941
942         EXIT;
943 }
944
945 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
946 {
947         client_obd_list_lock(&cli->cl_loi_list_lock);
948         cli->cl_avail_grant += grant;
949         client_obd_list_unlock(&cli->cl_loi_list_lock);
950 }
951
952 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
953 {
954         if (body->oa.o_valid & OBD_MD_FLGRANT) {
955                 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
956                 __osc_update_grant(cli, body->oa.o_grant);
957         }
958 }
959
960 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
961                               void *key, obd_count vallen, void *val,
962                               struct ptlrpc_request_set *set);
963
964 static int osc_shrink_grant_interpret(const struct lu_env *env,
965                                       struct ptlrpc_request *req,
966                                       void *aa, int rc)
967 {
968         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
969         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
970         struct ost_body *body;
971
972         if (rc != 0) {
973                 __osc_update_grant(cli, oa->o_grant);
974                 GOTO(out, rc);
975         }
976
977         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
978         LASSERT(body);
979         osc_update_grant(cli, body);
980 out:
981         OBDO_FREE(oa);
982         return rc;
983 }
984
985 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
986 {
987         client_obd_list_lock(&cli->cl_loi_list_lock);
988         oa->o_grant = cli->cl_avail_grant / 4;
989         cli->cl_avail_grant -= oa->o_grant;
990         client_obd_list_unlock(&cli->cl_loi_list_lock);
991         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
992                 oa->o_valid |= OBD_MD_FLFLAGS;
993                 oa->o_flags = 0;
994         }
995         oa->o_flags |= OBD_FL_SHRINK_GRANT;
996         osc_update_next_shrink(cli);
997 }
998
999 /* Shrink the current grant, either from some large amount to enough for a
1000  * full set of in-flight RPCs, or if we have already shrunk to that limit
1001  * then to enough for a single RPC.  This avoids keeping more grant than
1002  * needed, and avoids shrinking the grant piecemeal. */
1003 static int osc_shrink_grant(struct client_obd *cli)
1004 {
1005         long target = (cli->cl_max_rpcs_in_flight + 1) *
1006                       cli->cl_max_pages_per_rpc;
1007
1008         client_obd_list_lock(&cli->cl_loi_list_lock);
1009         if (cli->cl_avail_grant <= target)
1010                 target = cli->cl_max_pages_per_rpc;
1011         client_obd_list_unlock(&cli->cl_loi_list_lock);
1012
1013         return osc_shrink_grant_to_target(cli, target);
1014 }
1015
1016 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
1017 {
1018         int    rc = 0;
1019         struct ost_body     *body;
1020         ENTRY;
1021
1022         client_obd_list_lock(&cli->cl_loi_list_lock);
1023         /* Don't shrink if we are already above or below the desired limit
1024          * We don't want to shrink below a single RPC, as that will negatively
1025          * impact block allocation and long-term performance. */
1026         if (target < cli->cl_max_pages_per_rpc)
1027                 target = cli->cl_max_pages_per_rpc;
1028
1029         if (target >= cli->cl_avail_grant) {
1030                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1031                 RETURN(0);
1032         }
1033         client_obd_list_unlock(&cli->cl_loi_list_lock);
1034
1035         OBD_ALLOC_PTR(body);
1036         if (!body)
1037                 RETURN(-ENOMEM);
1038
1039         osc_announce_cached(cli, &body->oa, 0);
1040
1041         client_obd_list_lock(&cli->cl_loi_list_lock);
1042         body->oa.o_grant = cli->cl_avail_grant - target;
1043         cli->cl_avail_grant = target;
1044         client_obd_list_unlock(&cli->cl_loi_list_lock);
1045         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
1046                 body->oa.o_valid |= OBD_MD_FLFLAGS;
1047                 body->oa.o_flags = 0;
1048         }
1049         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1050         osc_update_next_shrink(cli);
1051
1052         rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1053                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1054                                 sizeof(*body), body, NULL);
1055         if (rc != 0)
1056                 __osc_update_grant(cli, body->oa.o_grant);
1057         OBD_FREE_PTR(body);
1058         RETURN(rc);
1059 }
1060
1061 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1062 static int osc_should_shrink_grant(struct client_obd *client)
1063 {
1064         cfs_time_t time = cfs_time_current();
1065         cfs_time_t next_shrink = client->cl_next_shrink_grant;
1066
1067         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1068              OBD_CONNECT_GRANT_SHRINK) == 0)
1069                 return 0;
1070
1071         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1072                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1073                     client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1074                         return 1;
1075                 else
1076                         osc_update_next_shrink(client);
1077         }
1078         return 0;
1079 }
1080
1081 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1082 {
1083         struct client_obd *client;
1084
1085         cfs_list_for_each_entry(client, &item->ti_obd_list,
1086                                 cl_grant_shrink_list) {
1087                 if (osc_should_shrink_grant(client))
1088                         osc_shrink_grant(client);
1089         }
1090         return 0;
1091 }
1092
1093 static int osc_add_shrink_grant(struct client_obd *client)
1094 {
1095         int rc;
1096
1097         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1098                                        TIMEOUT_GRANT,
1099                                        osc_grant_shrink_grant_cb, NULL,
1100                                        &client->cl_grant_shrink_list);
1101         if (rc) {
1102                 CERROR("add grant client %s error %d\n",
1103                         client->cl_import->imp_obd->obd_name, rc);
1104                 return rc;
1105         }
1106         CDEBUG(D_CACHE, "add grant client %s \n",
1107                client->cl_import->imp_obd->obd_name);
1108         osc_update_next_shrink(client);
1109         return 0;
1110 }
1111
1112 static int osc_del_shrink_grant(struct client_obd *client)
1113 {
1114         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1115                                          TIMEOUT_GRANT);
1116 }
1117
1118 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1119 {
1120         /*
1121          * ocd_grant is the total grant amount we're expect to hold: if we've
1122          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1123          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1124          *
1125          * race is tolerable here: if we're evicted, but imp_state already
1126          * left EVICTED state, then cl_dirty must be 0 already.
1127          */
1128         client_obd_list_lock(&cli->cl_loi_list_lock);
1129         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1130                 cli->cl_avail_grant = ocd->ocd_grant;
1131         else
1132                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1133
1134         if (cli->cl_avail_grant < 0) {
1135                 CWARN("%s: available grant < 0, the OSS is probably not running"
1136                       " with patch from bug20278 (%ld) \n",
1137                       cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1138                 /* workaround for 1.6 servers which do not have
1139                  * the patch from bug20278 */
1140                 cli->cl_avail_grant = ocd->ocd_grant;
1141         }
1142
1143         client_obd_list_unlock(&cli->cl_loi_list_lock);
1144
1145         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1146                cli->cl_import->imp_obd->obd_name,
1147                cli->cl_avail_grant, cli->cl_lost_grant);
1148
1149         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1150             cfs_list_empty(&cli->cl_grant_shrink_list))
1151                 osc_add_shrink_grant(cli);
1152 }
1153
1154 /* We assume that the reason this OSC got a short read is because it read
1155  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1156  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1157  * this stripe never got written at or beyond this stripe offset yet. */
1158 static void handle_short_read(int nob_read, obd_count page_count,
1159                               struct brw_page **pga)
1160 {
1161         char *ptr;
1162         int i = 0;
1163
1164         /* skip bytes read OK */
1165         while (nob_read > 0) {
1166                 LASSERT (page_count > 0);
1167
1168                 if (pga[i]->count > nob_read) {
1169                         /* EOF inside this page */
1170                         ptr = cfs_kmap(pga[i]->pg) +
1171                                 (pga[i]->off & ~CFS_PAGE_MASK);
1172                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1173                         cfs_kunmap(pga[i]->pg);
1174                         page_count--;
1175                         i++;
1176                         break;
1177                 }
1178
1179                 nob_read -= pga[i]->count;
1180                 page_count--;
1181                 i++;
1182         }
1183
1184         /* zero remaining pages */
1185         while (page_count-- > 0) {
1186                 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1187                 memset(ptr, 0, pga[i]->count);
1188                 cfs_kunmap(pga[i]->pg);
1189                 i++;
1190         }
1191 }
1192
1193 static int check_write_rcs(struct ptlrpc_request *req,
1194                            int requested_nob, int niocount,
1195                            obd_count page_count, struct brw_page **pga)
1196 {
1197         int     i;
1198         __u32   *remote_rcs;
1199
1200         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1201                                                   sizeof(*remote_rcs) *
1202                                                   niocount);
1203         if (remote_rcs == NULL) {
1204                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1205                 return(-EPROTO);
1206         }
1207
1208         /* return error if any niobuf was in error */
1209         for (i = 0; i < niocount; i++) {
1210                 if ((int)remote_rcs[i] < 0)
1211                         return(remote_rcs[i]);
1212
1213                 if (remote_rcs[i] != 0) {
1214                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1215                                 i, remote_rcs[i], req);
1216                         return(-EPROTO);
1217                 }
1218         }
1219
1220         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1221                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1222                        req->rq_bulk->bd_nob_transferred, requested_nob);
1223                 return(-EPROTO);
1224         }
1225
1226         return (0);
1227 }
1228
1229 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1230 {
1231         if (p1->flag != p2->flag) {
1232                 unsigned mask = ~(OBD_BRW_FROM_GRANT| OBD_BRW_NOCACHE|
1233                                   OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1234
1235                 /* warn if we try to combine flags that we don't know to be
1236                  * safe to combine */
1237                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1238                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1239                               "report this at http://bugs.whamcloud.com/\n",
1240                               p1->flag, p2->flag);
1241                 }
1242                 return 0;
1243         }
1244
1245         return (p1->off + p1->count == p2->off);
1246 }
1247
1248 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1249                                    struct brw_page **pga, int opc,
1250                                    cksum_type_t cksum_type)
1251 {
1252         __u32 cksum;
1253         int i = 0;
1254
1255         LASSERT (pg_count > 0);
1256         cksum = init_checksum(cksum_type);
1257         while (nob > 0 && pg_count > 0) {
1258                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1259                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1260                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1261
1262                 /* corrupt the data before we compute the checksum, to
1263                  * simulate an OST->client data error */
1264                 if (i == 0 && opc == OST_READ &&
1265                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1266                         memcpy(ptr + off, "bad1", min(4, nob));
1267                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1268                 cfs_kunmap(pga[i]->pg);
1269                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1270                                off, cksum);
1271
1272                 nob -= pga[i]->count;
1273                 pg_count--;
1274                 i++;
1275         }
1276         /* For sending we only compute the wrong checksum instead
1277          * of corrupting the data so it is still correct on a redo */
1278         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1279                 cksum++;
1280
1281         return fini_checksum(cksum, cksum_type);
1282 }
1283
1284 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1285                                 struct lov_stripe_md *lsm, obd_count page_count,
1286                                 struct brw_page **pga,
1287                                 struct ptlrpc_request **reqp,
1288                                 struct obd_capa *ocapa, int reserve,
1289                                 int resend)
1290 {
1291         struct ptlrpc_request   *req;
1292         struct ptlrpc_bulk_desc *desc;
1293         struct ost_body         *body;
1294         struct obd_ioobj        *ioobj;
1295         struct niobuf_remote    *niobuf;
1296         int niocount, i, requested_nob, opc, rc;
1297         struct osc_brw_async_args *aa;
1298         struct req_capsule      *pill;
1299         struct brw_page *pg_prev;
1300
1301         ENTRY;
1302         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1303                 RETURN(-ENOMEM); /* Recoverable */
1304         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1305                 RETURN(-EINVAL); /* Fatal */
1306
1307         if ((cmd & OBD_BRW_WRITE) != 0) {
1308                 opc = OST_WRITE;
1309                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1310                                                 cli->cl_import->imp_rq_pool,
1311                                                 &RQF_OST_BRW_WRITE);
1312         } else {
1313                 opc = OST_READ;
1314                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1315         }
1316         if (req == NULL)
1317                 RETURN(-ENOMEM);
1318
1319         for (niocount = i = 1; i < page_count; i++) {
1320                 if (!can_merge_pages(pga[i - 1], pga[i]))
1321                         niocount++;
1322         }
1323
1324         pill = &req->rq_pill;
1325         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1326                              sizeof(*ioobj));
1327         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1328                              niocount * sizeof(*niobuf));
1329         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1330
1331         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1332         if (rc) {
1333                 ptlrpc_request_free(req);
1334                 RETURN(rc);
1335         }
1336         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1337         ptlrpc_at_set_req_timeout(req);
1338
1339         if (opc == OST_WRITE)
1340                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1341                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1342         else
1343                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1344                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1345
1346         if (desc == NULL)
1347                 GOTO(out, rc = -ENOMEM);
1348         /* NB request now owns desc and will free it when it gets freed */
1349
1350         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1351         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1352         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1353         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1354
1355         lustre_set_wire_obdo(&body->oa, oa);
1356
1357         obdo_to_ioobj(oa, ioobj);
1358         ioobj->ioo_bufcnt = niocount;
1359         osc_pack_capa(req, body, ocapa);
1360         LASSERT (page_count > 0);
1361         pg_prev = pga[0];
1362         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1363                 struct brw_page *pg = pga[i];
1364                 int poff = pg->off & ~CFS_PAGE_MASK;
1365
1366                 LASSERT(pg->count > 0);
1367                 /* make sure there is no gap in the middle of page array */
1368                 LASSERTF(page_count == 1 ||
1369                          (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1370                           ergo(i > 0 && i < page_count - 1,
1371                                poff == 0 && pg->count == CFS_PAGE_SIZE)   &&
1372                           ergo(i == page_count - 1, poff == 0)),
1373                          "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1374                          i, page_count, pg, pg->off, pg->count);
1375 #ifdef __linux__
1376                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1377                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1378                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1379                          i, page_count,
1380                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1381                          pg_prev->pg, page_private(pg_prev->pg),
1382                          pg_prev->pg->index, pg_prev->off);
1383 #else
1384                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1385                          "i %d p_c %u\n", i, page_count);
1386 #endif
1387                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1388                         (pg->flag & OBD_BRW_SRVLOCK));
1389
1390                 ptlrpc_prep_bulk_page(desc, pg->pg, poff, pg->count);
1391                 requested_nob += pg->count;
1392
1393                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1394                         niobuf--;
1395                         niobuf->len += pg->count;
1396                 } else {
1397                         niobuf->offset = pg->off;
1398                         niobuf->len    = pg->count;
1399                         niobuf->flags  = pg->flag;
1400                 }
1401                 pg_prev = pg;
1402         }
1403
1404         LASSERTF((void *)(niobuf - niocount) ==
1405                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1406                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1407                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1408
1409         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1410         if (resend) {
1411                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1412                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1413                         body->oa.o_flags = 0;
1414                 }
1415                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1416         }
1417
1418         if (osc_should_shrink_grant(cli))
1419                 osc_shrink_grant_local(cli, &body->oa);
1420
1421         /* size[REQ_REC_OFF] still sizeof (*body) */
1422         if (opc == OST_WRITE) {
1423                 if (cli->cl_checksum &&
1424                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1425                         /* store cl_cksum_type in a local variable since
1426                          * it can be changed via lprocfs */
1427                         cksum_type_t cksum_type = cli->cl_cksum_type;
1428
1429                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1430                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1431                                 body->oa.o_flags = 0;
1432                         }
1433                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1434                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1435                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1436                                                              page_count, pga,
1437                                                              OST_WRITE,
1438                                                              cksum_type);
1439                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1440                                body->oa.o_cksum);
1441                         /* save this in 'oa', too, for later checking */
1442                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1443                         oa->o_flags |= cksum_type_pack(cksum_type);
1444                 } else {
1445                         /* clear out the checksum flag, in case this is a
1446                          * resend but cl_checksum is no longer set. b=11238 */
1447                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1448                 }
1449                 oa->o_cksum = body->oa.o_cksum;
1450                 /* 1 RC per niobuf */
1451                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1452                                      sizeof(__u32) * niocount);
1453         } else {
1454                 if (cli->cl_checksum &&
1455                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1456                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1457                                 body->oa.o_flags = 0;
1458                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1459                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1460                 }
1461         }
1462         ptlrpc_request_set_replen(req);
1463
1464         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1465         aa = ptlrpc_req_async_args(req);
1466         aa->aa_oa = oa;
1467         aa->aa_requested_nob = requested_nob;
1468         aa->aa_nio_count = niocount;
1469         aa->aa_page_count = page_count;
1470         aa->aa_resends = 0;
1471         aa->aa_ppga = pga;
1472         aa->aa_cli = cli;
1473         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1474         if (ocapa && reserve)
1475                 aa->aa_ocapa = capa_get(ocapa);
1476
1477         *reqp = req;
1478         RETURN(0);
1479
1480  out:
1481         ptlrpc_req_finished(req);
1482         RETURN(rc);
1483 }
1484
1485 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1486                                 __u32 client_cksum, __u32 server_cksum, int nob,
1487                                 obd_count page_count, struct brw_page **pga,
1488                                 cksum_type_t client_cksum_type)
1489 {
1490         __u32 new_cksum;
1491         char *msg;
1492         cksum_type_t cksum_type;
1493
1494         if (server_cksum == client_cksum) {
1495                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1496                 return 0;
1497         }
1498
1499         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1500                                        oa->o_flags : 0);
1501         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1502                                       cksum_type);
1503
1504         if (cksum_type != client_cksum_type)
1505                 msg = "the server did not use the checksum type specified in "
1506                       "the original request - likely a protocol problem";
1507         else if (new_cksum == server_cksum)
1508                 msg = "changed on the client after we checksummed it - "
1509                       "likely false positive due to mmap IO (bug 11742)";
1510         else if (new_cksum == client_cksum)
1511                 msg = "changed in transit before arrival at OST";
1512         else
1513                 msg = "changed in transit AND doesn't match the original - "
1514                       "likely false positive due to mmap IO (bug 11742)";
1515
1516         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1517                            " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1518                            msg, libcfs_nid2str(peer->nid),
1519                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1520                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1521                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1522                            oa->o_id,
1523                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1524                            pga[0]->off,
1525                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1526         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1527                "client csum now %x\n", client_cksum, client_cksum_type,
1528                server_cksum, cksum_type, new_cksum);
1529         return 1;
1530 }
1531
1532 /* Note rc enters this function as number of bytes transferred */
1533 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1534 {
1535         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1536         const lnet_process_id_t *peer =
1537                         &req->rq_import->imp_connection->c_peer;
1538         struct client_obd *cli = aa->aa_cli;
1539         struct ost_body *body;
1540         __u32 client_cksum = 0;
1541         ENTRY;
1542
1543         if (rc < 0 && rc != -EDQUOT) {
1544                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1545                 RETURN(rc);
1546         }
1547
1548         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1549         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1550         if (body == NULL) {
1551                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1552                 RETURN(-EPROTO);
1553         }
1554
1555         /* set/clear over quota flag for a uid/gid */
1556         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1557             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1558                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1559
1560                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1561                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1562                        body->oa.o_flags);
1563                 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1564         }
1565
1566         osc_update_grant(cli, body);
1567
1568         if (rc < 0)
1569                 RETURN(rc);
1570
1571         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1572                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1573
1574         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1575                 if (rc > 0) {
1576                         CERROR("Unexpected +ve rc %d\n", rc);
1577                         RETURN(-EPROTO);
1578                 }
1579                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1580
1581                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1582                         RETURN(-EAGAIN);
1583
1584                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1585                     check_write_checksum(&body->oa, peer, client_cksum,
1586                                          body->oa.o_cksum, aa->aa_requested_nob,
1587                                          aa->aa_page_count, aa->aa_ppga,
1588                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1589                         RETURN(-EAGAIN);
1590
1591                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1592                                      aa->aa_page_count, aa->aa_ppga);
1593                 GOTO(out, rc);
1594         }
1595
1596         /* The rest of this function executes only for OST_READs */
1597
1598         /* if unwrap_bulk failed, return -EAGAIN to retry */
1599         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1600         if (rc < 0)
1601                 GOTO(out, rc = -EAGAIN);
1602
1603         if (rc > aa->aa_requested_nob) {
1604                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1605                        aa->aa_requested_nob);
1606                 RETURN(-EPROTO);
1607         }
1608
1609         if (rc != req->rq_bulk->bd_nob_transferred) {
1610                 CERROR ("Unexpected rc %d (%d transferred)\n",
1611                         rc, req->rq_bulk->bd_nob_transferred);
1612                 return (-EPROTO);
1613         }
1614
1615         if (rc < aa->aa_requested_nob)
1616                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1617
1618         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1619                 static int cksum_counter;
1620                 __u32      server_cksum = body->oa.o_cksum;
1621                 char      *via;
1622                 char      *router;
1623                 cksum_type_t cksum_type;
1624
1625                 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1626                                                body->oa.o_flags : 0);
1627                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1628                                                  aa->aa_ppga, OST_READ,
1629                                                  cksum_type);
1630
1631                 if (peer->nid == req->rq_bulk->bd_sender) {
1632                         via = router = "";
1633                 } else {
1634                         via = " via ";
1635                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1636                 }
1637
1638                 if (server_cksum == ~0 && rc > 0) {
1639                         CERROR("Protocol error: server %s set the 'checksum' "
1640                                "bit, but didn't send a checksum.  Not fatal, "
1641                                "but please notify on http://bugs.whamcloud.com/\n",
1642                                libcfs_nid2str(peer->nid));
1643                 } else if (server_cksum != client_cksum) {
1644                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1645                                            "%s%s%s inode "DFID" object "
1646                                            LPU64"/"LPU64" extent "
1647                                            "["LPU64"-"LPU64"]\n",
1648                                            req->rq_import->imp_obd->obd_name,
1649                                            libcfs_nid2str(peer->nid),
1650                                            via, router,
1651                                            body->oa.o_valid & OBD_MD_FLFID ?
1652                                                 body->oa.o_parent_seq : (__u64)0,
1653                                            body->oa.o_valid & OBD_MD_FLFID ?
1654                                                 body->oa.o_parent_oid : 0,
1655                                            body->oa.o_valid & OBD_MD_FLFID ?
1656                                                 body->oa.o_parent_ver : 0,
1657                                            body->oa.o_id,
1658                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1659                                                 body->oa.o_seq : (__u64)0,
1660                                            aa->aa_ppga[0]->off,
1661                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1662                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1663                                                                         1);
1664                         CERROR("client %x, server %x, cksum_type %x\n",
1665                                client_cksum, server_cksum, cksum_type);
1666                         cksum_counter = 0;
1667                         aa->aa_oa->o_cksum = client_cksum;
1668                         rc = -EAGAIN;
1669                 } else {
1670                         cksum_counter++;
1671                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1672                         rc = 0;
1673                 }
1674         } else if (unlikely(client_cksum)) {
1675                 static int cksum_missed;
1676
1677                 cksum_missed++;
1678                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1679                         CERROR("Checksum %u requested from %s but not sent\n",
1680                                cksum_missed, libcfs_nid2str(peer->nid));
1681         } else {
1682                 rc = 0;
1683         }
1684 out:
1685         if (rc >= 0)
1686                 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1687
1688         RETURN(rc);
1689 }
1690
1691 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1692                             struct lov_stripe_md *lsm,
1693                             obd_count page_count, struct brw_page **pga,
1694                             struct obd_capa *ocapa)
1695 {
1696         struct ptlrpc_request *req;
1697         int                    rc;
1698         cfs_waitq_t            waitq;
1699         int                    generation, resends = 0;
1700         struct l_wait_info     lwi;
1701
1702         ENTRY;
1703
1704         cfs_waitq_init(&waitq);
1705         generation = exp->exp_obd->u.cli.cl_import->imp_generation;
1706
1707 restart_bulk:
1708         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1709                                   page_count, pga, &req, ocapa, 0, resends);
1710         if (rc != 0)
1711                 return (rc);
1712
1713         if (resends) {
1714                 req->rq_generation_set = 1;
1715                 req->rq_import_generation = generation;
1716         }
1717
1718         rc = ptlrpc_queue_wait(req);
1719
1720         if (rc == -ETIMEDOUT && req->rq_resend) {
1721                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1722                 ptlrpc_req_finished(req);
1723                 goto restart_bulk;
1724         }
1725
1726         rc = osc_brw_fini_request(req, rc);
1727
1728         ptlrpc_req_finished(req);
1729         /* When server return -EINPROGRESS, client should always retry
1730          * regardless of the number of times the bulk was resent already.*/
1731         if (osc_recoverable_error(rc)) {
1732                 resends++;
1733                 if (rc != -EINPROGRESS &&
1734                     !client_should_resend(resends, &exp->exp_obd->u.cli)) {
1735                         CERROR("%s: too many resend retries for object: "
1736                                ""LPU64":"LPU64", rc = %d.\n",
1737                                exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1738                         goto out;
1739                 }
1740                 if (generation !=
1741                     exp->exp_obd->u.cli.cl_import->imp_generation) {
1742                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
1743                                ""LPU64":"LPU64", rc = %d.\n",
1744                                exp->exp_obd->obd_name, oa->o_id, oa->o_seq, rc);
1745                         goto out;
1746                 }
1747
1748                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1749                                        NULL);
1750                 l_wait_event(waitq, 0, &lwi);
1751
1752                 goto restart_bulk;
1753         }
1754 out:
1755         if (rc == -EAGAIN || rc == -EINPROGRESS)
1756                 rc = -EIO;
1757         RETURN (rc);
1758 }
1759
1760 int osc_brw_redo_request(struct ptlrpc_request *request,
1761                          struct osc_brw_async_args *aa)
1762 {
1763         struct ptlrpc_request *new_req;
1764         struct ptlrpc_request_set *set = request->rq_set;
1765         struct osc_brw_async_args *new_aa;
1766         struct osc_async_page *oap;
1767         int rc = 0;
1768         ENTRY;
1769
1770         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1771
1772         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1773                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1774                                   aa->aa_cli, aa->aa_oa,
1775                                   NULL /* lsm unused by osc currently */,
1776                                   aa->aa_page_count, aa->aa_ppga,
1777                                   &new_req, aa->aa_ocapa, 0, 1);
1778         if (rc)
1779                 RETURN(rc);
1780
1781         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1782
1783         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1784                 if (oap->oap_request != NULL) {
1785                         LASSERTF(request == oap->oap_request,
1786                                  "request %p != oap_request %p\n",
1787                                  request, oap->oap_request);
1788                         if (oap->oap_interrupted) {
1789                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1790                                 ptlrpc_req_finished(new_req);
1791                                 RETURN(-EINTR);
1792                         }
1793                 }
1794         }
1795         /* New request takes over pga and oaps from old request.
1796          * Note that copying a list_head doesn't work, need to move it... */
1797         aa->aa_resends++;
1798         new_req->rq_interpret_reply = request->rq_interpret_reply;
1799         new_req->rq_async_args = request->rq_async_args;
1800         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1801         new_req->rq_generation_set = 1;
1802         new_req->rq_import_generation = request->rq_import_generation;
1803
1804         new_aa = ptlrpc_req_async_args(new_req);
1805
1806         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1807         cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1808         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1809
1810         cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1811                 if (oap->oap_request) {
1812                         ptlrpc_req_finished(oap->oap_request);
1813                         oap->oap_request = ptlrpc_request_addref(new_req);
1814                 }
1815         }
1816
1817         new_aa->aa_ocapa = aa->aa_ocapa;
1818         aa->aa_ocapa = NULL;
1819
1820         /* use ptlrpc_set_add_req is safe because interpret functions work
1821          * in check_set context. only one way exist with access to request
1822          * from different thread got -EINTR - this way protected with
1823          * cl_loi_list_lock */
1824         ptlrpc_set_add_req(set, new_req);
1825
1826         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1827
1828         DEBUG_REQ(D_INFO, new_req, "new request");
1829         RETURN(0);
1830 }
1831
1832 /*
1833  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1834  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1835  * fine for our small page arrays and doesn't require allocation.  its an
1836  * insertion sort that swaps elements that are strides apart, shrinking the
1837  * stride down until its '1' and the array is sorted.
1838  */
1839 static void sort_brw_pages(struct brw_page **array, int num)
1840 {
1841         int stride, i, j;
1842         struct brw_page *tmp;
1843
1844         if (num == 1)
1845                 return;
1846         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1847                 ;
1848
1849         do {
1850                 stride /= 3;
1851                 for (i = stride ; i < num ; i++) {
1852                         tmp = array[i];
1853                         j = i;
1854                         while (j >= stride && array[j - stride]->off > tmp->off) {
1855                                 array[j] = array[j - stride];
1856                                 j -= stride;
1857                         }
1858                         array[j] = tmp;
1859                 }
1860         } while (stride > 1);
1861 }
1862
1863 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1864 {
1865         int count = 1;
1866         int offset;
1867         int i = 0;
1868
1869         LASSERT (pages > 0);
1870         offset = pg[i]->off & ~CFS_PAGE_MASK;
1871
1872         for (;;) {
1873                 pages--;
1874                 if (pages == 0)         /* that's all */
1875                         return count;
1876
1877                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1878                         return count;   /* doesn't end on page boundary */
1879
1880                 i++;
1881                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1882                 if (offset != 0)        /* doesn't start on page boundary */
1883                         return count;
1884
1885                 count++;
1886         }
1887 }
1888
1889 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1890 {
1891         struct brw_page **ppga;
1892         int i;
1893
1894         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1895         if (ppga == NULL)
1896                 return NULL;
1897
1898         for (i = 0; i < count; i++)
1899                 ppga[i] = pga + i;
1900         return ppga;
1901 }
1902
1903 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1904 {
1905         LASSERT(ppga != NULL);
1906         OBD_FREE(ppga, sizeof(*ppga) * count);
1907 }
1908
1909 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1910                    obd_count page_count, struct brw_page *pga,
1911                    struct obd_trans_info *oti)
1912 {
1913         struct obdo *saved_oa = NULL;
1914         struct brw_page **ppga, **orig;
1915         struct obd_import *imp = class_exp2cliimp(exp);
1916         struct client_obd *cli;
1917         int rc, page_count_orig;
1918         ENTRY;
1919
1920         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1921         cli = &imp->imp_obd->u.cli;
1922
1923         if (cmd & OBD_BRW_CHECK) {
1924                 /* The caller just wants to know if there's a chance that this
1925                  * I/O can succeed */
1926
1927                 if (imp->imp_invalid)
1928                         RETURN(-EIO);
1929                 RETURN(0);
1930         }
1931
1932         /* test_brw with a failed create can trip this, maybe others. */
1933         LASSERT(cli->cl_max_pages_per_rpc);
1934
1935         rc = 0;
1936
1937         orig = ppga = osc_build_ppga(pga, page_count);
1938         if (ppga == NULL)
1939                 RETURN(-ENOMEM);
1940         page_count_orig = page_count;
1941
1942         sort_brw_pages(ppga, page_count);
1943         while (page_count) {
1944                 obd_count pages_per_brw;
1945
1946                 if (page_count > cli->cl_max_pages_per_rpc)
1947                         pages_per_brw = cli->cl_max_pages_per_rpc;
1948                 else
1949                         pages_per_brw = page_count;
1950
1951                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1952
1953                 if (saved_oa != NULL) {
1954                         /* restore previously saved oa */
1955                         *oinfo->oi_oa = *saved_oa;
1956                 } else if (page_count > pages_per_brw) {
1957                         /* save a copy of oa (brw will clobber it) */
1958                         OBDO_ALLOC(saved_oa);
1959                         if (saved_oa == NULL)
1960                                 GOTO(out, rc = -ENOMEM);
1961                         *saved_oa = *oinfo->oi_oa;
1962                 }
1963
1964                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1965                                       pages_per_brw, ppga, oinfo->oi_capa);
1966
1967                 if (rc != 0)
1968                         break;
1969
1970                 page_count -= pages_per_brw;
1971                 ppga += pages_per_brw;
1972         }
1973
1974 out:
1975         osc_release_ppga(orig, page_count_orig);
1976
1977         if (saved_oa != NULL)
1978                 OBDO_FREE(saved_oa);
1979
1980         RETURN(rc);
1981 }
1982
1983 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1984  * the dirty accounting.  Writeback completes or truncate happens before
1985  * writing starts.  Must be called with the loi lock held. */
1986 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1987                            int sent)
1988 {
1989         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1990 }
1991
1992
1993 /* This maintains the lists of pending pages to read/write for a given object
1994  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1995  * to quickly find objects that are ready to send an RPC. */
1996 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1997                          int cmd)
1998 {
1999         ENTRY;
2000
2001         if (lop->lop_num_pending == 0)
2002                 RETURN(0);
2003
2004         /* if we have an invalid import we want to drain the queued pages
2005          * by forcing them through rpcs that immediately fail and complete
2006          * the pages.  recovery relies on this to empty the queued pages
2007          * before canceling the locks and evicting down the llite pages */
2008         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2009                 RETURN(1);
2010
2011         /* stream rpcs in queue order as long as as there is an urgent page
2012          * queued.  this is our cheap solution for good batching in the case
2013          * where writepage marks some random page in the middle of the file
2014          * as urgent because of, say, memory pressure */
2015         if (!cfs_list_empty(&lop->lop_urgent)) {
2016                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
2017                 RETURN(1);
2018         }
2019
2020         if (cmd & OBD_BRW_WRITE) {
2021                 /* trigger a write rpc stream as long as there are dirtiers
2022                  * waiting for space.  as they're waiting, they're not going to
2023                  * create more pages to coalesce with what's waiting.. */
2024                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
2025                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
2026                         RETURN(1);
2027                 }
2028         }
2029         if (lop->lop_num_pending >= cli->cl_max_pages_per_rpc)
2030                 RETURN(1);
2031
2032         RETURN(0);
2033 }
2034
2035 static int lop_makes_hprpc(struct loi_oap_pages *lop)
2036 {
2037         struct osc_async_page *oap;
2038         ENTRY;
2039
2040         if (cfs_list_empty(&lop->lop_urgent))
2041                 RETURN(0);
2042
2043         oap = cfs_list_entry(lop->lop_urgent.next,
2044                          struct osc_async_page, oap_urgent_item);
2045
2046         if (oap->oap_async_flags & ASYNC_HP) {
2047                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
2048                 RETURN(1);
2049         }
2050
2051         RETURN(0);
2052 }
2053
2054 static void on_list(cfs_list_t *item, cfs_list_t *list,
2055                     int should_be_on)
2056 {
2057         if (cfs_list_empty(item) && should_be_on)
2058                 cfs_list_add_tail(item, list);
2059         else if (!cfs_list_empty(item) && !should_be_on)
2060                 cfs_list_del_init(item);
2061 }
2062
2063 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
2064  * can find pages to build into rpcs quickly */
2065 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
2066 {
2067         if (lop_makes_hprpc(&loi->loi_write_lop) ||
2068             lop_makes_hprpc(&loi->loi_read_lop)) {
2069                 /* HP rpc */
2070                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2071                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2072         } else {
2073                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2074                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2075                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2076                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2077         }
2078
2079         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2080                 loi->loi_write_lop.lop_num_pending);
2081
2082         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2083                 loi->loi_read_lop.lop_num_pending);
2084 }
2085
2086 static void lop_update_pending(struct client_obd *cli,
2087                                struct loi_oap_pages *lop, int cmd, int delta)
2088 {
2089         lop->lop_num_pending += delta;
2090         if (cmd & OBD_BRW_WRITE)
2091                 cli->cl_pending_w_pages += delta;
2092         else
2093                 cli->cl_pending_r_pages += delta;
2094 }
2095
2096 /**
2097  * this is called when a sync waiter receives an interruption.  Its job is to
2098  * get the caller woken as soon as possible.  If its page hasn't been put in an
2099  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2100  * desiring interruption which will forcefully complete the rpc once the rpc
2101  * has timed out.
2102  */
2103 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2104 {
2105         struct loi_oap_pages *lop;
2106         struct lov_oinfo *loi;
2107         int rc = -EBUSY;
2108         ENTRY;
2109
2110         LASSERT(!oap->oap_interrupted);
2111         oap->oap_interrupted = 1;
2112
2113         /* ok, it's been put in an rpc. only one oap gets a request reference */
2114         if (oap->oap_request != NULL) {
2115                 ptlrpc_mark_interrupted(oap->oap_request);
2116                 ptlrpcd_wake(oap->oap_request);
2117                 ptlrpc_req_finished(oap->oap_request);
2118                 oap->oap_request = NULL;
2119         }
2120
2121         /*
2122          * page completion may be called only if ->cpo_prep() method was
2123          * executed by osc_io_submit(), that also adds page the to pending list
2124          */
2125         if (!cfs_list_empty(&oap->oap_pending_item)) {
2126                 cfs_list_del_init(&oap->oap_pending_item);
2127                 cfs_list_del_init(&oap->oap_urgent_item);
2128
2129                 loi = oap->oap_loi;
2130                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2131                         &loi->loi_write_lop : &loi->loi_read_lop;
2132                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2133                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2134                 rc = oap->oap_caller_ops->ap_completion(env,
2135                                           oap->oap_caller_data,
2136                                           oap->oap_cmd, NULL, -EINTR);
2137         }
2138
2139         RETURN(rc);
2140 }
2141
2142 /* this is trying to propogate async writeback errors back up to the
2143  * application.  As an async write fails we record the error code for later if
2144  * the app does an fsync.  As long as errors persist we force future rpcs to be
2145  * sync so that the app can get a sync error and break the cycle of queueing
2146  * pages for which writeback will fail. */
2147 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2148                            int rc)
2149 {
2150         if (rc) {
2151                 if (!ar->ar_rc)
2152                         ar->ar_rc = rc;
2153
2154                 ar->ar_force_sync = 1;
2155                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2156                 return;
2157
2158         }
2159
2160         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2161                 ar->ar_force_sync = 0;
2162 }
2163
2164 void osc_oap_to_pending(struct osc_async_page *oap)
2165 {
2166         struct loi_oap_pages *lop;
2167
2168         if (oap->oap_cmd & OBD_BRW_WRITE)
2169                 lop = &oap->oap_loi->loi_write_lop;
2170         else
2171                 lop = &oap->oap_loi->loi_read_lop;
2172
2173         if (oap->oap_async_flags & ASYNC_HP)
2174                 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2175         else if (oap->oap_async_flags & ASYNC_URGENT)
2176                 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2177         cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2178         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2179 }
2180
2181 /* this must be called holding the loi list lock to give coverage to exit_cache,
2182  * async_flag maintenance, and oap_request */
2183 static void osc_ap_completion(const struct lu_env *env,
2184                               struct client_obd *cli, struct obdo *oa,
2185                               struct osc_async_page *oap, int sent, int rc)
2186 {
2187         __u64 xid = 0;
2188
2189         ENTRY;
2190         if (oap->oap_request != NULL) {
2191                 xid = ptlrpc_req_xid(oap->oap_request);
2192                 ptlrpc_req_finished(oap->oap_request);
2193                 oap->oap_request = NULL;
2194         }
2195
2196         cfs_spin_lock(&oap->oap_lock);
2197         oap->oap_async_flags = 0;
2198         cfs_spin_unlock(&oap->oap_lock);
2199         oap->oap_interrupted = 0;
2200
2201         if (oap->oap_cmd & OBD_BRW_WRITE) {
2202                 osc_process_ar(&cli->cl_ar, xid, rc);
2203                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2204         }
2205
2206         if (rc == 0 && oa != NULL) {
2207                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2208                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2209                 if (oa->o_valid & OBD_MD_FLMTIME)
2210                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2211                 if (oa->o_valid & OBD_MD_FLATIME)
2212                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2213                 if (oa->o_valid & OBD_MD_FLCTIME)
2214                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2215         }
2216
2217         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2218                                                 oap->oap_cmd, oa, rc);
2219
2220         /* cl_page_completion() drops PG_locked. so, a new I/O on the page could
2221          * start, but OSC calls it under lock and thus we can add oap back to
2222          * pending safely */
2223         if (rc)
2224                 /* upper layer wants to leave the page on pending queue */
2225                 osc_oap_to_pending(oap);
2226         else
2227                 osc_exit_cache(cli, oap, sent);
2228         EXIT;
2229 }
2230
2231 static int brw_queue_work(const struct lu_env *env, void *data)
2232 {
2233         struct client_obd *cli = data;
2234
2235         CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
2236
2237         client_obd_list_lock(&cli->cl_loi_list_lock);
2238         osc_check_rpcs0(env, cli, 1);
2239         client_obd_list_unlock(&cli->cl_loi_list_lock);
2240         RETURN(0);
2241 }
2242
2243 static int brw_interpret(const struct lu_env *env,
2244                          struct ptlrpc_request *req, void *data, int rc)
2245 {
2246         struct osc_brw_async_args *aa = data;
2247         struct client_obd *cli;
2248         int async;
2249         ENTRY;
2250
2251         rc = osc_brw_fini_request(req, rc);
2252         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2253         /* When server return -EINPROGRESS, client should always retry
2254          * regardless of the number of times the bulk was resent already. */
2255         if (osc_recoverable_error(rc)) {
2256                 if (req->rq_import_generation !=
2257                     req->rq_import->imp_generation) {
2258                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
2259                                ""LPU64":"LPU64", rc = %d.\n",
2260                                req->rq_import->imp_obd->obd_name,
2261                                aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
2262                 } else if (rc == -EINPROGRESS ||
2263                     client_should_resend(aa->aa_resends, aa->aa_cli)) {
2264                         rc = osc_brw_redo_request(req, aa);
2265                 } else {
2266                         CERROR("%s: too many resent retries for object: "
2267                                ""LPU64":"LPU64", rc = %d.\n",
2268                                req->rq_import->imp_obd->obd_name,
2269                                aa->aa_oa->o_id, aa->aa_oa->o_seq, rc);
2270                 }
2271
2272                 if (rc == 0)
2273                         RETURN(0);
2274                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2275                         rc = -EIO;
2276         }
2277
2278         if (aa->aa_ocapa) {
2279                 capa_put(aa->aa_ocapa);
2280                 aa->aa_ocapa = NULL;
2281         }
2282
2283         cli = aa->aa_cli;
2284         client_obd_list_lock(&cli->cl_loi_list_lock);
2285
2286         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2287          * is called so we know whether to go to sync BRWs or wait for more
2288          * RPCs to complete */
2289         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2290                 cli->cl_w_in_flight--;
2291         else
2292                 cli->cl_r_in_flight--;
2293
2294         async = cfs_list_empty(&aa->aa_oaps);
2295         if (!async) { /* from osc_send_oap_rpc() */
2296                 struct osc_async_page *oap, *tmp;
2297                 /* the caller may re-use the oap after the completion call so
2298                  * we need to clean it up a little */
2299                 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2300                                              oap_rpc_item) {
2301                         cfs_list_del_init(&oap->oap_rpc_item);
2302                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2303                 }
2304                 OBDO_FREE(aa->aa_oa);
2305         } else { /* from async_internal() */
2306                 obd_count i;
2307                 for (i = 0; i < aa->aa_page_count; i++)
2308                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2309         }
2310         osc_wake_cache_waiters(cli);
2311         osc_check_rpcs0(env, cli, 1);
2312         client_obd_list_unlock(&cli->cl_loi_list_lock);
2313
2314         if (!async)
2315                 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2316                                   req->rq_bulk->bd_nob_transferred);
2317         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2318         ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
2319
2320         RETURN(rc);
2321 }
2322
2323 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2324                                             struct client_obd *cli,
2325                                             cfs_list_t *rpc_list,
2326                                             int page_count, int cmd)
2327 {
2328         struct ptlrpc_request *req;
2329         struct brw_page **pga = NULL;
2330         struct osc_brw_async_args *aa;
2331         struct obdo *oa = NULL;
2332         const struct obd_async_page_ops *ops = NULL;
2333         struct osc_async_page *oap;
2334         struct osc_async_page *tmp;
2335         struct cl_req *clerq = NULL;
2336         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2337         struct ldlm_lock *lock = NULL;
2338         struct cl_req_attr crattr;
2339         int i, rc, mpflag = 0;
2340
2341         ENTRY;
2342         LASSERT(!cfs_list_empty(rpc_list));
2343
2344         if (cmd & OBD_BRW_MEMALLOC)
2345                 mpflag = cfs_memory_pressure_get_and_set();
2346
2347         memset(&crattr, 0, sizeof crattr);
2348         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2349         if (pga == NULL)
2350                 GOTO(out, req = ERR_PTR(-ENOMEM));
2351
2352         OBDO_ALLOC(oa);
2353         if (oa == NULL)
2354                 GOTO(out, req = ERR_PTR(-ENOMEM));
2355
2356         i = 0;
2357         cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2358                 struct cl_page *page = osc_oap2cl_page(oap);
2359                 if (ops == NULL) {
2360                         ops = oap->oap_caller_ops;
2361
2362                         clerq = cl_req_alloc(env, page, crt,
2363                                              1 /* only 1-object rpcs for
2364                                                 * now */);
2365                         if (IS_ERR(clerq))
2366                                 GOTO(out, req = (void *)clerq);
2367                         lock = oap->oap_ldlm_lock;
2368                 }
2369                 pga[i] = &oap->oap_brw_page;
2370                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2371                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2372                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2373                 i++;
2374                 cl_req_page_add(env, clerq, page);
2375         }
2376
2377         /* always get the data for the obdo for the rpc */
2378         LASSERT(ops != NULL);
2379         crattr.cra_oa = oa;
2380         crattr.cra_capa = NULL;
2381         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2382         if (lock) {
2383                 oa->o_handle = lock->l_remote_handle;
2384                 oa->o_valid |= OBD_MD_FLHANDLE;
2385         }
2386
2387         rc = cl_req_prep(env, clerq);
2388         if (rc != 0) {
2389                 CERROR("cl_req_prep failed: %d\n", rc);
2390                 GOTO(out, req = ERR_PTR(rc));
2391         }
2392
2393         sort_brw_pages(pga, page_count);
2394         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2395                                   pga, &req, crattr.cra_capa, 1, 0);
2396         if (rc != 0) {
2397                 CERROR("prep_req failed: %d\n", rc);
2398                 GOTO(out, req = ERR_PTR(rc));
2399         }
2400
2401         if (cmd & OBD_BRW_MEMALLOC)
2402                 req->rq_memalloc = 1;
2403
2404         /* Need to update the timestamps after the request is built in case
2405          * we race with setattr (locally or in queue at OST).  If OST gets
2406          * later setattr before earlier BRW (as determined by the request xid),
2407          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2408          * way to do this in a single call.  bug 10150 */
2409         cl_req_attr_set(env, clerq, &crattr,
2410                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2411
2412         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2413         aa = ptlrpc_req_async_args(req);
2414         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2415         cfs_list_splice(rpc_list, &aa->aa_oaps);
2416         CFS_INIT_LIST_HEAD(rpc_list);
2417         aa->aa_clerq = clerq;
2418 out:
2419         if (cmd & OBD_BRW_MEMALLOC)
2420                 cfs_memory_pressure_restore(mpflag);
2421
2422         capa_put(crattr.cra_capa);
2423         if (IS_ERR(req)) {
2424                 if (oa)
2425                         OBDO_FREE(oa);
2426                 if (pga)
2427                         OBD_FREE(pga, sizeof(*pga) * page_count);
2428                 /* this should happen rarely and is pretty bad, it makes the
2429                  * pending list not follow the dirty order */
2430                 client_obd_list_lock(&cli->cl_loi_list_lock);
2431                 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2432                         cfs_list_del_init(&oap->oap_rpc_item);
2433
2434                         /* queued sync pages can be torn down while the pages
2435                          * were between the pending list and the rpc */
2436                         if (oap->oap_interrupted) {
2437                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2438                                 osc_ap_completion(env, cli, NULL, oap, 0,
2439                                                   oap->oap_count);
2440                                 continue;
2441                         }
2442                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2443                 }
2444                 if (clerq && !IS_ERR(clerq))
2445                         cl_req_completion(env, clerq, PTR_ERR(req));
2446         }
2447         RETURN(req);
2448 }
2449
2450 /**
2451  * prepare pages for ASYNC io and put pages in send queue.
2452  *
2453  * \param cmd OBD_BRW_* macroses
2454  * \param lop pending pages
2455  *
2456  * \return zero if no page added to send queue.
2457  * \return 1 if pages successfully added to send queue.
2458  * \return negative on errors.
2459  */
2460 static int
2461 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2462                  struct lov_oinfo *loi, int cmd,
2463                  struct loi_oap_pages *lop, pdl_policy_t pol)
2464 {
2465         struct ptlrpc_request *req;
2466         obd_count page_count = 0;
2467         struct osc_async_page *oap = NULL, *tmp;
2468         struct osc_brw_async_args *aa;
2469         const struct obd_async_page_ops *ops;
2470         CFS_LIST_HEAD(rpc_list);
2471         int srvlock = 0, mem_tight = 0;
2472         struct cl_object *clob = NULL;
2473         obd_off starting_offset = OBD_OBJECT_EOF;
2474         unsigned int ending_offset;
2475         int starting_page_off = 0;
2476         ENTRY;
2477
2478         /* ASYNC_HP pages first. At present, when the lock the pages is
2479          * to be canceled, the pages covered by the lock will be sent out
2480          * with ASYNC_HP. We have to send out them as soon as possible. */
2481         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2482                 if (oap->oap_async_flags & ASYNC_HP)
2483                         cfs_list_move(&oap->oap_pending_item, &rpc_list);
2484                 else if (!(oap->oap_brw_flags & OBD_BRW_SYNC))
2485                         /* only do this for writeback pages. */
2486                         cfs_list_move_tail(&oap->oap_pending_item, &rpc_list);
2487                 if (++page_count >= cli->cl_max_pages_per_rpc)
2488                         break;
2489         }
2490         cfs_list_splice_init(&rpc_list, &lop->lop_pending);
2491         page_count = 0;
2492
2493         /* first we find the pages we're allowed to work with */
2494         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2495                                      oap_pending_item) {
2496                 ops = oap->oap_caller_ops;
2497
2498                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2499                          "magic 0x%x\n", oap, oap->oap_magic);
2500
2501                 if (clob == NULL) {
2502                         /* pin object in memory, so that completion call-backs
2503                          * can be safely called under client_obd_list lock. */
2504                         clob = osc_oap2cl_page(oap)->cp_obj;
2505                         cl_object_get(clob);
2506                 }
2507
2508                 if (page_count != 0 &&
2509                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2510                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2511                                " oap %p, page %p, srvlock %u\n",
2512                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2513                         break;
2514                 }
2515
2516                 /* If there is a gap at the start of this page, it can't merge
2517                  * with any previous page, so we'll hand the network a
2518                  * "fragmented" page array that it can't transfer in 1 RDMA */
2519                 if (oap->oap_obj_off < starting_offset) {
2520                         if (starting_page_off != 0)
2521                                 break;
2522
2523                         starting_page_off = oap->oap_page_off;
2524                         starting_offset = oap->oap_obj_off + starting_page_off;
2525                 } else if (oap->oap_page_off != 0)
2526                         break;
2527
2528                 /* in llite being 'ready' equates to the page being locked
2529                  * until completion unlocks it.  commit_write submits a page
2530                  * as not ready because its unlock will happen unconditionally
2531                  * as the call returns.  if we race with commit_write giving
2532                  * us that page we don't want to create a hole in the page
2533                  * stream, so we stop and leave the rpc to be fired by
2534                  * another dirtier or kupdated interval (the not ready page
2535                  * will still be on the dirty list).  we could call in
2536                  * at the end of ll_file_write to process the queue again. */
2537                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2538                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2539                                                     cmd);
2540                         if (rc < 0)
2541                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2542                                                 "instead of ready\n", oap,
2543                                                 oap->oap_page, rc);
2544                         switch (rc) {
2545                         case -EAGAIN:
2546                                 /* llite is telling us that the page is still
2547                                  * in commit_write and that we should try
2548                                  * and put it in an rpc again later.  we
2549                                  * break out of the loop so we don't create
2550                                  * a hole in the sequence of pages in the rpc
2551                                  * stream.*/
2552                                 oap = NULL;
2553                                 break;
2554                         case -EINTR:
2555                                 /* the io isn't needed.. tell the checks
2556                                  * below to complete the rpc with EINTR */
2557                                 cfs_spin_lock(&oap->oap_lock);
2558                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2559                                 cfs_spin_unlock(&oap->oap_lock);
2560                                 oap->oap_count = -EINTR;
2561                                 break;
2562                         case 0:
2563                                 cfs_spin_lock(&oap->oap_lock);
2564                                 oap->oap_async_flags |= ASYNC_READY;
2565                                 cfs_spin_unlock(&oap->oap_lock);
2566                                 break;
2567                         default:
2568                                 LASSERTF(0, "oap %p page %p returned %d "
2569                                             "from make_ready\n", oap,
2570                                             oap->oap_page, rc);
2571                                 break;
2572                         }
2573                 }
2574                 if (oap == NULL)
2575                         break;
2576
2577                 /* take the page out of our book-keeping */
2578                 cfs_list_del_init(&oap->oap_pending_item);
2579                 lop_update_pending(cli, lop, cmd, -1);
2580                 cfs_list_del_init(&oap->oap_urgent_item);
2581
2582                 /* ask the caller for the size of the io as the rpc leaves. */
2583                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2584                         oap->oap_count =
2585                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2586                                                       cmd);
2587                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2588                 }
2589                 if (oap->oap_count <= 0) {
2590                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2591                                oap->oap_count);
2592                         osc_ap_completion(env, cli, NULL,
2593                                           oap, 0, oap->oap_count);
2594                         continue;
2595                 }
2596
2597                 /* now put the page back in our accounting */
2598                 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2599                 if (page_count++ == 0)
2600                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2601
2602                 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2603                         mem_tight = 1;
2604
2605                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2606                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2607                  * have the same alignment as the initial writes that allocated
2608                  * extents on the server. */
2609                 ending_offset = oap->oap_obj_off + oap->oap_page_off +
2610                                 oap->oap_count;
2611                 if (!(ending_offset & (PTLRPC_MAX_BRW_SIZE - 1)))
2612                         break;
2613
2614                 if (page_count >= cli->cl_max_pages_per_rpc)
2615                         break;
2616
2617                 /* If there is a gap at the end of this page, it can't merge
2618                  * with any subsequent pages, so we'll hand the network a
2619                  * "fragmented" page array that it can't transfer in 1 RDMA */
2620                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2621                         break;
2622         }
2623
2624         loi_list_maint(cli, loi);
2625
2626         client_obd_list_unlock(&cli->cl_loi_list_lock);
2627
2628         if (clob != NULL)
2629                 cl_object_put(env, clob);
2630
2631         if (page_count == 0) {
2632                 client_obd_list_lock(&cli->cl_loi_list_lock);
2633                 RETURN(0);
2634         }
2635
2636         req = osc_build_req(env, cli, &rpc_list, page_count,
2637                             mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2638         if (IS_ERR(req)) {
2639                 LASSERT(cfs_list_empty(&rpc_list));
2640                 loi_list_maint(cli, loi);
2641                 RETURN(PTR_ERR(req));
2642         }
2643
2644         aa = ptlrpc_req_async_args(req);
2645
2646         starting_offset &= PTLRPC_MAX_BRW_SIZE - 1;
2647         if (cmd == OBD_BRW_READ) {
2648                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2649                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2650                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2651                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2652         } else {
2653                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2654                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2655                                  cli->cl_w_in_flight);
2656                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2657                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2658         }
2659
2660         client_obd_list_lock(&cli->cl_loi_list_lock);
2661
2662         if (cmd == OBD_BRW_READ)
2663                 cli->cl_r_in_flight++;
2664         else
2665                 cli->cl_w_in_flight++;
2666
2667         /* queued sync pages can be torn down while the pages
2668          * were between the pending list and the rpc */
2669         tmp = NULL;
2670         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2671                 /* only one oap gets a request reference */
2672                 if (tmp == NULL)
2673                         tmp = oap;
2674                 if (oap->oap_interrupted && !req->rq_intr) {
2675                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2676                                oap, req);
2677                         ptlrpc_mark_interrupted(req);
2678                 }
2679         }
2680         if (tmp != NULL)
2681                 tmp->oap_request = ptlrpc_request_addref(req);
2682
2683         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2684                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2685
2686         req->rq_interpret_reply = brw_interpret;
2687
2688         /* XXX: Maybe the caller can check the RPC bulk descriptor to see which
2689          *      CPU/NUMA node the majority of pages were allocated on, and try
2690          *      to assign the async RPC to the CPU core (PDL_POLICY_PREFERRED)
2691          *      to reduce cross-CPU memory traffic.
2692          *
2693          *      But on the other hand, we expect that multiple ptlrpcd threads
2694          *      and the initial write sponsor can run in parallel, especially
2695          *      when data checksum is enabled, which is CPU-bound operation and
2696          *      single ptlrpcd thread cannot process in time. So more ptlrpcd
2697          *      threads sharing BRW load (with PDL_POLICY_ROUND) seems better.
2698          */
2699         ptlrpcd_add_req(req, pol, -1);
2700         RETURN(1);
2701 }
2702
2703 #define LOI_DEBUG(LOI, STR, args...)                                     \
2704         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2705                !cfs_list_empty(&(LOI)->loi_ready_item) ||                \
2706                !cfs_list_empty(&(LOI)->loi_hp_ready_item),               \
2707                (LOI)->loi_write_lop.lop_num_pending,                     \
2708                !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent),        \
2709                (LOI)->loi_read_lop.lop_num_pending,                      \
2710                !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent),         \
2711                args)                                                     \
2712
2713 /* This is called by osc_check_rpcs() to find which objects have pages that
2714  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2715 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2716 {
2717         ENTRY;
2718
2719         /* First return objects that have blocked locks so that they
2720          * will be flushed quickly and other clients can get the lock,
2721          * then objects which have pages ready to be stuffed into RPCs */
2722         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2723                 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2724                                       struct lov_oinfo, loi_hp_ready_item));
2725         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2726                 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2727                                       struct lov_oinfo, loi_ready_item));
2728
2729         /* then if we have cache waiters, return all objects with queued
2730          * writes.  This is especially important when many small files
2731          * have filled up the cache and not been fired into rpcs because
2732          * they don't pass the nr_pending/object threshhold */
2733         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2734             !cfs_list_empty(&cli->cl_loi_write_list))
2735                 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2736                                       struct lov_oinfo, loi_write_item));
2737
2738         /* then return all queued objects when we have an invalid import
2739          * so that they get flushed */
2740         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2741                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2742                         RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2743                                               struct lov_oinfo,
2744                                               loi_write_item));
2745                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2746                         RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2747                                               struct lov_oinfo, loi_read_item));
2748         }
2749         RETURN(NULL);
2750 }
2751
2752 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2753 {
2754         struct osc_async_page *oap;
2755         int hprpc = 0;
2756
2757         if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2758                 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2759                                      struct osc_async_page, oap_urgent_item);
2760                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2761         }
2762
2763         if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2764                 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2765                                      struct osc_async_page, oap_urgent_item);
2766                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2767         }
2768
2769         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2770 }
2771
2772 /* called with the loi list lock held */
2773 static void osc_check_rpcs0(const struct lu_env *env, struct client_obd *cli, int ptlrpc)
2774 {
2775         struct lov_oinfo *loi;
2776         int rc = 0, race_counter = 0;
2777         pdl_policy_t pol;
2778         ENTRY;
2779
2780         pol = ptlrpc ? PDL_POLICY_SAME : PDL_POLICY_ROUND;
2781
2782         while ((loi = osc_next_loi(cli)) != NULL) {
2783                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2784
2785                 if (osc_max_rpc_in_flight(cli, loi))
2786                         break;
2787
2788                 /* attempt some read/write balancing by alternating between
2789                  * reads and writes in an object.  The makes_rpc checks here
2790                  * would be redundant if we were getting read/write work items
2791                  * instead of objects.  we don't want send_oap_rpc to drain a
2792                  * partial read pending queue when we're given this object to
2793                  * do io on writes while there are cache waiters */
2794                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2795                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2796                                               &loi->loi_write_lop, pol);
2797                         if (rc < 0) {
2798                                 CERROR("Write request failed with %d\n", rc);
2799
2800                                 /* osc_send_oap_rpc failed, mostly because of
2801                                  * memory pressure.
2802                                  *
2803                                  * It can't break here, because if:
2804                                  *  - a page was submitted by osc_io_submit, so
2805                                  *    page locked;
2806                                  *  - no request in flight
2807                                  *  - no subsequent request
2808                                  * The system will be in live-lock state,
2809                                  * because there is no chance to call
2810                                  * osc_io_unplug() and osc_check_rpcs() any
2811                                  * more. pdflush can't help in this case,
2812                                  * because it might be blocked at grabbing
2813                                  * the page lock as we mentioned.
2814                                  *
2815                                  * Anyway, continue to drain pages. */
2816                                 /* break; */
2817                         }
2818
2819                         if (rc > 0)
2820                                 race_counter = 0;
2821                         else if (rc == 0)
2822                                 race_counter++;
2823                 }
2824                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2825                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2826                                               &loi->loi_read_lop, pol);
2827                         if (rc < 0)
2828                                 CERROR("Read request failed with %d\n", rc);
2829
2830                         if (rc > 0)
2831                                 race_counter = 0;
2832                         else if (rc == 0)
2833                                 race_counter++;
2834                 }
2835
2836                 /* attempt some inter-object balancing by issuing rpcs
2837                  * for each object in turn */
2838                 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2839                         cfs_list_del_init(&loi->loi_hp_ready_item);
2840                 if (!cfs_list_empty(&loi->loi_ready_item))
2841                         cfs_list_del_init(&loi->loi_ready_item);
2842                 if (!cfs_list_empty(&loi->loi_write_item))
2843                         cfs_list_del_init(&loi->loi_write_item);
2844                 if (!cfs_list_empty(&loi->loi_read_item))
2845                         cfs_list_del_init(&loi->loi_read_item);
2846
2847                 loi_list_maint(cli, loi);
2848
2849                 /* send_oap_rpc fails with 0 when make_ready tells it to
2850                  * back off.  llite's make_ready does this when it tries
2851                  * to lock a page queued for write that is already locked.
2852                  * we want to try sending rpcs from many objects, but we
2853                  * don't want to spin failing with 0.  */
2854                 if (race_counter == 10)
2855                         break;
2856         }
2857 }
2858
2859 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2860 {
2861         osc_check_rpcs0(env, cli, 0);
2862 }
2863
2864 /**
2865  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2866  * is available.
2867  */
2868 int osc_enter_cache_try(const struct lu_env *env,
2869                         struct client_obd *cli, struct lov_oinfo *loi,
2870                         struct osc_async_page *oap, int transient)
2871 {
2872         int has_grant;
2873
2874         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2875         if (has_grant) {
2876                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2877                 if (transient) {
2878                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2879                         cfs_atomic_inc(&obd_dirty_transit_pages);
2880                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2881                 }
2882         }
2883         return has_grant;
2884 }
2885
2886 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2887  * grant or cache space. */
2888 static int osc_enter_cache(const struct lu_env *env,
2889                            struct client_obd *cli, struct lov_oinfo *loi,
2890                            struct osc_async_page *oap)
2891 {
2892         struct osc_cache_waiter ocw;
2893         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
2894         int rc = -EDQUOT;
2895         ENTRY;
2896
2897         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2898                "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2899                cli->cl_dirty_max, obd_max_dirty_pages,
2900                cli->cl_lost_grant, cli->cl_avail_grant);
2901
2902         /* force the caller to try sync io.  this can jump the list
2903          * of queued writes and create a discontiguous rpc stream */
2904         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
2905             cli->cl_dirty_max < CFS_PAGE_SIZE     ||
2906             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
2907                 RETURN(-EDQUOT);
2908
2909         /* Hopefully normal case - cache space and write credits available */
2910         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2911             cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2912             osc_enter_cache_try(env, cli, loi, oap, 0))
2913                 RETURN(0);
2914
2915         /* We can get here for two reasons: too many dirty pages in cache, or
2916          * run out of grants. In both cases we should write dirty pages out.
2917          * Adding a cache waiter will trigger urgent write-out no matter what
2918          * RPC size will be.
2919          * The exiting condition is no avail grants and no dirty pages caching,
2920          * that really means there is no space on the OST. */
2921         cfs_waitq_init(&ocw.ocw_waitq);
2922         ocw.ocw_oap = oap;
2923         while (cli->cl_dirty > 0) {
2924                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2925                 ocw.ocw_rc = 0;
2926
2927                 loi_list_maint(cli, loi);
2928                 osc_check_rpcs(env, cli);
2929                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2930
2931                 CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
2932                        cli->cl_import->imp_obd->obd_name, &ocw, oap);
2933
2934                 rc = l_wait_event(ocw.ocw_waitq, cfs_list_empty(&ocw.ocw_entry), &lwi);
2935
2936                 client_obd_list_lock(&cli->cl_loi_list_lock);
2937                 cfs_list_del_init(&ocw.ocw_entry);
2938                 if (rc < 0)
2939                         break;
2940
2941                 rc = ocw.ocw_rc;
2942                 if (rc != -EDQUOT)
2943                         break;
2944         }
2945
2946         RETURN(rc);
2947 }
2948
2949
2950 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2951                         struct lov_oinfo *loi, cfs_page_t *page,
2952                         obd_off offset, const struct obd_async_page_ops *ops,
2953                         void *data, void **res, int nocache,
2954                         struct lustre_handle *lockh)
2955 {
2956         struct osc_async_page *oap;
2957
2958         ENTRY;
2959
2960         if (!page)
2961                 return cfs_size_round(sizeof(*oap));
2962
2963         oap = *res;
2964         oap->oap_magic = OAP_MAGIC;
2965         oap->oap_cli = &exp->exp_obd->u.cli;
2966         oap->oap_loi = loi;
2967
2968         oap->oap_caller_ops = ops;
2969         oap->oap_caller_data = data;
2970
2971         oap->oap_page = page;
2972         oap->oap_obj_off = offset;
2973         if (!client_is_remote(exp) &&
2974             cfs_capable(CFS_CAP_SYS_RESOURCE))
2975                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2976
2977         LASSERT(!(offset & ~CFS_PAGE_MASK));
2978
2979         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2980         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2981         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2982         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2983
2984         cfs_spin_lock_init(&oap->oap_lock);
2985         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2986         RETURN(0);
2987 }
2988
2989 int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
2990                        struct lov_stripe_md *lsm, struct lov_oinfo *loi,
2991                        struct osc_async_page *oap, int cmd, int off,
2992                        int count, obd_flag brw_flags, enum async_flags async_flags)
2993 {
2994         struct client_obd *cli = &exp->exp_obd->u.cli;
2995         int rc = 0;
2996         ENTRY;
2997
2998         if (oap->oap_magic != OAP_MAGIC)
2999                 RETURN(-EINVAL);
3000
3001         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
3002                 RETURN(-EIO);
3003
3004         if (!cfs_list_empty(&oap->oap_pending_item) ||
3005             !cfs_list_empty(&oap->oap_urgent_item) ||
3006             !cfs_list_empty(&oap->oap_rpc_item))
3007                 RETURN(-EBUSY);
3008
3009         /* check if the file's owner/group is over quota */
3010         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
3011                 struct cl_object *obj;
3012                 struct cl_attr    attr; /* XXX put attr into thread info */
3013                 unsigned int qid[MAXQUOTAS];
3014
3015                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
3016
3017                 cl_object_attr_lock(obj);
3018                 rc = cl_object_attr_get(env, obj, &attr);
3019                 cl_object_attr_unlock(obj);
3020
3021                 qid[USRQUOTA] = attr.cat_uid;
3022                 qid[GRPQUOTA] = attr.cat_gid;
3023                 if (rc == 0 &&
3024                     osc_quota_chkdq(cli, qid) == NO_QUOTA)
3025                         rc = -EDQUOT;
3026                 if (rc)
3027                         RETURN(rc);
3028         }
3029
3030         if (loi == NULL)
3031                 loi = lsm->lsm_oinfo[0];
3032
3033         client_obd_list_lock(&cli->cl_loi_list_lock);
3034
3035         LASSERT(off + count <= CFS_PAGE_SIZE);
3036         oap->oap_cmd = cmd;
3037         oap->oap_page_off = off;
3038         oap->oap_count = count;
3039         oap->oap_brw_flags = brw_flags;
3040         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
3041         if (cfs_memory_pressure_get())
3042                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
3043         cfs_spin_lock(&oap->oap_lock);
3044         oap->oap_async_flags = async_flags;
3045         cfs_spin_unlock(&oap->oap_lock);
3046
3047         if (cmd & OBD_BRW_WRITE) {
3048                 rc = osc_enter_cache(env, cli, loi, oap);
3049                 if (rc) {
3050                         client_obd_list_unlock(&cli->cl_loi_list_lock);
3051                         RETURN(rc);
3052                 }
3053         }
3054
3055         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3056                   cmd);
3057
3058         osc_oap_to_pending(oap);
3059         loi_list_maint(cli, loi);
3060         if (!osc_max_rpc_in_flight(cli, loi) &&
3061             lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
3062                 LASSERT(cli->cl_writeback_work != NULL);
3063                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
3064
3065                 CDEBUG(D_CACHE, "Queued writeback work for client obd %p/%d.\n",
3066                        cli, rc);
3067         }
3068         client_obd_list_unlock(&cli->cl_loi_list_lock);
3069
3070         RETURN(0);
3071 }
3072
3073 /* aka (~was & now & flag), but this is more clear :) */
3074 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3075
3076 int osc_set_async_flags_base(struct client_obd *cli,
3077                              struct lov_oinfo *loi, struct osc_async_page *oap,
3078                              obd_flag async_flags)
3079 {
3080         struct loi_oap_pages *lop;
3081         int flags = 0;
3082         ENTRY;
3083
3084         LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3085
3086         if (oap->oap_cmd & OBD_BRW_WRITE) {
3087                 lop = &loi->loi_write_lop;
3088         } else {
3089                 lop = &loi->loi_read_lop;
3090         }
3091
3092         if ((oap->oap_async_flags & async_flags) == async_flags)
3093                 RETURN(0);
3094
3095         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3096                 flags |= ASYNC_READY;
3097
3098         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3099             cfs_list_empty(&oap->oap_rpc_item)) {
3100                 if (oap->oap_async_flags & ASYNC_HP)
3101                         cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3102                 else
3103                         cfs_list_add_tail(&oap->oap_urgent_item,
3104                                           &lop->lop_urgent);
3105                 flags |= ASYNC_URGENT;
3106                 loi_list_maint(cli, loi);
3107         }
3108         cfs_spin_lock(&oap->oap_lock);
3109         oap->oap_async_flags |= flags;
3110         cfs_spin_unlock(&oap->oap_lock);
3111
3112         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3113                         oap->oap_async_flags);
3114         RETURN(0);
3115 }
3116
3117 int osc_teardown_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
3118                             struct lov_oinfo *loi, struct osc_async_page *oap)
3119 {
3120         struct client_obd *cli = &exp->exp_obd->u.cli;
3121         struct loi_oap_pages *lop;
3122         int rc = 0;
3123         ENTRY;
3124
3125         if (oap->oap_magic != OAP_MAGIC)
3126                 RETURN(-EINVAL);
3127
3128         if (loi == NULL)
3129                 loi = lsm->lsm_oinfo[0];
3130
3131         if (oap->oap_cmd & OBD_BRW_WRITE) {
3132                 lop = &loi->loi_write_lop;
3133         } else {
3134                 lop = &loi->loi_read_lop;
3135         }
3136
3137         client_obd_list_lock(&cli->cl_loi_list_lock);
3138
3139         if (!cfs_list_empty(&oap->oap_rpc_item))
3140                 GOTO(out, rc = -EBUSY);
3141
3142         osc_exit_cache(cli, oap, 0);
3143         osc_wake_cache_waiters(cli);
3144
3145         if (!cfs_list_empty(&oap->oap_urgent_item)) {
3146                 cfs_list_del_init(&oap->oap_urgent_item);
3147                 cfs_spin_lock(&oap->oap_lock);
3148                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3149                 cfs_spin_unlock(&oap->oap_lock);
3150         }
3151         if (!cfs_list_empty(&oap->oap_pending_item)) {
3152                 cfs_list_del_init(&oap->oap_pending_item);
3153                 lop_update_pending(cli, lop, oap->oap_cmd, -1);