Whamcloud - gitweb
LU-241 Support crc32c with hardware accelerated instruction as one of lustre checksums
[fs/lustre-release.git] / lustre / osc / osc_request.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  */
32 /*
33  * Copyright (c) 2011 Whamcloud, Inc.
34  */
35 /*
36  * This file is part of Lustre, http://www.lustre.org/
37  * Lustre is a trademark of Sun Microsystems, Inc.
38  */
39
40 #ifndef EXPORT_SYMTAB
41 # define EXPORT_SYMTAB
42 #endif
43 #define DEBUG_SUBSYSTEM S_OSC
44
45 #include <libcfs/libcfs.h>
46
47 #ifndef __KERNEL__
48 # include <liblustre.h>
49 #endif
50
51 #include <lustre_dlm.h>
52 #include <lustre_net.h>
53 #include <lustre/lustre_user.h>
54 #include <obd_cksum.h>
55 #include <obd_ost.h>
56 #include <obd_lov.h>
57
58 #ifdef  __CYGWIN__
59 # include <ctype.h>
60 #endif
61
62 #include <lustre_ha.h>
63 #include <lprocfs_status.h>
64 #include <lustre_log.h>
65 #include <lustre_debug.h>
66 #include <lustre_param.h>
67 #include "osc_internal.h"
68
69 static quota_interface_t *quota_interface = NULL;
70 extern quota_interface_t osc_quota_interface;
71
72 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
73 static int brw_interpret(const struct lu_env *env,
74                          struct ptlrpc_request *req, void *data, int rc);
75 int osc_cleanup(struct obd_device *obd);
76
77 /* Pack OSC object metadata for disk storage (LE byte order). */
78 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
79                       struct lov_stripe_md *lsm)
80 {
81         int lmm_size;
82         ENTRY;
83
84         lmm_size = sizeof(**lmmp);
85         if (!lmmp)
86                 RETURN(lmm_size);
87
88         if (*lmmp && !lsm) {
89                 OBD_FREE(*lmmp, lmm_size);
90                 *lmmp = NULL;
91                 RETURN(0);
92         }
93
94         if (!*lmmp) {
95                 OBD_ALLOC(*lmmp, lmm_size);
96                 if (!*lmmp)
97                         RETURN(-ENOMEM);
98         }
99
100         if (lsm) {
101                 LASSERT(lsm->lsm_object_id);
102                 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
103                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
104                 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
105         }
106
107         RETURN(lmm_size);
108 }
109
110 /* Unpack OSC object metadata from disk storage (LE byte order). */
111 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
112                         struct lov_mds_md *lmm, int lmm_bytes)
113 {
114         int lsm_size;
115         struct obd_import *imp = class_exp2cliimp(exp);
116         ENTRY;
117
118         if (lmm != NULL) {
119                 if (lmm_bytes < sizeof (*lmm)) {
120                         CERROR("lov_mds_md too small: %d, need %d\n",
121                                lmm_bytes, (int)sizeof(*lmm));
122                         RETURN(-EINVAL);
123                 }
124                 /* XXX LOV_MAGIC etc check? */
125
126                 if (lmm->lmm_object_id == 0) {
127                         CERROR("lov_mds_md: zero lmm_object_id\n");
128                         RETURN(-EINVAL);
129                 }
130         }
131
132         lsm_size = lov_stripe_md_size(1);
133         if (lsmp == NULL)
134                 RETURN(lsm_size);
135
136         if (*lsmp != NULL && lmm == NULL) {
137                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
138                 OBD_FREE(*lsmp, lsm_size);
139                 *lsmp = NULL;
140                 RETURN(0);
141         }
142
143         if (*lsmp == NULL) {
144                 OBD_ALLOC(*lsmp, lsm_size);
145                 if (*lsmp == NULL)
146                         RETURN(-ENOMEM);
147                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
148                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
149                         OBD_FREE(*lsmp, lsm_size);
150                         RETURN(-ENOMEM);
151                 }
152                 loi_init((*lsmp)->lsm_oinfo[0]);
153         }
154
155         if (lmm != NULL) {
156                 /* XXX zero *lsmp? */
157                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
158                 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
159                 LASSERT((*lsmp)->lsm_object_id);
160                 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
161         }
162
163         if (imp != NULL &&
164             (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
165                 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
166         else
167                 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
168
169         RETURN(lsm_size);
170 }
171
172 static inline void osc_pack_capa(struct ptlrpc_request *req,
173                                  struct ost_body *body, void *capa)
174 {
175         struct obd_capa *oc = (struct obd_capa *)capa;
176         struct lustre_capa *c;
177
178         if (!capa)
179                 return;
180
181         c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
182         LASSERT(c);
183         capa_cpy(c, oc);
184         body->oa.o_valid |= OBD_MD_FLOSSCAPA;
185         DEBUG_CAPA(D_SEC, c, "pack");
186 }
187
188 static inline void osc_pack_req_body(struct ptlrpc_request *req,
189                                      struct obd_info *oinfo)
190 {
191         struct ost_body *body;
192
193         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
194         LASSERT(body);
195
196         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
197         osc_pack_capa(req, body, oinfo->oi_capa);
198 }
199
200 static inline void osc_set_capa_size(struct ptlrpc_request *req,
201                                      const struct req_msg_field *field,
202                                      struct obd_capa *oc)
203 {
204         if (oc == NULL)
205                 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
206         else
207                 /* it is already calculated as sizeof struct obd_capa */
208                 ;
209 }
210
211 static int osc_getattr_interpret(const struct lu_env *env,
212                                  struct ptlrpc_request *req,
213                                  struct osc_async_args *aa, int rc)
214 {
215         struct ost_body *body;
216         ENTRY;
217
218         if (rc != 0)
219                 GOTO(out, rc);
220
221         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
222         if (body) {
223                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
224                 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
225
226                 /* This should really be sent by the OST */
227                 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
228                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
229         } else {
230                 CDEBUG(D_INFO, "can't unpack ost_body\n");
231                 rc = -EPROTO;
232                 aa->aa_oi->oi_oa->o_valid = 0;
233         }
234 out:
235         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
236         RETURN(rc);
237 }
238
239 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
240                              struct ptlrpc_request_set *set)
241 {
242         struct ptlrpc_request *req;
243         struct osc_async_args *aa;
244         int                    rc;
245         ENTRY;
246
247         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
248         if (req == NULL)
249                 RETURN(-ENOMEM);
250
251         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
252         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
253         if (rc) {
254                 ptlrpc_request_free(req);
255                 RETURN(rc);
256         }
257
258         osc_pack_req_body(req, oinfo);
259
260         ptlrpc_request_set_replen(req);
261         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
262
263         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
264         aa = ptlrpc_req_async_args(req);
265         aa->aa_oi = oinfo;
266
267         ptlrpc_set_add_req(set, req);
268         RETURN(0);
269 }
270
271 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
272 {
273         struct ptlrpc_request *req;
274         struct ost_body       *body;
275         int                    rc;
276         ENTRY;
277
278         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
279         if (req == NULL)
280                 RETURN(-ENOMEM);
281
282         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
283         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
284         if (rc) {
285                 ptlrpc_request_free(req);
286                 RETURN(rc);
287         }
288
289         osc_pack_req_body(req, oinfo);
290
291         ptlrpc_request_set_replen(req);
292
293         rc = ptlrpc_queue_wait(req);
294         if (rc)
295                 GOTO(out, rc);
296
297         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
298         if (body == NULL)
299                 GOTO(out, rc = -EPROTO);
300
301         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
302         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
303
304         /* This should really be sent by the OST */
305         oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
306         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
307
308         EXIT;
309  out:
310         ptlrpc_req_finished(req);
311         return rc;
312 }
313
314 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
315                        struct obd_trans_info *oti)
316 {
317         struct ptlrpc_request *req;
318         struct ost_body       *body;
319         int                    rc;
320         ENTRY;
321
322         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
323
324         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
325         if (req == NULL)
326                 RETURN(-ENOMEM);
327
328         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
329         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
330         if (rc) {
331                 ptlrpc_request_free(req);
332                 RETURN(rc);
333         }
334
335         osc_pack_req_body(req, oinfo);
336
337         ptlrpc_request_set_replen(req);
338
339         rc = ptlrpc_queue_wait(req);
340         if (rc)
341                 GOTO(out, rc);
342
343         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
344         if (body == NULL)
345                 GOTO(out, rc = -EPROTO);
346
347         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
348
349         EXIT;
350 out:
351         ptlrpc_req_finished(req);
352         RETURN(rc);
353 }
354
355 static int osc_setattr_interpret(const struct lu_env *env,
356                                  struct ptlrpc_request *req,
357                                  struct osc_setattr_args *sa, int rc)
358 {
359         struct ost_body *body;
360         ENTRY;
361
362         if (rc != 0)
363                 GOTO(out, rc);
364
365         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
366         if (body == NULL)
367                 GOTO(out, rc = -EPROTO);
368
369         lustre_get_wire_obdo(sa->sa_oa, &body->oa);
370 out:
371         rc = sa->sa_upcall(sa->sa_cookie, rc);
372         RETURN(rc);
373 }
374
375 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
376                            struct obd_trans_info *oti,
377                            obd_enqueue_update_f upcall, void *cookie,
378                            struct ptlrpc_request_set *rqset)
379 {
380         struct ptlrpc_request   *req;
381         struct osc_setattr_args *sa;
382         int                      rc;
383         ENTRY;
384
385         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
386         if (req == NULL)
387                 RETURN(-ENOMEM);
388
389         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
390         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
391         if (rc) {
392                 ptlrpc_request_free(req);
393                 RETURN(rc);
394         }
395
396         if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
397                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
398
399         osc_pack_req_body(req, oinfo);
400
401         ptlrpc_request_set_replen(req);
402
403         /* do mds to ost setattr asynchronously */
404         if (!rqset) {
405                 /* Do not wait for response. */
406                 ptlrpcd_add_req(req, PSCOPE_OTHER);
407         } else {
408                 req->rq_interpret_reply =
409                         (ptlrpc_interpterer_t)osc_setattr_interpret;
410
411                 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
412                 sa = ptlrpc_req_async_args(req);
413                 sa->sa_oa = oinfo->oi_oa;
414                 sa->sa_upcall = upcall;
415                 sa->sa_cookie = cookie;
416
417                 if (rqset == PTLRPCD_SET)
418                         ptlrpcd_add_req(req, PSCOPE_OTHER);
419                 else
420                         ptlrpc_set_add_req(rqset, req);
421         }
422
423         RETURN(0);
424 }
425
426 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
427                              struct obd_trans_info *oti,
428                              struct ptlrpc_request_set *rqset)
429 {
430         return osc_setattr_async_base(exp, oinfo, oti,
431                                       oinfo->oi_cb_up, oinfo, rqset);
432 }
433
434 int osc_real_create(struct obd_export *exp, struct obdo *oa,
435                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
436 {
437         struct ptlrpc_request *req;
438         struct ost_body       *body;
439         struct lov_stripe_md  *lsm;
440         int                    rc;
441         ENTRY;
442
443         LASSERT(oa);
444         LASSERT(ea);
445
446         lsm = *ea;
447         if (!lsm) {
448                 rc = obd_alloc_memmd(exp, &lsm);
449                 if (rc < 0)
450                         RETURN(rc);
451         }
452
453         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
454         if (req == NULL)
455                 GOTO(out, rc = -ENOMEM);
456
457         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
458         if (rc) {
459                 ptlrpc_request_free(req);
460                 GOTO(out, rc);
461         }
462
463         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
464         LASSERT(body);
465         lustre_set_wire_obdo(&body->oa, oa);
466
467         ptlrpc_request_set_replen(req);
468
469         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
470             oa->o_flags == OBD_FL_DELORPHAN) {
471                 DEBUG_REQ(D_HA, req,
472                           "delorphan from OST integration");
473                 /* Don't resend the delorphan req */
474                 req->rq_no_resend = req->rq_no_delay = 1;
475         }
476
477         rc = ptlrpc_queue_wait(req);
478         if (rc)
479                 GOTO(out_req, rc);
480
481         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
482         if (body == NULL)
483                 GOTO(out_req, rc = -EPROTO);
484
485         lustre_get_wire_obdo(oa, &body->oa);
486
487         /* This should really be sent by the OST */
488         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
489         oa->o_valid |= OBD_MD_FLBLKSZ;
490
491         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
492          * have valid lsm_oinfo data structs, so don't go touching that.
493          * This needs to be fixed in a big way.
494          */
495         lsm->lsm_object_id = oa->o_id;
496         lsm->lsm_object_seq = oa->o_seq;
497         *ea = lsm;
498
499         if (oti != NULL) {
500                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
501
502                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
503                         if (!oti->oti_logcookies)
504                                 oti_alloc_cookies(oti, 1);
505                         *oti->oti_logcookies = oa->o_lcookie;
506                 }
507         }
508
509         CDEBUG(D_HA, "transno: "LPD64"\n",
510                lustre_msg_get_transno(req->rq_repmsg));
511 out_req:
512         ptlrpc_req_finished(req);
513 out:
514         if (rc && !*ea)
515                 obd_free_memmd(exp, &lsm);
516         RETURN(rc);
517 }
518
519 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
520                    obd_enqueue_update_f upcall, void *cookie,
521                    struct ptlrpc_request_set *rqset)
522 {
523         struct ptlrpc_request   *req;
524         struct osc_setattr_args *sa;
525         struct ost_body         *body;
526         int                      rc;
527         ENTRY;
528
529         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
530         if (req == NULL)
531                 RETURN(-ENOMEM);
532
533         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
534         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
535         if (rc) {
536                 ptlrpc_request_free(req);
537                 RETURN(rc);
538         }
539         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
540         ptlrpc_at_set_req_timeout(req);
541
542         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
543         LASSERT(body);
544         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
545         osc_pack_capa(req, body, oinfo->oi_capa);
546
547         ptlrpc_request_set_replen(req);
548
549
550         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
551         CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
552         sa = ptlrpc_req_async_args(req);
553         sa->sa_oa     = oinfo->oi_oa;
554         sa->sa_upcall = upcall;
555         sa->sa_cookie = cookie;
556         if (rqset == PTLRPCD_SET)
557                 ptlrpcd_add_req(req, PSCOPE_OTHER);
558         else
559                 ptlrpc_set_add_req(rqset, req);
560
561         RETURN(0);
562 }
563
564 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
565                      struct obd_trans_info *oti,
566                      struct ptlrpc_request_set *rqset)
567 {
568         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
569         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
570         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
571         return osc_punch_base(exp, oinfo,
572                               oinfo->oi_cb_up, oinfo, rqset);
573 }
574
575 static int osc_sync_interpret(const struct lu_env *env,
576                               struct ptlrpc_request *req,
577                               void *arg, int rc)
578 {
579         struct osc_async_args *aa = arg;
580         struct ost_body *body;
581         ENTRY;
582
583         if (rc)
584                 GOTO(out, rc);
585
586         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
587         if (body == NULL) {
588                 CERROR ("can't unpack ost_body\n");
589                 GOTO(out, rc = -EPROTO);
590         }
591
592         *aa->aa_oi->oi_oa = body->oa;
593 out:
594         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
595         RETURN(rc);
596 }
597
598 static int osc_sync(struct obd_export *exp, struct obd_info *oinfo,
599                     obd_size start, obd_size end,
600                     struct ptlrpc_request_set *set)
601 {
602         struct ptlrpc_request *req;
603         struct ost_body       *body;
604         struct osc_async_args *aa;
605         int                    rc;
606         ENTRY;
607
608         if (!oinfo->oi_oa) {
609                 CDEBUG(D_INFO, "oa NULL\n");
610                 RETURN(-EINVAL);
611         }
612
613         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
614         if (req == NULL)
615                 RETURN(-ENOMEM);
616
617         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
618         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
619         if (rc) {
620                 ptlrpc_request_free(req);
621                 RETURN(rc);
622         }
623
624         /* overload the size and blocks fields in the oa with start/end */
625         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
626         LASSERT(body);
627         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
628         body->oa.o_size = start;
629         body->oa.o_blocks = end;
630         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
631         osc_pack_capa(req, body, oinfo->oi_capa);
632
633         ptlrpc_request_set_replen(req);
634         req->rq_interpret_reply = osc_sync_interpret;
635
636         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
637         aa = ptlrpc_req_async_args(req);
638         aa->aa_oi = oinfo;
639
640         ptlrpc_set_add_req(set, req);
641         RETURN (0);
642 }
643
644 /* Find and cancel locally locks matched by @mode in the resource found by
645  * @objid. Found locks are added into @cancel list. Returns the amount of
646  * locks added to @cancels list. */
647 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
648                                    cfs_list_t *cancels,
649                                    ldlm_mode_t mode, int lock_flags)
650 {
651         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
652         struct ldlm_res_id res_id;
653         struct ldlm_resource *res;
654         int count;
655         ENTRY;
656
657         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
658         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
659         if (res == NULL)
660                 RETURN(0);
661
662         LDLM_RESOURCE_ADDREF(res);
663         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
664                                            lock_flags, 0, NULL);
665         LDLM_RESOURCE_DELREF(res);
666         ldlm_resource_putref(res);
667         RETURN(count);
668 }
669
670 static int osc_destroy_interpret(const struct lu_env *env,
671                                  struct ptlrpc_request *req, void *data,
672                                  int rc)
673 {
674         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
675
676         cfs_atomic_dec(&cli->cl_destroy_in_flight);
677         cfs_waitq_signal(&cli->cl_destroy_waitq);
678         return 0;
679 }
680
681 static int osc_can_send_destroy(struct client_obd *cli)
682 {
683         if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
684             cli->cl_max_rpcs_in_flight) {
685                 /* The destroy request can be sent */
686                 return 1;
687         }
688         if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
689             cli->cl_max_rpcs_in_flight) {
690                 /*
691                  * The counter has been modified between the two atomic
692                  * operations.
693                  */
694                 cfs_waitq_signal(&cli->cl_destroy_waitq);
695         }
696         return 0;
697 }
698
699 /* Destroy requests can be async always on the client, and we don't even really
700  * care about the return code since the client cannot do anything at all about
701  * a destroy failure.
702  * When the MDS is unlinking a filename, it saves the file objects into a
703  * recovery llog, and these object records are cancelled when the OST reports
704  * they were destroyed and sync'd to disk (i.e. transaction committed).
705  * If the client dies, or the OST is down when the object should be destroyed,
706  * the records are not cancelled, and when the OST reconnects to the MDS next,
707  * it will retrieve the llog unlink logs and then sends the log cancellation
708  * cookies to the MDS after committing destroy transactions. */
709 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
710                        struct lov_stripe_md *ea, struct obd_trans_info *oti,
711                        struct obd_export *md_export, void *capa)
712 {
713         struct client_obd     *cli = &exp->exp_obd->u.cli;
714         struct ptlrpc_request *req;
715         struct ost_body       *body;
716         CFS_LIST_HEAD(cancels);
717         int rc, count;
718         ENTRY;
719
720         if (!oa) {
721                 CDEBUG(D_INFO, "oa NULL\n");
722                 RETURN(-EINVAL);
723         }
724
725         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
726                                         LDLM_FL_DISCARD_DATA);
727
728         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
729         if (req == NULL) {
730                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
731                 RETURN(-ENOMEM);
732         }
733
734         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
735         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
736                                0, &cancels, count);
737         if (rc) {
738                 ptlrpc_request_free(req);
739                 RETURN(rc);
740         }
741
742         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
743         ptlrpc_at_set_req_timeout(req);
744
745         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
746                 oa->o_lcookie = *oti->oti_logcookies;
747         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
748         LASSERT(body);
749         lustre_set_wire_obdo(&body->oa, oa);
750
751         osc_pack_capa(req, body, (struct obd_capa *)capa);
752         ptlrpc_request_set_replen(req);
753
754         /* don't throttle destroy RPCs for the MDT */
755         if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
756                 req->rq_interpret_reply = osc_destroy_interpret;
757                 if (!osc_can_send_destroy(cli)) {
758                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
759                                                           NULL);
760
761                         /*
762                          * Wait until the number of on-going destroy RPCs drops
763                          * under max_rpc_in_flight
764                          */
765                         l_wait_event_exclusive(cli->cl_destroy_waitq,
766                                                osc_can_send_destroy(cli), &lwi);
767                 }
768         }
769
770         /* Do not wait for response */
771         ptlrpcd_add_req(req, PSCOPE_OTHER);
772         RETURN(0);
773 }
774
775 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
776                                 long writing_bytes)
777 {
778         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
779
780         LASSERT(!(oa->o_valid & bits));
781
782         oa->o_valid |= bits;
783         client_obd_list_lock(&cli->cl_loi_list_lock);
784         oa->o_dirty = cli->cl_dirty;
785         if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
786                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
787                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
788                 oa->o_undirty = 0;
789         } else if (cfs_atomic_read(&obd_dirty_pages) -
790                    cfs_atomic_read(&obd_dirty_transit_pages) >
791                    obd_max_dirty_pages + 1){
792                 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
793                  * not covered by a lock thus they may safely race and trip
794                  * this CERROR() unless we add in a small fudge factor (+1). */
795                 CERROR("dirty %d - %d > system dirty_max %d\n",
796                        cfs_atomic_read(&obd_dirty_pages),
797                        cfs_atomic_read(&obd_dirty_transit_pages),
798                        obd_max_dirty_pages);
799                 oa->o_undirty = 0;
800         } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
801                 CERROR("dirty %lu - dirty_max %lu too big???\n",
802                        cli->cl_dirty, cli->cl_dirty_max);
803                 oa->o_undirty = 0;
804         } else {
805                 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
806                                 (cli->cl_max_rpcs_in_flight + 1);
807                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
808         }
809         oa->o_grant = cli->cl_avail_grant;
810         oa->o_dropped = cli->cl_lost_grant;
811         cli->cl_lost_grant = 0;
812         client_obd_list_unlock(&cli->cl_loi_list_lock);
813         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
814                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
815
816 }
817
818 static void osc_update_next_shrink(struct client_obd *cli)
819 {
820         cli->cl_next_shrink_grant =
821                 cfs_time_shift(cli->cl_grant_shrink_interval);
822         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
823                cli->cl_next_shrink_grant);
824 }
825
826 /* caller must hold loi_list_lock */
827 static void osc_consume_write_grant(struct client_obd *cli,
828                                     struct brw_page *pga)
829 {
830         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
831         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
832         cfs_atomic_inc(&obd_dirty_pages);
833         cli->cl_dirty += CFS_PAGE_SIZE;
834         cli->cl_avail_grant -= CFS_PAGE_SIZE;
835         pga->flag |= OBD_BRW_FROM_GRANT;
836         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
837                CFS_PAGE_SIZE, pga, pga->pg);
838         LASSERT(cli->cl_avail_grant >= 0);
839         osc_update_next_shrink(cli);
840 }
841
842 /* the companion to osc_consume_write_grant, called when a brw has completed.
843  * must be called with the loi lock held. */
844 static void osc_release_write_grant(struct client_obd *cli,
845                                     struct brw_page *pga, int sent)
846 {
847         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
848         ENTRY;
849
850         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
851         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
852                 EXIT;
853                 return;
854         }
855
856         pga->flag &= ~OBD_BRW_FROM_GRANT;
857         cfs_atomic_dec(&obd_dirty_pages);
858         cli->cl_dirty -= CFS_PAGE_SIZE;
859         if (pga->flag & OBD_BRW_NOCACHE) {
860                 pga->flag &= ~OBD_BRW_NOCACHE;
861                 cfs_atomic_dec(&obd_dirty_transit_pages);
862                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
863         }
864         if (!sent) {
865                 cli->cl_lost_grant += CFS_PAGE_SIZE;
866                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
867                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
868         } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
869                 /* For short writes we shouldn't count parts of pages that
870                  * span a whole block on the OST side, or our accounting goes
871                  * wrong.  Should match the code in filter_grant_check. */
872                 int offset = pga->off & ~CFS_PAGE_MASK;
873                 int count = pga->count + (offset & (blocksize - 1));
874                 int end = (offset + pga->count) & (blocksize - 1);
875                 if (end)
876                         count += blocksize - end;
877
878                 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
879                 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
880                        CFS_PAGE_SIZE - count, cli->cl_lost_grant,
881                        cli->cl_avail_grant, cli->cl_dirty);
882         }
883
884         EXIT;
885 }
886
887 static unsigned long rpcs_in_flight(struct client_obd *cli)
888 {
889         return cli->cl_r_in_flight + cli->cl_w_in_flight;
890 }
891
892 /* caller must hold loi_list_lock */
893 void osc_wake_cache_waiters(struct client_obd *cli)
894 {
895         cfs_list_t *l, *tmp;
896         struct osc_cache_waiter *ocw;
897
898         ENTRY;
899         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
900                 /* if we can't dirty more, we must wait until some is written */
901                 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
902                    (cfs_atomic_read(&obd_dirty_pages) + 1 >
903                     obd_max_dirty_pages)) {
904                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
905                                "osc max %ld, sys max %d\n", cli->cl_dirty,
906                                cli->cl_dirty_max, obd_max_dirty_pages);
907                         return;
908                 }
909
910                 /* if still dirty cache but no grant wait for pending RPCs that
911                  * may yet return us some grant before doing sync writes */
912                 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
913                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
914                                cli->cl_w_in_flight);
915                         return;
916                 }
917
918                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
919                 cfs_list_del_init(&ocw->ocw_entry);
920                 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
921                         /* no more RPCs in flight to return grant, do sync IO */
922                         ocw->ocw_rc = -EDQUOT;
923                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
924                 } else {
925                         osc_consume_write_grant(cli,
926                                                 &ocw->ocw_oap->oap_brw_page);
927                 }
928
929                 cfs_waitq_signal(&ocw->ocw_waitq);
930         }
931
932         EXIT;
933 }
934
935 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
936 {
937         client_obd_list_lock(&cli->cl_loi_list_lock);
938         cli->cl_avail_grant += grant;
939         client_obd_list_unlock(&cli->cl_loi_list_lock);
940 }
941
942 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
943 {
944         if (body->oa.o_valid & OBD_MD_FLGRANT) {
945                 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
946                 __osc_update_grant(cli, body->oa.o_grant);
947         }
948 }
949
950 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
951                               void *key, obd_count vallen, void *val,
952                               struct ptlrpc_request_set *set);
953
954 static int osc_shrink_grant_interpret(const struct lu_env *env,
955                                       struct ptlrpc_request *req,
956                                       void *aa, int rc)
957 {
958         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
959         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
960         struct ost_body *body;
961
962         if (rc != 0) {
963                 __osc_update_grant(cli, oa->o_grant);
964                 GOTO(out, rc);
965         }
966
967         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
968         LASSERT(body);
969         osc_update_grant(cli, body);
970 out:
971         OBDO_FREE(oa);
972         return rc;
973 }
974
975 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
976 {
977         client_obd_list_lock(&cli->cl_loi_list_lock);
978         oa->o_grant = cli->cl_avail_grant / 4;
979         cli->cl_avail_grant -= oa->o_grant;
980         client_obd_list_unlock(&cli->cl_loi_list_lock);
981         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
982                 oa->o_valid |= OBD_MD_FLFLAGS;
983                 oa->o_flags = 0;
984         }
985         oa->o_flags |= OBD_FL_SHRINK_GRANT;
986         osc_update_next_shrink(cli);
987 }
988
989 /* Shrink the current grant, either from some large amount to enough for a
990  * full set of in-flight RPCs, or if we have already shrunk to that limit
991  * then to enough for a single RPC.  This avoids keeping more grant than
992  * needed, and avoids shrinking the grant piecemeal. */
993 static int osc_shrink_grant(struct client_obd *cli)
994 {
995         long target = (cli->cl_max_rpcs_in_flight + 1) *
996                       cli->cl_max_pages_per_rpc;
997
998         client_obd_list_lock(&cli->cl_loi_list_lock);
999         if (cli->cl_avail_grant <= target)
1000                 target = cli->cl_max_pages_per_rpc;
1001         client_obd_list_unlock(&cli->cl_loi_list_lock);
1002
1003         return osc_shrink_grant_to_target(cli, target);
1004 }
1005
1006 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
1007 {
1008         int    rc = 0;
1009         struct ost_body     *body;
1010         ENTRY;
1011
1012         client_obd_list_lock(&cli->cl_loi_list_lock);
1013         /* Don't shrink if we are already above or below the desired limit
1014          * We don't want to shrink below a single RPC, as that will negatively
1015          * impact block allocation and long-term performance. */
1016         if (target < cli->cl_max_pages_per_rpc)
1017                 target = cli->cl_max_pages_per_rpc;
1018
1019         if (target >= cli->cl_avail_grant) {
1020                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1021                 RETURN(0);
1022         }
1023         client_obd_list_unlock(&cli->cl_loi_list_lock);
1024
1025         OBD_ALLOC_PTR(body);
1026         if (!body)
1027                 RETURN(-ENOMEM);
1028
1029         osc_announce_cached(cli, &body->oa, 0);
1030
1031         client_obd_list_lock(&cli->cl_loi_list_lock);
1032         body->oa.o_grant = cli->cl_avail_grant - target;
1033         cli->cl_avail_grant = target;
1034         client_obd_list_unlock(&cli->cl_loi_list_lock);
1035         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
1036                 body->oa.o_valid |= OBD_MD_FLFLAGS;
1037                 body->oa.o_flags = 0;
1038         }
1039         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1040         osc_update_next_shrink(cli);
1041
1042         rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1043                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1044                                 sizeof(*body), body, NULL);
1045         if (rc != 0)
1046                 __osc_update_grant(cli, body->oa.o_grant);
1047         OBD_FREE_PTR(body);
1048         RETURN(rc);
1049 }
1050
1051 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1052 static int osc_should_shrink_grant(struct client_obd *client)
1053 {
1054         cfs_time_t time = cfs_time_current();
1055         cfs_time_t next_shrink = client->cl_next_shrink_grant;
1056
1057         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1058              OBD_CONNECT_GRANT_SHRINK) == 0)
1059                 return 0;
1060
1061         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1062                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1063                     client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1064                         return 1;
1065                 else
1066                         osc_update_next_shrink(client);
1067         }
1068         return 0;
1069 }
1070
1071 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1072 {
1073         struct client_obd *client;
1074
1075         cfs_list_for_each_entry(client, &item->ti_obd_list,
1076                                 cl_grant_shrink_list) {
1077                 if (osc_should_shrink_grant(client))
1078                         osc_shrink_grant(client);
1079         }
1080         return 0;
1081 }
1082
1083 static int osc_add_shrink_grant(struct client_obd *client)
1084 {
1085         int rc;
1086
1087         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1088                                        TIMEOUT_GRANT,
1089                                        osc_grant_shrink_grant_cb, NULL,
1090                                        &client->cl_grant_shrink_list);
1091         if (rc) {
1092                 CERROR("add grant client %s error %d\n",
1093                         client->cl_import->imp_obd->obd_name, rc);
1094                 return rc;
1095         }
1096         CDEBUG(D_CACHE, "add grant client %s \n",
1097                client->cl_import->imp_obd->obd_name);
1098         osc_update_next_shrink(client);
1099         return 0;
1100 }
1101
1102 static int osc_del_shrink_grant(struct client_obd *client)
1103 {
1104         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1105                                          TIMEOUT_GRANT);
1106 }
1107
1108 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1109 {
1110         /*
1111          * ocd_grant is the total grant amount we're expect to hold: if we've
1112          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1113          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1114          *
1115          * race is tolerable here: if we're evicted, but imp_state already
1116          * left EVICTED state, then cl_dirty must be 0 already.
1117          */
1118         client_obd_list_lock(&cli->cl_loi_list_lock);
1119         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1120                 cli->cl_avail_grant = ocd->ocd_grant;
1121         else
1122                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1123
1124         if (cli->cl_avail_grant < 0) {
1125                 CWARN("%s: available grant < 0, the OSS is probably not running"
1126                       " with patch from bug20278 (%ld) \n",
1127                       cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1128                 /* workaround for 1.6 servers which do not have
1129                  * the patch from bug20278 */
1130                 cli->cl_avail_grant = ocd->ocd_grant;
1131         }
1132
1133         client_obd_list_unlock(&cli->cl_loi_list_lock);
1134
1135         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1136                cli->cl_import->imp_obd->obd_name,
1137                cli->cl_avail_grant, cli->cl_lost_grant);
1138
1139         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1140             cfs_list_empty(&cli->cl_grant_shrink_list))
1141                 osc_add_shrink_grant(cli);
1142 }
1143
1144 /* We assume that the reason this OSC got a short read is because it read
1145  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1146  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1147  * this stripe never got written at or beyond this stripe offset yet. */
1148 static void handle_short_read(int nob_read, obd_count page_count,
1149                               struct brw_page **pga)
1150 {
1151         char *ptr;
1152         int i = 0;
1153
1154         /* skip bytes read OK */
1155         while (nob_read > 0) {
1156                 LASSERT (page_count > 0);
1157
1158                 if (pga[i]->count > nob_read) {
1159                         /* EOF inside this page */
1160                         ptr = cfs_kmap(pga[i]->pg) +
1161                                 (pga[i]->off & ~CFS_PAGE_MASK);
1162                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1163                         cfs_kunmap(pga[i]->pg);
1164                         page_count--;
1165                         i++;
1166                         break;
1167                 }
1168
1169                 nob_read -= pga[i]->count;
1170                 page_count--;
1171                 i++;
1172         }
1173
1174         /* zero remaining pages */
1175         while (page_count-- > 0) {
1176                 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1177                 memset(ptr, 0, pga[i]->count);
1178                 cfs_kunmap(pga[i]->pg);
1179                 i++;
1180         }
1181 }
1182
1183 static int check_write_rcs(struct ptlrpc_request *req,
1184                            int requested_nob, int niocount,
1185                            obd_count page_count, struct brw_page **pga)
1186 {
1187         int     i;
1188         __u32   *remote_rcs;
1189
1190         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1191                                                   sizeof(*remote_rcs) *
1192                                                   niocount);
1193         if (remote_rcs == NULL) {
1194                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1195                 return(-EPROTO);
1196         }
1197
1198         /* return error if any niobuf was in error */
1199         for (i = 0; i < niocount; i++) {
1200                 if ((int)remote_rcs[i] < 0)
1201                         return(remote_rcs[i]);
1202
1203                 if (remote_rcs[i] != 0) {
1204                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1205                                 i, remote_rcs[i], req);
1206                         return(-EPROTO);
1207                 }
1208         }
1209
1210         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1211                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1212                        req->rq_bulk->bd_nob_transferred, requested_nob);
1213                 return(-EPROTO);
1214         }
1215
1216         return (0);
1217 }
1218
1219 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1220 {
1221         if (p1->flag != p2->flag) {
1222                 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1223                                   OBD_BRW_NOCACHE|OBD_BRW_SYNC|OBD_BRW_ASYNC);
1224
1225                 /* warn if we try to combine flags that we don't know to be
1226                  * safe to combine */
1227                 if ((p1->flag & mask) != (p2->flag & mask))
1228                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
1229                                "same brw?\n", p1->flag, p2->flag);
1230                 return 0;
1231         }
1232
1233         return (p1->off + p1->count == p2->off);
1234 }
1235
1236 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1237                                    struct brw_page **pga, int opc,
1238                                    cksum_type_t cksum_type)
1239 {
1240         __u32 cksum;
1241         int i = 0;
1242
1243         LASSERT (pg_count > 0);
1244         cksum = init_checksum(cksum_type);
1245         while (nob > 0 && pg_count > 0) {
1246                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1247                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1248                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1249
1250                 /* corrupt the data before we compute the checksum, to
1251                  * simulate an OST->client data error */
1252                 if (i == 0 && opc == OST_READ &&
1253                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1254                         memcpy(ptr + off, "bad1", min(4, nob));
1255                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1256                 cfs_kunmap(pga[i]->pg);
1257                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1258                                off, cksum);
1259
1260                 nob -= pga[i]->count;
1261                 pg_count--;
1262                 i++;
1263         }
1264         /* For sending we only compute the wrong checksum instead
1265          * of corrupting the data so it is still correct on a redo */
1266         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1267                 cksum++;
1268
1269         return cksum;
1270 }
1271
1272 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1273                                 struct lov_stripe_md *lsm, obd_count page_count,
1274                                 struct brw_page **pga,
1275                                 struct ptlrpc_request **reqp,
1276                                 struct obd_capa *ocapa, int reserve,
1277                                 int resend)
1278 {
1279         struct ptlrpc_request   *req;
1280         struct ptlrpc_bulk_desc *desc;
1281         struct ost_body         *body;
1282         struct obd_ioobj        *ioobj;
1283         struct niobuf_remote    *niobuf;
1284         int niocount, i, requested_nob, opc, rc;
1285         struct osc_brw_async_args *aa;
1286         struct req_capsule      *pill;
1287         struct brw_page *pg_prev;
1288
1289         ENTRY;
1290         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1291                 RETURN(-ENOMEM); /* Recoverable */
1292         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1293                 RETURN(-EINVAL); /* Fatal */
1294
1295         if ((cmd & OBD_BRW_WRITE) != 0) {
1296                 opc = OST_WRITE;
1297                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1298                                                 cli->cl_import->imp_rq_pool,
1299                                                 &RQF_OST_BRW_WRITE);
1300         } else {
1301                 opc = OST_READ;
1302                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1303         }
1304         if (req == NULL)
1305                 RETURN(-ENOMEM);
1306
1307         for (niocount = i = 1; i < page_count; i++) {
1308                 if (!can_merge_pages(pga[i - 1], pga[i]))
1309                         niocount++;
1310         }
1311
1312         pill = &req->rq_pill;
1313         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1314                              sizeof(*ioobj));
1315         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1316                              niocount * sizeof(*niobuf));
1317         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1318
1319         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1320         if (rc) {
1321                 ptlrpc_request_free(req);
1322                 RETURN(rc);
1323         }
1324         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1325         ptlrpc_at_set_req_timeout(req);
1326
1327         if (opc == OST_WRITE)
1328                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1329                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1330         else
1331                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1332                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1333
1334         if (desc == NULL)
1335                 GOTO(out, rc = -ENOMEM);
1336         /* NB request now owns desc and will free it when it gets freed */
1337
1338         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1339         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1340         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1341         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1342
1343         lustre_set_wire_obdo(&body->oa, oa);
1344
1345         obdo_to_ioobj(oa, ioobj);
1346         ioobj->ioo_bufcnt = niocount;
1347         osc_pack_capa(req, body, ocapa);
1348         LASSERT (page_count > 0);
1349         pg_prev = pga[0];
1350         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1351                 struct brw_page *pg = pga[i];
1352                 int poff = pg->off & ~CFS_PAGE_MASK;
1353
1354                 LASSERT(pg->count > 0);
1355                 /* make sure there is no gap in the middle of page array */
1356                 LASSERTF(page_count == 1 ||
1357                          (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1358                           ergo(i > 0 && i < page_count - 1,
1359                                poff == 0 && pg->count == CFS_PAGE_SIZE)   &&
1360                           ergo(i == page_count - 1, poff == 0)),
1361                          "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1362                          i, page_count, pg, pg->off, pg->count);
1363 #ifdef __linux__
1364                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1365                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1366                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1367                          i, page_count,
1368                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1369                          pg_prev->pg, page_private(pg_prev->pg),
1370                          pg_prev->pg->index, pg_prev->off);
1371 #else
1372                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1373                          "i %d p_c %u\n", i, page_count);
1374 #endif
1375                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1376                         (pg->flag & OBD_BRW_SRVLOCK));
1377
1378                 ptlrpc_prep_bulk_page(desc, pg->pg, poff, pg->count);
1379                 requested_nob += pg->count;
1380
1381                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1382                         niobuf--;
1383                         niobuf->len += pg->count;
1384                 } else {
1385                         niobuf->offset = pg->off;
1386                         niobuf->len    = pg->count;
1387                         niobuf->flags  = pg->flag;
1388                 }
1389                 pg_prev = pg;
1390         }
1391
1392         LASSERTF((void *)(niobuf - niocount) ==
1393                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1394                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1395                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1396
1397         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1398         if (resend) {
1399                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1400                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1401                         body->oa.o_flags = 0;
1402                 }
1403                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1404         }
1405
1406         if (osc_should_shrink_grant(cli))
1407                 osc_shrink_grant_local(cli, &body->oa);
1408
1409         /* size[REQ_REC_OFF] still sizeof (*body) */
1410         if (opc == OST_WRITE) {
1411                 if (unlikely(cli->cl_checksum) &&
1412                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1413                         /* store cl_cksum_type in a local variable since
1414                          * it can be changed via lprocfs */
1415                         cksum_type_t cksum_type = cli->cl_cksum_type;
1416
1417                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1418                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1419                                 body->oa.o_flags = 0;
1420                         }
1421                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1422                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1423                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1424                                                              page_count, pga,
1425                                                              OST_WRITE,
1426                                                              cksum_type);
1427                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1428                                body->oa.o_cksum);
1429                         /* save this in 'oa', too, for later checking */
1430                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1431                         oa->o_flags |= cksum_type_pack(cksum_type);
1432                 } else {
1433                         /* clear out the checksum flag, in case this is a
1434                          * resend but cl_checksum is no longer set. b=11238 */
1435                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1436                 }
1437                 oa->o_cksum = body->oa.o_cksum;
1438                 /* 1 RC per niobuf */
1439                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1440                                      sizeof(__u32) * niocount);
1441         } else {
1442                 if (unlikely(cli->cl_checksum) &&
1443                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1444                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1445                                 body->oa.o_flags = 0;
1446                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1447                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1448                 }
1449         }
1450         ptlrpc_request_set_replen(req);
1451
1452         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1453         aa = ptlrpc_req_async_args(req);
1454         aa->aa_oa = oa;
1455         aa->aa_requested_nob = requested_nob;
1456         aa->aa_nio_count = niocount;
1457         aa->aa_page_count = page_count;
1458         aa->aa_resends = 0;
1459         aa->aa_ppga = pga;
1460         aa->aa_cli = cli;
1461         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1462         if (ocapa && reserve)
1463                 aa->aa_ocapa = capa_get(ocapa);
1464
1465         *reqp = req;
1466         RETURN(0);
1467
1468  out:
1469         ptlrpc_req_finished(req);
1470         RETURN(rc);
1471 }
1472
1473 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1474                                 __u32 client_cksum, __u32 server_cksum, int nob,
1475                                 obd_count page_count, struct brw_page **pga,
1476                                 cksum_type_t client_cksum_type)
1477 {
1478         __u32 new_cksum;
1479         char *msg;
1480         cksum_type_t cksum_type;
1481
1482         if (server_cksum == client_cksum) {
1483                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1484                 return 0;
1485         }
1486
1487         /* If this is mmaped file - it can be changed at any time */
1488         if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
1489                 return 1;
1490
1491         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1492                                        oa->o_flags : 0);
1493         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1494                                       cksum_type);
1495
1496         if (cksum_type != client_cksum_type)
1497                 msg = "the server did not use the checksum type specified in "
1498                       "the original request - likely a protocol problem";
1499         else if (new_cksum == server_cksum)
1500                 msg = "changed on the client after we checksummed it - "
1501                       "likely false positive due to mmap IO (bug 11742)";
1502         else if (new_cksum == client_cksum)
1503                 msg = "changed in transit before arrival at OST";
1504         else
1505                 msg = "changed in transit AND doesn't match the original - "
1506                       "likely false positive due to mmap IO (bug 11742)";
1507
1508         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1509                            " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1510                            msg, libcfs_nid2str(peer->nid),
1511                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1512                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1513                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1514                            oa->o_id,
1515                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1516                            pga[0]->off,
1517                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1518         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1519                "client csum now %x\n", client_cksum, client_cksum_type,
1520                server_cksum, cksum_type, new_cksum);
1521         return 1;
1522 }
1523
1524 /* Note rc enters this function as number of bytes transferred */
1525 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1526 {
1527         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1528         const lnet_process_id_t *peer =
1529                         &req->rq_import->imp_connection->c_peer;
1530         struct client_obd *cli = aa->aa_cli;
1531         struct ost_body *body;
1532         __u32 client_cksum = 0;
1533         ENTRY;
1534
1535         if (rc < 0 && rc != -EDQUOT) {
1536                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1537                 RETURN(rc);
1538         }
1539
1540         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1541         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1542         if (body == NULL) {
1543                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1544                 RETURN(-EPROTO);
1545         }
1546
1547 #ifdef HAVE_QUOTA_SUPPORT
1548         /* set/clear over quota flag for a uid/gid */
1549         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1550             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1551                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1552
1553                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1554                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1555                        body->oa.o_flags);
1556                 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1557                              body->oa.o_flags);
1558         }
1559 #endif
1560
1561         osc_update_grant(cli, body);
1562
1563         if (rc < 0)
1564                 RETURN(rc);
1565
1566         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1567                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1568
1569         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1570                 if (rc > 0) {
1571                         CERROR("Unexpected +ve rc %d\n", rc);
1572                         RETURN(-EPROTO);
1573                 }
1574                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1575
1576                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1577                         RETURN(-EAGAIN);
1578
1579                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1580                     check_write_checksum(&body->oa, peer, client_cksum,
1581                                          body->oa.o_cksum, aa->aa_requested_nob,
1582                                          aa->aa_page_count, aa->aa_ppga,
1583                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1584                         RETURN(-EAGAIN);
1585
1586                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1587                                      aa->aa_page_count, aa->aa_ppga);
1588                 GOTO(out, rc);
1589         }
1590
1591         /* The rest of this function executes only for OST_READs */
1592
1593         /* if unwrap_bulk failed, return -EAGAIN to retry */
1594         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1595         if (rc < 0)
1596                 GOTO(out, rc = -EAGAIN);
1597
1598         if (rc > aa->aa_requested_nob) {
1599                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1600                        aa->aa_requested_nob);
1601                 RETURN(-EPROTO);
1602         }
1603
1604         if (rc != req->rq_bulk->bd_nob_transferred) {
1605                 CERROR ("Unexpected rc %d (%d transferred)\n",
1606                         rc, req->rq_bulk->bd_nob_transferred);
1607                 return (-EPROTO);
1608         }
1609
1610         if (rc < aa->aa_requested_nob)
1611                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1612
1613         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1614                 static int cksum_counter;
1615                 __u32      server_cksum = body->oa.o_cksum;
1616                 char      *via;
1617                 char      *router;
1618                 cksum_type_t cksum_type;
1619
1620                 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1621                                                body->oa.o_flags : 0);
1622                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1623                                                  aa->aa_ppga, OST_READ,
1624                                                  cksum_type);
1625
1626                 if (peer->nid == req->rq_bulk->bd_sender) {
1627                         via = router = "";
1628                 } else {
1629                         via = " via ";
1630                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1631                 }
1632
1633                 if (server_cksum == ~0 && rc > 0) {
1634                         CERROR("Protocol error: server %s set the 'checksum' "
1635                                "bit, but didn't send a checksum.  Not fatal, "
1636                                "but please notify on http://bugs.whamcloud.com/\n",
1637                                libcfs_nid2str(peer->nid));
1638                 } else if (server_cksum != client_cksum) {
1639                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1640                                            "%s%s%s inode "DFID" object "
1641                                            LPU64"/"LPU64" extent "
1642                                            "["LPU64"-"LPU64"]\n",
1643                                            req->rq_import->imp_obd->obd_name,
1644                                            libcfs_nid2str(peer->nid),
1645                                            via, router,
1646                                            body->oa.o_valid & OBD_MD_FLFID ?
1647                                                 body->oa.o_parent_seq : (__u64)0,
1648                                            body->oa.o_valid & OBD_MD_FLFID ?
1649                                                 body->oa.o_parent_oid : 0,
1650                                            body->oa.o_valid & OBD_MD_FLFID ?
1651                                                 body->oa.o_parent_ver : 0,
1652                                            body->oa.o_id,
1653                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1654                                                 body->oa.o_seq : (__u64)0,
1655                                            aa->aa_ppga[0]->off,
1656                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1657                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1658                                                                         1);
1659                         CERROR("client %x, server %x, cksum_type %x\n",
1660                                client_cksum, server_cksum, cksum_type);
1661                         cksum_counter = 0;
1662                         aa->aa_oa->o_cksum = client_cksum;
1663                         rc = -EAGAIN;
1664                 } else {
1665                         cksum_counter++;
1666                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1667                         rc = 0;
1668                 }
1669         } else if (unlikely(client_cksum)) {
1670                 static int cksum_missed;
1671
1672                 cksum_missed++;
1673                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1674                         CERROR("Checksum %u requested from %s but not sent\n",
1675                                cksum_missed, libcfs_nid2str(peer->nid));
1676         } else {
1677                 rc = 0;
1678         }
1679 out:
1680         if (rc >= 0)
1681                 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1682
1683         RETURN(rc);
1684 }
1685
1686 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1687                             struct lov_stripe_md *lsm,
1688                             obd_count page_count, struct brw_page **pga,
1689                             struct obd_capa *ocapa)
1690 {
1691         struct ptlrpc_request *req;
1692         int                    rc;
1693         cfs_waitq_t            waitq;
1694         int                    resends = 0;
1695         struct l_wait_info     lwi;
1696
1697         ENTRY;
1698
1699         cfs_waitq_init(&waitq);
1700
1701 restart_bulk:
1702         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1703                                   page_count, pga, &req, ocapa, 0, resends);
1704         if (rc != 0)
1705                 return (rc);
1706
1707         rc = ptlrpc_queue_wait(req);
1708
1709         if (rc == -ETIMEDOUT && req->rq_resend) {
1710                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1711                 ptlrpc_req_finished(req);
1712                 goto restart_bulk;
1713         }
1714
1715         rc = osc_brw_fini_request(req, rc);
1716
1717         ptlrpc_req_finished(req);
1718         if (osc_recoverable_error(rc)) {
1719                 resends++;
1720                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1721                         CERROR("too many resend retries, returning error\n");
1722                         RETURN(-EIO);
1723                 }
1724
1725                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1726                 l_wait_event(waitq, 0, &lwi);
1727
1728                 goto restart_bulk;
1729         }
1730
1731         RETURN (rc);
1732 }
1733
1734 int osc_brw_redo_request(struct ptlrpc_request *request,
1735                          struct osc_brw_async_args *aa)
1736 {
1737         struct ptlrpc_request *new_req;
1738         struct ptlrpc_request_set *set = request->rq_set;
1739         struct osc_brw_async_args *new_aa;
1740         struct osc_async_page *oap;
1741         int rc = 0;
1742         ENTRY;
1743
1744         if (!client_should_resend(aa->aa_resends, aa->aa_cli)) {
1745                 CERROR("too many resent retries, returning error\n");
1746                 RETURN(-EIO);
1747         }
1748
1749         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1750
1751         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1752                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1753                                   aa->aa_cli, aa->aa_oa,
1754                                   NULL /* lsm unused by osc currently */,
1755                                   aa->aa_page_count, aa->aa_ppga,
1756                                   &new_req, aa->aa_ocapa, 0, 1);
1757         if (rc)
1758                 RETURN(rc);
1759
1760         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1761
1762         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1763                 if (oap->oap_request != NULL) {
1764                         LASSERTF(request == oap->oap_request,
1765                                  "request %p != oap_request %p\n",
1766                                  request, oap->oap_request);
1767                         if (oap->oap_interrupted) {
1768                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1769                                 ptlrpc_req_finished(new_req);
1770                                 RETURN(-EINTR);
1771                         }
1772                 }
1773         }
1774         /* New request takes over pga and oaps from old request.
1775          * Note that copying a list_head doesn't work, need to move it... */
1776         aa->aa_resends++;
1777         new_req->rq_interpret_reply = request->rq_interpret_reply;
1778         new_req->rq_async_args = request->rq_async_args;
1779         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1780
1781         new_aa = ptlrpc_req_async_args(new_req);
1782
1783         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1784         cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1785         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1786
1787         cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1788                 if (oap->oap_request) {
1789                         ptlrpc_req_finished(oap->oap_request);
1790                         oap->oap_request = ptlrpc_request_addref(new_req);
1791                 }
1792         }
1793
1794         new_aa->aa_ocapa = aa->aa_ocapa;
1795         aa->aa_ocapa = NULL;
1796
1797         /* use ptlrpc_set_add_req is safe because interpret functions work
1798          * in check_set context. only one way exist with access to request
1799          * from different thread got -EINTR - this way protected with
1800          * cl_loi_list_lock */
1801         ptlrpc_set_add_req(set, new_req);
1802
1803         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1804
1805         DEBUG_REQ(D_INFO, new_req, "new request");
1806         RETURN(0);
1807 }
1808
1809 /*
1810  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1811  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1812  * fine for our small page arrays and doesn't require allocation.  its an
1813  * insertion sort that swaps elements that are strides apart, shrinking the
1814  * stride down until its '1' and the array is sorted.
1815  */
1816 static void sort_brw_pages(struct brw_page **array, int num)
1817 {
1818         int stride, i, j;
1819         struct brw_page *tmp;
1820
1821         if (num == 1)
1822                 return;
1823         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1824                 ;
1825
1826         do {
1827                 stride /= 3;
1828                 for (i = stride ; i < num ; i++) {
1829                         tmp = array[i];
1830                         j = i;
1831                         while (j >= stride && array[j - stride]->off > tmp->off) {
1832                                 array[j] = array[j - stride];
1833                                 j -= stride;
1834                         }
1835                         array[j] = tmp;
1836                 }
1837         } while (stride > 1);
1838 }
1839
1840 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1841 {
1842         int count = 1;
1843         int offset;
1844         int i = 0;
1845
1846         LASSERT (pages > 0);
1847         offset = pg[i]->off & ~CFS_PAGE_MASK;
1848
1849         for (;;) {
1850                 pages--;
1851                 if (pages == 0)         /* that's all */
1852                         return count;
1853
1854                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1855                         return count;   /* doesn't end on page boundary */
1856
1857                 i++;
1858                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1859                 if (offset != 0)        /* doesn't start on page boundary */
1860                         return count;
1861
1862                 count++;
1863         }
1864 }
1865
1866 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1867 {
1868         struct brw_page **ppga;
1869         int i;
1870
1871         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1872         if (ppga == NULL)
1873                 return NULL;
1874
1875         for (i = 0; i < count; i++)
1876                 ppga[i] = pga + i;
1877         return ppga;
1878 }
1879
1880 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1881 {
1882         LASSERT(ppga != NULL);
1883         OBD_FREE(ppga, sizeof(*ppga) * count);
1884 }
1885
1886 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1887                    obd_count page_count, struct brw_page *pga,
1888                    struct obd_trans_info *oti)
1889 {
1890         struct obdo *saved_oa = NULL;
1891         struct brw_page **ppga, **orig;
1892         struct obd_import *imp = class_exp2cliimp(exp);
1893         struct client_obd *cli;
1894         int rc, page_count_orig;
1895         ENTRY;
1896
1897         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1898         cli = &imp->imp_obd->u.cli;
1899
1900         if (cmd & OBD_BRW_CHECK) {
1901                 /* The caller just wants to know if there's a chance that this
1902                  * I/O can succeed */
1903
1904                 if (imp->imp_invalid)
1905                         RETURN(-EIO);
1906                 RETURN(0);
1907         }
1908
1909         /* test_brw with a failed create can trip this, maybe others. */
1910         LASSERT(cli->cl_max_pages_per_rpc);
1911
1912         rc = 0;
1913
1914         orig = ppga = osc_build_ppga(pga, page_count);
1915         if (ppga == NULL)
1916                 RETURN(-ENOMEM);
1917         page_count_orig = page_count;
1918
1919         sort_brw_pages(ppga, page_count);
1920         while (page_count) {
1921                 obd_count pages_per_brw;
1922
1923                 if (page_count > cli->cl_max_pages_per_rpc)
1924                         pages_per_brw = cli->cl_max_pages_per_rpc;
1925                 else
1926                         pages_per_brw = page_count;
1927
1928                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1929
1930                 if (saved_oa != NULL) {
1931                         /* restore previously saved oa */
1932                         *oinfo->oi_oa = *saved_oa;
1933                 } else if (page_count > pages_per_brw) {
1934                         /* save a copy of oa (brw will clobber it) */
1935                         OBDO_ALLOC(saved_oa);
1936                         if (saved_oa == NULL)
1937                                 GOTO(out, rc = -ENOMEM);
1938                         *saved_oa = *oinfo->oi_oa;
1939                 }
1940
1941                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1942                                       pages_per_brw, ppga, oinfo->oi_capa);
1943
1944                 if (rc != 0)
1945                         break;
1946
1947                 page_count -= pages_per_brw;
1948                 ppga += pages_per_brw;
1949         }
1950
1951 out:
1952         osc_release_ppga(orig, page_count_orig);
1953
1954         if (saved_oa != NULL)
1955                 OBDO_FREE(saved_oa);
1956
1957         RETURN(rc);
1958 }
1959
1960 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1961  * the dirty accounting.  Writeback completes or truncate happens before
1962  * writing starts.  Must be called with the loi lock held. */
1963 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1964                            int sent)
1965 {
1966         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1967 }
1968
1969
1970 /* This maintains the lists of pending pages to read/write for a given object
1971  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1972  * to quickly find objects that are ready to send an RPC. */
1973 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1974                          int cmd)
1975 {
1976         int optimal;
1977         ENTRY;
1978
1979         if (lop->lop_num_pending == 0)
1980                 RETURN(0);
1981
1982         /* if we have an invalid import we want to drain the queued pages
1983          * by forcing them through rpcs that immediately fail and complete
1984          * the pages.  recovery relies on this to empty the queued pages
1985          * before canceling the locks and evicting down the llite pages */
1986         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1987                 RETURN(1);
1988
1989         /* stream rpcs in queue order as long as as there is an urgent page
1990          * queued.  this is our cheap solution for good batching in the case
1991          * where writepage marks some random page in the middle of the file
1992          * as urgent because of, say, memory pressure */
1993         if (!cfs_list_empty(&lop->lop_urgent)) {
1994                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1995                 RETURN(1);
1996         }
1997         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1998         optimal = cli->cl_max_pages_per_rpc;
1999         if (cmd & OBD_BRW_WRITE) {
2000                 /* trigger a write rpc stream as long as there are dirtiers
2001                  * waiting for space.  as they're waiting, they're not going to
2002                  * create more pages to coalesce with what's waiting.. */
2003                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
2004                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
2005                         RETURN(1);
2006                 }
2007                 /* +16 to avoid triggering rpcs that would want to include pages
2008                  * that are being queued but which can't be made ready until
2009                  * the queuer finishes with the page. this is a wart for
2010                  * llite::commit_write() */
2011                 optimal += 16;
2012         }
2013         if (lop->lop_num_pending >= optimal)
2014                 RETURN(1);
2015
2016         RETURN(0);
2017 }
2018
2019 static int lop_makes_hprpc(struct loi_oap_pages *lop)
2020 {
2021         struct osc_async_page *oap;
2022         ENTRY;
2023
2024         if (cfs_list_empty(&lop->lop_urgent))
2025                 RETURN(0);
2026
2027         oap = cfs_list_entry(lop->lop_urgent.next,
2028                          struct osc_async_page, oap_urgent_item);
2029
2030         if (oap->oap_async_flags & ASYNC_HP) {
2031                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
2032                 RETURN(1);
2033         }
2034
2035         RETURN(0);
2036 }
2037
2038 static void on_list(cfs_list_t *item, cfs_list_t *list,
2039                     int should_be_on)
2040 {
2041         if (cfs_list_empty(item) && should_be_on)
2042                 cfs_list_add_tail(item, list);
2043         else if (!cfs_list_empty(item) && !should_be_on)
2044                 cfs_list_del_init(item);
2045 }
2046
2047 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
2048  * can find pages to build into rpcs quickly */
2049 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
2050 {
2051         if (lop_makes_hprpc(&loi->loi_write_lop) ||
2052             lop_makes_hprpc(&loi->loi_read_lop)) {
2053                 /* HP rpc */
2054                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2055                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2056         } else {
2057                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2058                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2059                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2060                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2061         }
2062
2063         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2064                 loi->loi_write_lop.lop_num_pending);
2065
2066         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2067                 loi->loi_read_lop.lop_num_pending);
2068 }
2069
2070 static void lop_update_pending(struct client_obd *cli,
2071                                struct loi_oap_pages *lop, int cmd, int delta)
2072 {
2073         lop->lop_num_pending += delta;
2074         if (cmd & OBD_BRW_WRITE)
2075                 cli->cl_pending_w_pages += delta;
2076         else
2077                 cli->cl_pending_r_pages += delta;
2078 }
2079
2080 /**
2081  * this is called when a sync waiter receives an interruption.  Its job is to
2082  * get the caller woken as soon as possible.  If its page hasn't been put in an
2083  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2084  * desiring interruption which will forcefully complete the rpc once the rpc
2085  * has timed out.
2086  */
2087 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2088 {
2089         struct loi_oap_pages *lop;
2090         struct lov_oinfo *loi;
2091         int rc = -EBUSY;
2092         ENTRY;
2093
2094         LASSERT(!oap->oap_interrupted);
2095         oap->oap_interrupted = 1;
2096
2097         /* ok, it's been put in an rpc. only one oap gets a request reference */
2098         if (oap->oap_request != NULL) {
2099                 ptlrpc_mark_interrupted(oap->oap_request);
2100                 ptlrpcd_wake(oap->oap_request);
2101                 ptlrpc_req_finished(oap->oap_request);
2102                 oap->oap_request = NULL;
2103         }
2104
2105         /*
2106          * page completion may be called only if ->cpo_prep() method was
2107          * executed by osc_io_submit(), that also adds page the to pending list
2108          */
2109         if (!cfs_list_empty(&oap->oap_pending_item)) {
2110                 cfs_list_del_init(&oap->oap_pending_item);
2111                 cfs_list_del_init(&oap->oap_urgent_item);
2112
2113                 loi = oap->oap_loi;
2114                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2115                         &loi->loi_write_lop : &loi->loi_read_lop;
2116                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2117                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2118                 rc = oap->oap_caller_ops->ap_completion(env,
2119                                           oap->oap_caller_data,
2120                                           oap->oap_cmd, NULL, -EINTR);
2121         }
2122
2123         RETURN(rc);
2124 }
2125
2126 /* this is trying to propogate async writeback errors back up to the
2127  * application.  As an async write fails we record the error code for later if
2128  * the app does an fsync.  As long as errors persist we force future rpcs to be
2129  * sync so that the app can get a sync error and break the cycle of queueing
2130  * pages for which writeback will fail. */
2131 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2132                            int rc)
2133 {
2134         if (rc) {
2135                 if (!ar->ar_rc)
2136                         ar->ar_rc = rc;
2137
2138                 ar->ar_force_sync = 1;
2139                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2140                 return;
2141
2142         }
2143
2144         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2145                 ar->ar_force_sync = 0;
2146 }
2147
2148 void osc_oap_to_pending(struct osc_async_page *oap)
2149 {
2150         struct loi_oap_pages *lop;
2151
2152         if (oap->oap_cmd & OBD_BRW_WRITE)
2153                 lop = &oap->oap_loi->loi_write_lop;
2154         else
2155                 lop = &oap->oap_loi->loi_read_lop;
2156
2157         if (oap->oap_async_flags & ASYNC_HP)
2158                 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2159         else if (oap->oap_async_flags & ASYNC_URGENT)
2160                 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2161         cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2162         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2163 }
2164
2165 /* this must be called holding the loi list lock to give coverage to exit_cache,
2166  * async_flag maintenance, and oap_request */
2167 static void osc_ap_completion(const struct lu_env *env,
2168                               struct client_obd *cli, struct obdo *oa,
2169                               struct osc_async_page *oap, int sent, int rc)
2170 {
2171         __u64 xid = 0;
2172
2173         ENTRY;
2174         if (oap->oap_request != NULL) {
2175                 xid = ptlrpc_req_xid(oap->oap_request);
2176                 ptlrpc_req_finished(oap->oap_request);
2177                 oap->oap_request = NULL;
2178         }
2179
2180         cfs_spin_lock(&oap->oap_lock);
2181         oap->oap_async_flags = 0;
2182         cfs_spin_unlock(&oap->oap_lock);
2183         oap->oap_interrupted = 0;
2184
2185         if (oap->oap_cmd & OBD_BRW_WRITE) {
2186                 osc_process_ar(&cli->cl_ar, xid, rc);
2187                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2188         }
2189
2190         if (rc == 0 && oa != NULL) {
2191                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2192                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2193                 if (oa->o_valid & OBD_MD_FLMTIME)
2194                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2195                 if (oa->o_valid & OBD_MD_FLATIME)
2196                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2197                 if (oa->o_valid & OBD_MD_FLCTIME)
2198                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2199         }
2200
2201         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2202                                                 oap->oap_cmd, oa, rc);
2203
2204         /* ll_ap_completion (from llite) drops PG_locked. so, a new
2205          * I/O on the page could start, but OSC calls it under lock
2206          * and thus we can add oap back to pending safely */
2207         if (rc)
2208                 /* upper layer wants to leave the page on pending queue */
2209                 osc_oap_to_pending(oap);
2210         else
2211                 osc_exit_cache(cli, oap, sent);
2212         EXIT;
2213 }
2214
2215 static int brw_interpret(const struct lu_env *env,
2216                          struct ptlrpc_request *req, void *data, int rc)
2217 {
2218         struct osc_brw_async_args *aa = data;
2219         struct client_obd *cli;
2220         int async;
2221         ENTRY;
2222
2223         rc = osc_brw_fini_request(req, rc);
2224         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2225         if (osc_recoverable_error(rc)) {
2226                 /* Only retry once for mmaped files since the mmaped page
2227                  * might be modified at anytime. We have to retry at least
2228                  * once in case there WAS really a corruption of the page
2229                  * on the network, that was not caused by mmap() modifying
2230                  * the page. Bug11742 */
2231                 if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
2232                     aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
2233                     aa->aa_oa->o_flags & OBD_FL_MMAP) {
2234                         rc = 0;
2235                 } else {
2236                         rc = osc_brw_redo_request(req, aa);
2237                         if (rc == 0)
2238                                 RETURN(0);
2239                 }
2240         }
2241
2242         if (aa->aa_ocapa) {
2243                 capa_put(aa->aa_ocapa);
2244                 aa->aa_ocapa = NULL;
2245         }
2246
2247         cli = aa->aa_cli;
2248
2249         client_obd_list_lock(&cli->cl_loi_list_lock);
2250
2251         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2252          * is called so we know whether to go to sync BRWs or wait for more
2253          * RPCs to complete */
2254         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2255                 cli->cl_w_in_flight--;
2256         else
2257                 cli->cl_r_in_flight--;
2258
2259         async = cfs_list_empty(&aa->aa_oaps);
2260         if (!async) { /* from osc_send_oap_rpc() */
2261                 struct osc_async_page *oap, *tmp;
2262                 /* the caller may re-use the oap after the completion call so
2263                  * we need to clean it up a little */
2264                 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2265                                              oap_rpc_item) {
2266                         cfs_list_del_init(&oap->oap_rpc_item);
2267                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2268                 }
2269                 OBDO_FREE(aa->aa_oa);
2270         } else { /* from async_internal() */
2271                 obd_count i;
2272                 for (i = 0; i < aa->aa_page_count; i++)
2273                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2274         }
2275         osc_wake_cache_waiters(cli);
2276         osc_check_rpcs(env, cli);
2277         client_obd_list_unlock(&cli->cl_loi_list_lock);
2278         if (!async)
2279                 cl_req_completion(env, aa->aa_clerq, rc);
2280         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2281
2282         RETURN(rc);
2283 }
2284
2285 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2286                                             struct client_obd *cli,
2287                                             cfs_list_t *rpc_list,
2288                                             int page_count, int cmd)
2289 {
2290         struct ptlrpc_request *req;
2291         struct brw_page **pga = NULL;
2292         struct osc_brw_async_args *aa;
2293         struct obdo *oa = NULL;
2294         const struct obd_async_page_ops *ops = NULL;
2295         void *caller_data = NULL;
2296         struct osc_async_page *oap;
2297         struct osc_async_page *tmp;
2298         struct ost_body *body;
2299         struct cl_req *clerq = NULL;
2300         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2301         struct ldlm_lock *lock = NULL;
2302         struct cl_req_attr crattr;
2303         int i, rc, mpflag = 0;
2304
2305         ENTRY;
2306         LASSERT(!cfs_list_empty(rpc_list));
2307
2308         if (cmd & OBD_BRW_MEMALLOC)
2309                 mpflag = cfs_memory_pressure_get_and_set();
2310
2311         memset(&crattr, 0, sizeof crattr);
2312         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2313         if (pga == NULL)
2314                 GOTO(out, req = ERR_PTR(-ENOMEM));
2315
2316         OBDO_ALLOC(oa);
2317         if (oa == NULL)
2318                 GOTO(out, req = ERR_PTR(-ENOMEM));
2319
2320         i = 0;
2321         cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2322                 struct cl_page *page = osc_oap2cl_page(oap);
2323                 if (ops == NULL) {
2324                         ops = oap->oap_caller_ops;
2325                         caller_data = oap->oap_caller_data;
2326
2327                         clerq = cl_req_alloc(env, page, crt,
2328                                              1 /* only 1-object rpcs for
2329                                                 * now */);
2330                         if (IS_ERR(clerq))
2331                                 GOTO(out, req = (void *)clerq);
2332                         lock = oap->oap_ldlm_lock;
2333                 }
2334                 pga[i] = &oap->oap_brw_page;
2335                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2336                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2337                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2338                 i++;
2339                 cl_req_page_add(env, clerq, page);
2340         }
2341
2342         /* always get the data for the obdo for the rpc */
2343         LASSERT(ops != NULL);
2344         crattr.cra_oa = oa;
2345         crattr.cra_capa = NULL;
2346         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2347         if (lock) {
2348                 oa->o_handle = lock->l_remote_handle;
2349                 oa->o_valid |= OBD_MD_FLHANDLE;
2350         }
2351
2352         rc = cl_req_prep(env, clerq);
2353         if (rc != 0) {
2354                 CERROR("cl_req_prep failed: %d\n", rc);
2355                 GOTO(out, req = ERR_PTR(rc));
2356         }
2357
2358         sort_brw_pages(pga, page_count);
2359         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2360                                   pga, &req, crattr.cra_capa, 1, 0);
2361         if (rc != 0) {
2362                 CERROR("prep_req failed: %d\n", rc);
2363                 GOTO(out, req = ERR_PTR(rc));
2364         }
2365
2366         if (cmd & OBD_BRW_MEMALLOC)
2367                 req->rq_memalloc = 1;
2368
2369         /* Need to update the timestamps after the request is built in case
2370          * we race with setattr (locally or in queue at OST).  If OST gets
2371          * later setattr before earlier BRW (as determined by the request xid),
2372          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2373          * way to do this in a single call.  bug 10150 */
2374         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2375         cl_req_attr_set(env, clerq, &crattr,
2376                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2377
2378         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2379         aa = ptlrpc_req_async_args(req);
2380         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2381         cfs_list_splice(rpc_list, &aa->aa_oaps);
2382         CFS_INIT_LIST_HEAD(rpc_list);
2383         aa->aa_clerq = clerq;
2384 out:
2385         if (cmd & OBD_BRW_MEMALLOC)
2386                 cfs_memory_pressure_restore(mpflag);
2387
2388         capa_put(crattr.cra_capa);
2389         if (IS_ERR(req)) {
2390                 if (oa)
2391                         OBDO_FREE(oa);
2392                 if (pga)
2393                         OBD_FREE(pga, sizeof(*pga) * page_count);
2394                 /* this should happen rarely and is pretty bad, it makes the
2395                  * pending list not follow the dirty order */
2396                 client_obd_list_lock(&cli->cl_loi_list_lock);
2397                 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2398                         cfs_list_del_init(&oap->oap_rpc_item);
2399
2400                         /* queued sync pages can be torn down while the pages
2401                          * were between the pending list and the rpc */
2402                         if (oap->oap_interrupted) {
2403                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2404                                 osc_ap_completion(env, cli, NULL, oap, 0,
2405                                                   oap->oap_count);
2406                                 continue;
2407                         }
2408                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2409                 }
2410                 if (clerq && !IS_ERR(clerq))
2411                         cl_req_completion(env, clerq, PTR_ERR(req));
2412         }
2413         RETURN(req);
2414 }
2415
2416 /**
2417  * prepare pages for ASYNC io and put pages in send queue.
2418  *
2419  * \param cmd OBD_BRW_* macroses
2420  * \param lop pending pages
2421  *
2422  * \return zero if no page added to send queue.
2423  * \return 1 if pages successfully added to send queue.
2424  * \return negative on errors.
2425  */
2426 static int
2427 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2428                  struct lov_oinfo *loi,
2429                  int cmd, struct loi_oap_pages *lop)
2430 {
2431         struct ptlrpc_request *req;
2432         obd_count page_count = 0;
2433         struct osc_async_page *oap = NULL, *tmp;
2434         struct osc_brw_async_args *aa;
2435         const struct obd_async_page_ops *ops;
2436         CFS_LIST_HEAD(rpc_list);
2437         int srvlock = 0, mem_tight = 0;
2438         struct cl_object *clob = NULL;
2439         obd_off starting_offset = OBD_OBJECT_EOF;
2440         unsigned int ending_offset;
2441         int starting_page_off = 0;
2442         ENTRY;
2443
2444         /* ASYNC_HP pages first. At present, when the lock the pages is
2445          * to be canceled, the pages covered by the lock will be sent out
2446          * with ASYNC_HP. We have to send out them as soon as possible. */
2447         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2448                 if (oap->oap_async_flags & ASYNC_HP)
2449                         cfs_list_move(&oap->oap_pending_item, &lop->lop_pending);
2450                 if (++page_count >= cli->cl_max_pages_per_rpc)
2451                         break;
2452         }
2453         page_count = 0;
2454
2455         /* first we find the pages we're allowed to work with */
2456         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2457                                      oap_pending_item) {
2458                 ops = oap->oap_caller_ops;
2459
2460                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2461                          "magic 0x%x\n", oap, oap->oap_magic);
2462
2463                 if (clob == NULL) {
2464                         /* pin object in memory, so that completion call-backs
2465                          * can be safely called under client_obd_list lock. */
2466                         clob = osc_oap2cl_page(oap)->cp_obj;
2467                         cl_object_get(clob);
2468                 }
2469
2470                 if (page_count != 0 &&
2471                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2472                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2473                                " oap %p, page %p, srvlock %u\n",
2474                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2475                         break;
2476                 }
2477
2478                 /* If there is a gap at the start of this page, it can't merge
2479                  * with any previous page, so we'll hand the network a
2480                  * "fragmented" page array that it can't transfer in 1 RDMA */
2481                 if (oap->oap_obj_off < starting_offset) {
2482                         if (starting_page_off != 0)
2483                                 break;
2484
2485                         starting_page_off = oap->oap_page_off;
2486                         starting_offset = oap->oap_obj_off + starting_page_off;
2487                 } else if (oap->oap_page_off != 0)
2488                         break;
2489
2490                 /* in llite being 'ready' equates to the page being locked
2491                  * until completion unlocks it.  commit_write submits a page
2492                  * as not ready because its unlock will happen unconditionally
2493                  * as the call returns.  if we race with commit_write giving
2494                  * us that page we don't want to create a hole in the page
2495                  * stream, so we stop and leave the rpc to be fired by
2496                  * another dirtier or kupdated interval (the not ready page
2497                  * will still be on the dirty list).  we could call in
2498                  * at the end of ll_file_write to process the queue again. */
2499                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2500                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2501                                                     cmd);
2502                         if (rc < 0)
2503                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2504                                                 "instead of ready\n", oap,
2505                                                 oap->oap_page, rc);
2506                         switch (rc) {
2507                         case -EAGAIN:
2508                                 /* llite is telling us that the page is still
2509                                  * in commit_write and that we should try
2510                                  * and put it in an rpc again later.  we
2511                                  * break out of the loop so we don't create
2512                                  * a hole in the sequence of pages in the rpc
2513                                  * stream.*/
2514                                 oap = NULL;
2515                                 break;
2516                         case -EINTR:
2517                                 /* the io isn't needed.. tell the checks
2518                                  * below to complete the rpc with EINTR */
2519                                 cfs_spin_lock(&oap->oap_lock);
2520                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2521                                 cfs_spin_unlock(&oap->oap_lock);
2522                                 oap->oap_count = -EINTR;
2523                                 break;
2524                         case 0:
2525                                 cfs_spin_lock(&oap->oap_lock);
2526                                 oap->oap_async_flags |= ASYNC_READY;
2527                                 cfs_spin_unlock(&oap->oap_lock);
2528                                 break;
2529                         default:
2530                                 LASSERTF(0, "oap %p page %p returned %d "
2531                                             "from make_ready\n", oap,
2532                                             oap->oap_page, rc);
2533                                 break;
2534                         }
2535                 }
2536                 if (oap == NULL)
2537                         break;
2538                 /*
2539                  * Page submitted for IO has to be locked. Either by
2540                  * ->ap_make_ready() or by higher layers.
2541                  */
2542 #if defined(__KERNEL__) && defined(__linux__)
2543                 {
2544                         struct cl_page *page;
2545
2546                         page = osc_oap2cl_page(oap);
2547
2548                         if (page->cp_type == CPT_CACHEABLE &&
2549                             !(PageLocked(oap->oap_page) &&
2550                               (CheckWriteback(oap->oap_page, cmd)))) {
2551                                 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2552                                        oap->oap_page,
2553                                        (long)oap->oap_page->flags,
2554                                        oap->oap_async_flags);
2555                                 LBUG();
2556                         }
2557                 }
2558 #endif
2559
2560                 /* take the page out of our book-keeping */
2561                 cfs_list_del_init(&oap->oap_pending_item);
2562                 lop_update_pending(cli, lop, cmd, -1);
2563                 cfs_list_del_init(&oap->oap_urgent_item);
2564
2565                 /* ask the caller for the size of the io as the rpc leaves. */
2566                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2567                         oap->oap_count =
2568                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2569                                                       cmd);
2570                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2571                 }
2572                 if (oap->oap_count <= 0) {
2573                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2574                                oap->oap_count);
2575                         osc_ap_completion(env, cli, NULL,
2576                                           oap, 0, oap->oap_count);
2577                         continue;
2578                 }
2579
2580                 /* now put the page back in our accounting */
2581                 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2582                 if (page_count++ == 0)
2583                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2584
2585                 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2586                         mem_tight = 1;
2587
2588                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2589                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2590                  * have the same alignment as the initial writes that allocated
2591                  * extents on the server. */
2592                 ending_offset = oap->oap_obj_off + oap->oap_page_off +
2593                                 oap->oap_count;
2594                 if (!(ending_offset & (PTLRPC_MAX_BRW_SIZE - 1)))
2595                         break;
2596
2597                 if (page_count >= cli->cl_max_pages_per_rpc)
2598                         break;
2599
2600                 /* If there is a gap at the end of this page, it can't merge
2601                  * with any subsequent pages, so we'll hand the network a
2602                  * "fragmented" page array that it can't transfer in 1 RDMA */
2603                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2604                         break;
2605         }
2606
2607         osc_wake_cache_waiters(cli);
2608
2609         loi_list_maint(cli, loi);
2610
2611         client_obd_list_unlock(&cli->cl_loi_list_lock);
2612
2613         if (clob != NULL)
2614                 cl_object_put(env, clob);
2615
2616         if (page_count == 0) {
2617                 client_obd_list_lock(&cli->cl_loi_list_lock);
2618                 RETURN(0);
2619         }
2620
2621         req = osc_build_req(env, cli, &rpc_list, page_count,
2622                             mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2623         if (IS_ERR(req)) {
2624                 LASSERT(cfs_list_empty(&rpc_list));
2625                 loi_list_maint(cli, loi);
2626                 RETURN(PTR_ERR(req));
2627         }
2628
2629         aa = ptlrpc_req_async_args(req);
2630
2631         starting_offset &= PTLRPC_MAX_BRW_SIZE - 1;
2632         if (cmd == OBD_BRW_READ) {
2633                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2634                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2635                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2636                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2637         } else {
2638                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2639                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2640                                  cli->cl_w_in_flight);
2641                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2642                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2643         }
2644         ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2645
2646         client_obd_list_lock(&cli->cl_loi_list_lock);
2647
2648         if (cmd == OBD_BRW_READ)
2649                 cli->cl_r_in_flight++;
2650         else
2651                 cli->cl_w_in_flight++;
2652
2653         /* queued sync pages can be torn down while the pages
2654          * were between the pending list and the rpc */
2655         tmp = NULL;
2656         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2657                 /* only one oap gets a request reference */
2658                 if (tmp == NULL)
2659                         tmp = oap;
2660                 if (oap->oap_interrupted && !req->rq_intr) {
2661                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2662                                oap, req);
2663                         ptlrpc_mark_interrupted(req);
2664                 }
2665         }
2666         if (tmp != NULL)
2667                 tmp->oap_request = ptlrpc_request_addref(req);
2668
2669         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2670                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2671
2672         req->rq_interpret_reply = brw_interpret;
2673         ptlrpcd_add_req(req, PSCOPE_BRW);
2674         RETURN(1);
2675 }
2676
2677 #define LOI_DEBUG(LOI, STR, args...)                                     \
2678         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2679                !cfs_list_empty(&(LOI)->loi_ready_item) ||                \
2680                !cfs_list_empty(&(LOI)->loi_hp_ready_item),               \
2681                (LOI)->loi_write_lop.lop_num_pending,                     \
2682                !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent),        \
2683                (LOI)->loi_read_lop.lop_num_pending,                      \
2684                !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent),         \
2685                args)                                                     \
2686
2687 /* This is called by osc_check_rpcs() to find which objects have pages that
2688  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2689 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2690 {
2691         ENTRY;
2692
2693         /* First return objects that have blocked locks so that they
2694          * will be flushed quickly and other clients can get the lock,
2695          * then objects which have pages ready to be stuffed into RPCs */
2696         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2697                 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2698                                       struct lov_oinfo, loi_hp_ready_item));
2699         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2700                 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2701                                       struct lov_oinfo, loi_ready_item));
2702
2703         /* then if we have cache waiters, return all objects with queued
2704          * writes.  This is especially important when many small files
2705          * have filled up the cache and not been fired into rpcs because
2706          * they don't pass the nr_pending/object threshhold */
2707         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2708             !cfs_list_empty(&cli->cl_loi_write_list))
2709                 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2710                                       struct lov_oinfo, loi_write_item));
2711
2712         /* then return all queued objects when we have an invalid import
2713          * so that they get flushed */
2714         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2715                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2716                         RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2717                                               struct lov_oinfo,
2718                                               loi_write_item));
2719                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2720                         RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2721                                               struct lov_oinfo, loi_read_item));
2722         }
2723         RETURN(NULL);
2724 }
2725
2726 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2727 {
2728         struct osc_async_page *oap;
2729         int hprpc = 0;
2730
2731         if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2732                 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2733                                      struct osc_async_page, oap_urgent_item);
2734                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2735         }
2736
2737         if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2738                 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2739                                      struct osc_async_page, oap_urgent_item);
2740                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2741         }
2742
2743         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2744 }
2745
2746 /* called with the loi list lock held */
2747 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2748 {
2749         struct lov_oinfo *loi;
2750         int rc = 0, race_counter = 0;
2751         ENTRY;
2752
2753         while ((loi = osc_next_loi(cli)) != NULL) {
2754                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2755
2756                 if (osc_max_rpc_in_flight(cli, loi))
2757                         break;
2758
2759                 /* attempt some read/write balancing by alternating between
2760                  * reads and writes in an object.  The makes_rpc checks here
2761                  * would be redundant if we were getting read/write work items
2762                  * instead of objects.  we don't want send_oap_rpc to drain a
2763                  * partial read pending queue when we're given this object to
2764                  * do io on writes while there are cache waiters */
2765                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2766                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2767                                               &loi->loi_write_lop);
2768                         if (rc < 0) {
2769                                 CERROR("Write request failed with %d\n", rc);
2770
2771                                 /* osc_send_oap_rpc failed, mostly because of
2772                                  * memory pressure.
2773                                  *
2774                                  * It can't break here, because if:
2775                                  *  - a page was submitted by osc_io_submit, so
2776                                  *    page locked;
2777                                  *  - no request in flight
2778                                  *  - no subsequent request
2779                                  * The system will be in live-lock state,
2780                                  * because there is no chance to call
2781                                  * osc_io_unplug() and osc_check_rpcs() any
2782                                  * more. pdflush can't help in this case,
2783                                  * because it might be blocked at grabbing
2784                                  * the page lock as we mentioned.
2785                                  *
2786                                  * Anyway, continue to drain pages. */
2787                                 /* break; */
2788                         }
2789
2790                         if (rc > 0)
2791                                 race_counter = 0;
2792                         else if (rc == 0)
2793                                 race_counter++;
2794                 }
2795                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2796                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2797                                               &loi->loi_read_lop);
2798                         if (rc < 0)
2799                                 CERROR("Read request failed with %d\n", rc);
2800
2801                         if (rc > 0)
2802                                 race_counter = 0;
2803                         else if (rc == 0)
2804                                 race_counter++;
2805                 }
2806
2807                 /* attempt some inter-object balancing by issuing rpcs
2808                  * for each object in turn */
2809                 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2810                         cfs_list_del_init(&loi->loi_hp_ready_item);
2811                 if (!cfs_list_empty(&loi->loi_ready_item))
2812                         cfs_list_del_init(&loi->loi_ready_item);
2813                 if (!cfs_list_empty(&loi->loi_write_item))
2814                         cfs_list_del_init(&loi->loi_write_item);
2815                 if (!cfs_list_empty(&loi->loi_read_item))
2816                         cfs_list_del_init(&loi->loi_read_item);
2817
2818                 loi_list_maint(cli, loi);
2819
2820                 /* send_oap_rpc fails with 0 when make_ready tells it to
2821                  * back off.  llite's make_ready does this when it tries
2822                  * to lock a page queued for write that is already locked.
2823                  * we want to try sending rpcs from many objects, but we
2824                  * don't want to spin failing with 0.  */
2825                 if (race_counter == 10)
2826                         break;
2827         }
2828         EXIT;
2829 }
2830
2831 /* we're trying to queue a page in the osc so we're subject to the
2832  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2833  * If the osc's queued pages are already at that limit, then we want to sleep
2834  * until there is space in the osc's queue for us.  We also may be waiting for
2835  * write credits from the OST if there are RPCs in flight that may return some
2836  * before we fall back to sync writes.
2837  *
2838  * We need this know our allocation was granted in the presence of signals */
2839 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2840 {
2841         int rc;
2842         ENTRY;
2843         client_obd_list_lock(&cli->cl_loi_list_lock);
2844         rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2845         client_obd_list_unlock(&cli->cl_loi_list_lock);
2846         RETURN(rc);
2847 };
2848
2849 /**
2850  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2851  * is available.
2852  */
2853 int osc_enter_cache_try(const struct lu_env *env,
2854                         struct client_obd *cli, struct lov_oinfo *loi,
2855                         struct osc_async_page *oap, int transient)
2856 {
2857         int has_grant;
2858
2859         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2860         if (has_grant) {
2861                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2862                 if (transient) {
2863                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2864                         cfs_atomic_inc(&obd_dirty_transit_pages);
2865                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2866                 }
2867         }
2868         return has_grant;
2869 }
2870
2871 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2872  * grant or cache space. */
2873 static int osc_enter_cache(const struct lu_env *env,
2874                            struct client_obd *cli, struct lov_oinfo *loi,
2875                            struct osc_async_page *oap)
2876 {
2877         struct osc_cache_waiter ocw;
2878         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
2879
2880         ENTRY;
2881
2882         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2883                "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2884                cli->cl_dirty_max, obd_max_dirty_pages,
2885                cli->cl_lost_grant, cli->cl_avail_grant);
2886
2887         /* force the caller to try sync io.  this can jump the list
2888          * of queued writes and create a discontiguous rpc stream */
2889         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
2890             cli->cl_dirty_max < CFS_PAGE_SIZE     ||
2891             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
2892                 RETURN(-EDQUOT);
2893
2894         /* Hopefully normal case - cache space and write credits available */
2895         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2896             cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2897             osc_enter_cache_try(env, cli, loi, oap, 0))
2898                 RETURN(0);
2899
2900         /* It is safe to block as a cache waiter as long as there is grant
2901          * space available or the hope of additional grant being returned
2902          * when an in flight write completes.  Using the write back cache
2903          * if possible is preferable to sending the data synchronously
2904          * because write pages can then be merged in to large requests.
2905          * The addition of this cache waiter will causing pending write
2906          * pages to be sent immediately. */
2907         if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2908                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2909                 cfs_waitq_init(&ocw.ocw_waitq);
2910                 ocw.ocw_oap = oap;
2911                 ocw.ocw_rc = 0;
2912
2913                 loi_list_maint(cli, loi);
2914                 osc_check_rpcs(env, cli);
2915                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2916
2917                 CDEBUG(D_CACHE, "sleeping for cache space\n");
2918                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2919
2920                 client_obd_list_lock(&cli->cl_loi_list_lock);
2921                 if (!cfs_list_empty(&ocw.ocw_entry)) {
2922                         cfs_list_del(&ocw.ocw_entry);
2923                         RETURN(-EINTR);
2924                 }
2925                 RETURN(ocw.ocw_rc);
2926         }
2927
2928         RETURN(-EDQUOT);
2929 }
2930
2931
2932 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2933                         struct lov_oinfo *loi, cfs_page_t *page,
2934                         obd_off offset, const struct obd_async_page_ops *ops,
2935                         void *data, void **res, int nocache,
2936                         struct lustre_handle *lockh)
2937 {
2938         struct osc_async_page *oap;
2939
2940         ENTRY;
2941
2942         if (!page)
2943                 return cfs_size_round(sizeof(*oap));
2944
2945         oap = *res;
2946         oap->oap_magic = OAP_MAGIC;
2947         oap->oap_cli = &exp->exp_obd->u.cli;
2948         oap->oap_loi = loi;
2949
2950         oap->oap_caller_ops = ops;
2951         oap->oap_caller_data = data;
2952
2953         oap->oap_page = page;
2954         oap->oap_obj_off = offset;
2955         if (!client_is_remote(exp) &&
2956             cfs_capable(CFS_CAP_SYS_RESOURCE))
2957                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2958
2959         LASSERT(!(offset & ~CFS_PAGE_MASK));
2960
2961         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2962         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2963         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2964         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2965
2966         cfs_spin_lock_init(&oap->oap_lock);
2967         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2968         RETURN(0);
2969 }
2970
2971 int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
2972                        struct lov_stripe_md *lsm, struct lov_oinfo *loi,
2973                        struct osc_async_page *oap, int cmd, int off,
2974                        int count, obd_flag brw_flags, enum async_flags async_flags)
2975 {
2976         struct client_obd *cli = &exp->exp_obd->u.cli;
2977         int rc = 0;
2978         ENTRY;
2979
2980         if (oap->oap_magic != OAP_MAGIC)
2981                 RETURN(-EINVAL);
2982
2983         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2984                 RETURN(-EIO);
2985
2986         if (!cfs_list_empty(&oap->oap_pending_item) ||
2987             !cfs_list_empty(&oap->oap_urgent_item) ||
2988             !cfs_list_empty(&oap->oap_rpc_item))
2989                 RETURN(-EBUSY);
2990
2991         /* check if the file's owner/group is over quota */
2992         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2993                 struct cl_object *obj;
2994                 struct cl_attr    attr; /* XXX put attr into thread info */
2995                 unsigned int qid[MAXQUOTAS];
2996
2997                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2998
2999                 cl_object_attr_lock(obj);
3000                 rc = cl_object_attr_get(env, obj, &attr);
3001                 cl_object_attr_unlock(obj);
3002
3003                 qid[USRQUOTA] = attr.cat_uid;
3004                 qid[GRPQUOTA] = attr.cat_gid;
3005                 if (rc == 0 &&
3006                     lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
3007                         rc = -EDQUOT;
3008                 if (rc)
3009                         RETURN(rc);
3010         }
3011
3012         if (loi == NULL)
3013                 loi = lsm->lsm_oinfo[0];
3014
3015         client_obd_list_lock(&cli->cl_loi_list_lock);
3016
3017         LASSERT(off + count <= CFS_PAGE_SIZE);
3018         oap->oap_cmd = cmd;
3019         oap->oap_page_off = off;
3020         oap->oap_count = count;
3021         oap->oap_brw_flags = brw_flags;
3022         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
3023         if (cfs_memory_pressure_get())
3024                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
3025         cfs_spin_lock(&oap->oap_lock);
3026         oap->oap_async_flags = async_flags;
3027         cfs_spin_unlock(&oap->oap_lock);
3028
3029         if (cmd & OBD_BRW_WRITE) {
3030                 rc = osc_enter_cache(env, cli, loi, oap);
3031                 if (rc) {
3032                         client_obd_list_unlock(&cli->cl_loi_list_lock);
3033                         RETURN(rc);
3034                 }
3035         }
3036
3037         osc_oap_to_pending(oap);
3038         loi_list_maint(cli, loi);
3039
3040         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3041                   cmd);
3042
3043         osc_check_rpcs(env, cli);
3044         client_obd_list_unlock(&cli->cl_loi_list_lock);
3045
3046         RETURN(0);
3047 }
3048
3049 /* aka (~was & now & flag), but this is more clear :) */
3050 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3051
3052 int osc_set_async_flags_base(struct client_obd *cli,
3053                              struct lov_oinfo *loi, struct osc_async_page *oap,
3054                              obd_flag async_flags)
3055 {
3056         struct loi_oap_pages *lop;
3057         int flags = 0;
3058         ENTRY;
3059
3060         LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3061
3062         if (oap->oap_cmd & OBD_BRW_WRITE) {
3063                 lop = &loi->loi_write_lop;
3064         } else {
3065                 lop = &loi->loi_read_lop;
3066         }
3067
3068         if ((oap->oap_async_flags & async_flags) == async_flags)
3069                 RETURN(0);
3070
3071         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3072                 flags |= ASYNC_READY;
3073
3074         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3075             cfs_list_empty(&oap->oap_rpc_item)) {
3076                 if (oap->oap_async_flags & ASYNC_HP)
3077                         cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3078                 else
3079                         cfs_list_add_tail(&oap->oap_urgent_item,
3080                                           &lop->lop_urgent);
3081                 flags |= ASYNC_URGENT;
3082                 loi_list_maint(cli, loi);
3083         }
3084         cfs_spin_lock(&oap->oap_lock);
3085         oap->oap_async_flags |= flags;
3086         cfs_spin_unlock(&oap->oap_lock);
3087
3088         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3089                         oap->oap_async_flags);
3090         RETURN(0);
3091 }
3092
3093 int osc_teardown_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
3094                             struct lov_oinfo *loi, struct osc_async_page *oap)
3095 {
3096         struct client_obd *cli = &exp->exp_obd->u.cli;
3097         struct loi_oap_pages *lop;
3098         int rc = 0;
3099         ENTRY;
3100
3101         if (oap->oap_magic != OAP_MAGIC)
3102                 RETURN(-EINVAL);
3103
3104         if (loi == NULL)
3105                 loi = lsm->lsm_oinfo[0];
3106
3107         if (oap->oap_cmd & OBD_BRW_WRITE) {
3108                 lop = &loi->loi_write_lop;
3109         } else {
3110                 lop = &loi->loi_read_lop;
3111         }
3112
3113         client_obd_list_lock(&cli->cl_loi_list_lock);
3114
3115         if (!cfs_list_empty(&oap->oap_rpc_item))
3116                 GOTO(out, rc = -EBUSY);
3117
3118         osc_exit_cache(cli, oap, 0);
3119         osc_wake_cache_waiters(cli);
3120
3121         if (!cfs_list_empty(&oap->oap_urgent_item)) {
3122                 cfs_list_del_init(&oap->oap_urgent_item);
3123                 cfs_spin_lock(&oap->oap_lock);
3124                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3125                 cfs_spin_unlock(&oap->oap_lock);
3126         }
3127         if (!cfs_list_empty(&oap->oap_pending_item)) {
3128                 cfs_list_del_init(&oap->oap_pending_item);
3129                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3130         }
3131         loi_list_maint(cli, loi);
3132         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3133 out:
3134         client_obd_list_unlock(&cli->cl_loi_list_lock);
3135         RETURN(rc);
3136 }
3137
3138 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
3139                                         struct ldlm_enqueue_info *einfo)
3140 {
3141         void *data = einfo->ei_cbdata;
3142         int set = 0;
3143
3144         LASSERT(lock != NULL);
3145         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3146         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3147         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3148         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3149
3150         lock_res_and_lock(lock);
3151         cfs_spin_lock(&osc_ast_guard);
3152
3153         if (lock->l_ast_data == NULL)
3154                 lock->l_ast_data = data;
3155         if (lock->l_ast_data == data)
3156                 set = 1;
3157
3158         cfs_spin_unlock(&osc_ast_guard);
3159         unlock_res_and_lock(lock);
3160
3161         return set;
3162 }
3163
3164 static int osc_set_data_with_check(struct lustre_handle *lockh,