Whamcloud - gitweb
LU-106 procfs: many proc entries are not accessed safely
[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, PDL_POLICY_ROUND, -1);
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, PDL_POLICY_ROUND, -1);
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, PDL_POLICY_ROUND, -1);
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, PDL_POLICY_ROUND, -1);
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| OBD_BRW_NOCACHE|
1223                                   OBD_BRW_SYNC|OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1224
1225                 /* warn if we try to combine flags that we don't know to be
1226                  * safe to combine */
1227                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1228                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1229                               "report this at http://bugs.whamcloud.com/\n",
1230                               p1->flag, p2->flag);
1231                 }
1232                 return 0;
1233         }
1234
1235         return (p1->off + p1->count == p2->off);
1236 }
1237
1238 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1239                                    struct brw_page **pga, int opc,
1240                                    cksum_type_t cksum_type)
1241 {
1242         __u32 cksum;
1243         int i = 0;
1244
1245         LASSERT (pg_count > 0);
1246         cksum = init_checksum(cksum_type);
1247         while (nob > 0 && pg_count > 0) {
1248                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1249                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1250                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1251
1252                 /* corrupt the data before we compute the checksum, to
1253                  * simulate an OST->client data error */
1254                 if (i == 0 && opc == OST_READ &&
1255                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1256                         memcpy(ptr + off, "bad1", min(4, nob));
1257                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1258                 cfs_kunmap(pga[i]->pg);
1259                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1260                                off, cksum);
1261
1262                 nob -= pga[i]->count;
1263                 pg_count--;
1264                 i++;
1265         }
1266         /* For sending we only compute the wrong checksum instead
1267          * of corrupting the data so it is still correct on a redo */
1268         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1269                 cksum++;
1270
1271         return cksum;
1272 }
1273
1274 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1275                                 struct lov_stripe_md *lsm, obd_count page_count,
1276                                 struct brw_page **pga,
1277                                 struct ptlrpc_request **reqp,
1278                                 struct obd_capa *ocapa, int reserve,
1279                                 int resend)
1280 {
1281         struct ptlrpc_request   *req;
1282         struct ptlrpc_bulk_desc *desc;
1283         struct ost_body         *body;
1284         struct obd_ioobj        *ioobj;
1285         struct niobuf_remote    *niobuf;
1286         int niocount, i, requested_nob, opc, rc;
1287         struct osc_brw_async_args *aa;
1288         struct req_capsule      *pill;
1289         struct brw_page *pg_prev;
1290
1291         ENTRY;
1292         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1293                 RETURN(-ENOMEM); /* Recoverable */
1294         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1295                 RETURN(-EINVAL); /* Fatal */
1296
1297         if ((cmd & OBD_BRW_WRITE) != 0) {
1298                 opc = OST_WRITE;
1299                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1300                                                 cli->cl_import->imp_rq_pool,
1301                                                 &RQF_OST_BRW_WRITE);
1302         } else {
1303                 opc = OST_READ;
1304                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1305         }
1306         if (req == NULL)
1307                 RETURN(-ENOMEM);
1308
1309         for (niocount = i = 1; i < page_count; i++) {
1310                 if (!can_merge_pages(pga[i - 1], pga[i]))
1311                         niocount++;
1312         }
1313
1314         pill = &req->rq_pill;
1315         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1316                              sizeof(*ioobj));
1317         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1318                              niocount * sizeof(*niobuf));
1319         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1320
1321         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1322         if (rc) {
1323                 ptlrpc_request_free(req);
1324                 RETURN(rc);
1325         }
1326         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1327         ptlrpc_at_set_req_timeout(req);
1328
1329         if (opc == OST_WRITE)
1330                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1331                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1332         else
1333                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1334                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1335
1336         if (desc == NULL)
1337                 GOTO(out, rc = -ENOMEM);
1338         /* NB request now owns desc and will free it when it gets freed */
1339
1340         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1341         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1342         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1343         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1344
1345         lustre_set_wire_obdo(&body->oa, oa);
1346
1347         obdo_to_ioobj(oa, ioobj);
1348         ioobj->ioo_bufcnt = niocount;
1349         osc_pack_capa(req, body, ocapa);
1350         LASSERT (page_count > 0);
1351         pg_prev = pga[0];
1352         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1353                 struct brw_page *pg = pga[i];
1354                 int poff = pg->off & ~CFS_PAGE_MASK;
1355
1356                 LASSERT(pg->count > 0);
1357                 /* make sure there is no gap in the middle of page array */
1358                 LASSERTF(page_count == 1 ||
1359                          (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1360                           ergo(i > 0 && i < page_count - 1,
1361                                poff == 0 && pg->count == CFS_PAGE_SIZE)   &&
1362                           ergo(i == page_count - 1, poff == 0)),
1363                          "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1364                          i, page_count, pg, pg->off, pg->count);
1365 #ifdef __linux__
1366                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1367                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1368                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1369                          i, page_count,
1370                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1371                          pg_prev->pg, page_private(pg_prev->pg),
1372                          pg_prev->pg->index, pg_prev->off);
1373 #else
1374                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1375                          "i %d p_c %u\n", i, page_count);
1376 #endif
1377                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1378                         (pg->flag & OBD_BRW_SRVLOCK));
1379
1380                 ptlrpc_prep_bulk_page(desc, pg->pg, poff, pg->count);
1381                 requested_nob += pg->count;
1382
1383                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1384                         niobuf--;
1385                         niobuf->len += pg->count;
1386                 } else {
1387                         niobuf->offset = pg->off;
1388                         niobuf->len    = pg->count;
1389                         niobuf->flags  = pg->flag;
1390                 }
1391                 pg_prev = pg;
1392         }
1393
1394         LASSERTF((void *)(niobuf - niocount) ==
1395                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1396                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1397                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1398
1399         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1400         if (resend) {
1401                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1402                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1403                         body->oa.o_flags = 0;
1404                 }
1405                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1406         }
1407
1408         if (osc_should_shrink_grant(cli))
1409                 osc_shrink_grant_local(cli, &body->oa);
1410
1411         /* size[REQ_REC_OFF] still sizeof (*body) */
1412         if (opc == OST_WRITE) {
1413                 if (unlikely(cli->cl_checksum) &&
1414                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1415                         /* store cl_cksum_type in a local variable since
1416                          * it can be changed via lprocfs */
1417                         cksum_type_t cksum_type = cli->cl_cksum_type;
1418
1419                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1420                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1421                                 body->oa.o_flags = 0;
1422                         }
1423                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1424                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1425                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1426                                                              page_count, pga,
1427                                                              OST_WRITE,
1428                                                              cksum_type);
1429                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1430                                body->oa.o_cksum);
1431                         /* save this in 'oa', too, for later checking */
1432                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1433                         oa->o_flags |= cksum_type_pack(cksum_type);
1434                 } else {
1435                         /* clear out the checksum flag, in case this is a
1436                          * resend but cl_checksum is no longer set. b=11238 */
1437                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1438                 }
1439                 oa->o_cksum = body->oa.o_cksum;
1440                 /* 1 RC per niobuf */
1441                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1442                                      sizeof(__u32) * niocount);
1443         } else {
1444                 if (unlikely(cli->cl_checksum) &&
1445                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1446                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1447                                 body->oa.o_flags = 0;
1448                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1449                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1450                 }
1451         }
1452         ptlrpc_request_set_replen(req);
1453
1454         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1455         aa = ptlrpc_req_async_args(req);
1456         aa->aa_oa = oa;
1457         aa->aa_requested_nob = requested_nob;
1458         aa->aa_nio_count = niocount;
1459         aa->aa_page_count = page_count;
1460         aa->aa_resends = 0;
1461         aa->aa_ppga = pga;
1462         aa->aa_cli = cli;
1463         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1464         if (ocapa && reserve)
1465                 aa->aa_ocapa = capa_get(ocapa);
1466
1467         *reqp = req;
1468         RETURN(0);
1469
1470  out:
1471         ptlrpc_req_finished(req);
1472         RETURN(rc);
1473 }
1474
1475 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1476                                 __u32 client_cksum, __u32 server_cksum, int nob,
1477                                 obd_count page_count, struct brw_page **pga,
1478                                 cksum_type_t client_cksum_type)
1479 {
1480         __u32 new_cksum;
1481         char *msg;
1482         cksum_type_t cksum_type;
1483
1484         if (server_cksum == client_cksum) {
1485                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1486                 return 0;
1487         }
1488
1489         /* If this is mmaped file - it can be changed at any time */
1490         if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
1491                 return 1;
1492
1493         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1494                                        oa->o_flags : 0);
1495         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1496                                       cksum_type);
1497
1498         if (cksum_type != client_cksum_type)
1499                 msg = "the server did not use the checksum type specified in "
1500                       "the original request - likely a protocol problem";
1501         else if (new_cksum == server_cksum)
1502                 msg = "changed on the client after we checksummed it - "
1503                       "likely false positive due to mmap IO (bug 11742)";
1504         else if (new_cksum == client_cksum)
1505                 msg = "changed in transit before arrival at OST";
1506         else
1507                 msg = "changed in transit AND doesn't match the original - "
1508                       "likely false positive due to mmap IO (bug 11742)";
1509
1510         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1511                            " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1512                            msg, libcfs_nid2str(peer->nid),
1513                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1514                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1515                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1516                            oa->o_id,
1517                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1518                            pga[0]->off,
1519                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1520         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1521                "client csum now %x\n", client_cksum, client_cksum_type,
1522                server_cksum, cksum_type, new_cksum);
1523         return 1;
1524 }
1525
1526 /* Note rc enters this function as number of bytes transferred */
1527 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1528 {
1529         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1530         const lnet_process_id_t *peer =
1531                         &req->rq_import->imp_connection->c_peer;
1532         struct client_obd *cli = aa->aa_cli;
1533         struct ost_body *body;
1534         __u32 client_cksum = 0;
1535         ENTRY;
1536
1537         if (rc < 0 && rc != -EDQUOT) {
1538                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1539                 RETURN(rc);
1540         }
1541
1542         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1543         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1544         if (body == NULL) {
1545                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1546                 RETURN(-EPROTO);
1547         }
1548
1549 #ifdef HAVE_QUOTA_SUPPORT
1550         /* set/clear over quota flag for a uid/gid */
1551         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1552             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1553                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1554
1555                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1556                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1557                        body->oa.o_flags);
1558                 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1559                              body->oa.o_flags);
1560         }
1561 #endif
1562
1563         osc_update_grant(cli, body);
1564
1565         if (rc < 0)
1566                 RETURN(rc);
1567
1568         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1569                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1570
1571         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1572                 if (rc > 0) {
1573                         CERROR("Unexpected +ve rc %d\n", rc);
1574                         RETURN(-EPROTO);
1575                 }
1576                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1577
1578                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1579                         RETURN(-EAGAIN);
1580
1581                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1582                     check_write_checksum(&body->oa, peer, client_cksum,
1583                                          body->oa.o_cksum, aa->aa_requested_nob,
1584                                          aa->aa_page_count, aa->aa_ppga,
1585                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1586                         RETURN(-EAGAIN);
1587
1588                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1589                                      aa->aa_page_count, aa->aa_ppga);
1590                 GOTO(out, rc);
1591         }
1592
1593         /* The rest of this function executes only for OST_READs */
1594
1595         /* if unwrap_bulk failed, return -EAGAIN to retry */
1596         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1597         if (rc < 0)
1598                 GOTO(out, rc = -EAGAIN);
1599
1600         if (rc > aa->aa_requested_nob) {
1601                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1602                        aa->aa_requested_nob);
1603                 RETURN(-EPROTO);
1604         }
1605
1606         if (rc != req->rq_bulk->bd_nob_transferred) {
1607                 CERROR ("Unexpected rc %d (%d transferred)\n",
1608                         rc, req->rq_bulk->bd_nob_transferred);
1609                 return (-EPROTO);
1610         }
1611
1612         if (rc < aa->aa_requested_nob)
1613                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1614
1615         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1616                 static int cksum_counter;
1617                 __u32      server_cksum = body->oa.o_cksum;
1618                 char      *via;
1619                 char      *router;
1620                 cksum_type_t cksum_type;
1621
1622                 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1623                                                body->oa.o_flags : 0);
1624                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1625                                                  aa->aa_ppga, OST_READ,
1626                                                  cksum_type);
1627
1628                 if (peer->nid == req->rq_bulk->bd_sender) {
1629                         via = router = "";
1630                 } else {
1631                         via = " via ";
1632                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1633                 }
1634
1635                 if (server_cksum == ~0 && rc > 0) {
1636                         CERROR("Protocol error: server %s set the 'checksum' "
1637                                "bit, but didn't send a checksum.  Not fatal, "
1638                                "but please notify on http://bugs.whamcloud.com/\n",
1639                                libcfs_nid2str(peer->nid));
1640                 } else if (server_cksum != client_cksum) {
1641                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1642                                            "%s%s%s inode "DFID" object "
1643                                            LPU64"/"LPU64" extent "
1644                                            "["LPU64"-"LPU64"]\n",
1645                                            req->rq_import->imp_obd->obd_name,
1646                                            libcfs_nid2str(peer->nid),
1647                                            via, router,
1648                                            body->oa.o_valid & OBD_MD_FLFID ?
1649                                                 body->oa.o_parent_seq : (__u64)0,
1650                                            body->oa.o_valid & OBD_MD_FLFID ?
1651                                                 body->oa.o_parent_oid : 0,
1652                                            body->oa.o_valid & OBD_MD_FLFID ?
1653                                                 body->oa.o_parent_ver : 0,
1654                                            body->oa.o_id,
1655                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1656                                                 body->oa.o_seq : (__u64)0,
1657                                            aa->aa_ppga[0]->off,
1658                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1659                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1660                                                                         1);
1661                         CERROR("client %x, server %x, cksum_type %x\n",
1662                                client_cksum, server_cksum, cksum_type);
1663                         cksum_counter = 0;
1664                         aa->aa_oa->o_cksum = client_cksum;
1665                         rc = -EAGAIN;
1666                 } else {
1667                         cksum_counter++;
1668                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1669                         rc = 0;
1670                 }
1671         } else if (unlikely(client_cksum)) {
1672                 static int cksum_missed;
1673
1674                 cksum_missed++;
1675                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1676                         CERROR("Checksum %u requested from %s but not sent\n",
1677                                cksum_missed, libcfs_nid2str(peer->nid));
1678         } else {
1679                 rc = 0;
1680         }
1681 out:
1682         if (rc >= 0)
1683                 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1684
1685         RETURN(rc);
1686 }
1687
1688 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1689                             struct lov_stripe_md *lsm,
1690                             obd_count page_count, struct brw_page **pga,
1691                             struct obd_capa *ocapa)
1692 {
1693         struct ptlrpc_request *req;
1694         int                    rc;
1695         cfs_waitq_t            waitq;
1696         int                    resends = 0;
1697         struct l_wait_info     lwi;
1698
1699         ENTRY;
1700
1701         cfs_waitq_init(&waitq);
1702
1703 restart_bulk:
1704         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1705                                   page_count, pga, &req, ocapa, 0, resends);
1706         if (rc != 0)
1707                 return (rc);
1708
1709         rc = ptlrpc_queue_wait(req);
1710
1711         if (rc == -ETIMEDOUT && req->rq_resend) {
1712                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1713                 ptlrpc_req_finished(req);
1714                 goto restart_bulk;
1715         }
1716
1717         rc = osc_brw_fini_request(req, rc);
1718
1719         ptlrpc_req_finished(req);
1720         if (osc_recoverable_error(rc)) {
1721                 resends++;
1722                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1723                         CERROR("too many resend retries, returning error\n");
1724                         RETURN(-EIO);
1725                 }
1726
1727                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1728                 l_wait_event(waitq, 0, &lwi);
1729
1730                 goto restart_bulk;
1731         }
1732
1733         RETURN (rc);
1734 }
1735
1736 int osc_brw_redo_request(struct ptlrpc_request *request,
1737                          struct osc_brw_async_args *aa)
1738 {
1739         struct ptlrpc_request *new_req;
1740         struct ptlrpc_request_set *set = request->rq_set;
1741         struct osc_brw_async_args *new_aa;
1742         struct osc_async_page *oap;
1743         int rc = 0;
1744         ENTRY;
1745
1746         if (!client_should_resend(aa->aa_resends, aa->aa_cli)) {
1747                 CERROR("too many resent retries, returning error\n");
1748                 RETURN(-EIO);
1749         }
1750
1751         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1752
1753         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1754                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1755                                   aa->aa_cli, aa->aa_oa,
1756                                   NULL /* lsm unused by osc currently */,
1757                                   aa->aa_page_count, aa->aa_ppga,
1758                                   &new_req, aa->aa_ocapa, 0, 1);
1759         if (rc)
1760                 RETURN(rc);
1761
1762         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1763
1764         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1765                 if (oap->oap_request != NULL) {
1766                         LASSERTF(request == oap->oap_request,
1767                                  "request %p != oap_request %p\n",
1768                                  request, oap->oap_request);
1769                         if (oap->oap_interrupted) {
1770                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1771                                 ptlrpc_req_finished(new_req);
1772                                 RETURN(-EINTR);
1773                         }
1774                 }
1775         }
1776         /* New request takes over pga and oaps from old request.
1777          * Note that copying a list_head doesn't work, need to move it... */
1778         aa->aa_resends++;
1779         new_req->rq_interpret_reply = request->rq_interpret_reply;
1780         new_req->rq_async_args = request->rq_async_args;
1781         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1782
1783         new_aa = ptlrpc_req_async_args(new_req);
1784
1785         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1786         cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1787         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1788
1789         cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1790                 if (oap->oap_request) {
1791                         ptlrpc_req_finished(oap->oap_request);
1792                         oap->oap_request = ptlrpc_request_addref(new_req);
1793                 }
1794         }
1795
1796         new_aa->aa_ocapa = aa->aa_ocapa;
1797         aa->aa_ocapa = NULL;
1798
1799         /* use ptlrpc_set_add_req is safe because interpret functions work
1800          * in check_set context. only one way exist with access to request
1801          * from different thread got -EINTR - this way protected with
1802          * cl_loi_list_lock */
1803         ptlrpc_set_add_req(set, new_req);
1804
1805         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1806
1807         DEBUG_REQ(D_INFO, new_req, "new request");
1808         RETURN(0);
1809 }
1810
1811 /*
1812  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1813  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1814  * fine for our small page arrays and doesn't require allocation.  its an
1815  * insertion sort that swaps elements that are strides apart, shrinking the
1816  * stride down until its '1' and the array is sorted.
1817  */
1818 static void sort_brw_pages(struct brw_page **array, int num)
1819 {
1820         int stride, i, j;
1821         struct brw_page *tmp;
1822
1823         if (num == 1)
1824                 return;
1825         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1826                 ;
1827
1828         do {
1829                 stride /= 3;
1830                 for (i = stride ; i < num ; i++) {
1831                         tmp = array[i];
1832                         j = i;
1833                         while (j >= stride && array[j - stride]->off > tmp->off) {
1834                                 array[j] = array[j - stride];
1835                                 j -= stride;
1836                         }
1837                         array[j] = tmp;
1838                 }
1839         } while (stride > 1);
1840 }
1841
1842 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1843 {
1844         int count = 1;
1845         int offset;
1846         int i = 0;
1847
1848         LASSERT (pages > 0);
1849         offset = pg[i]->off & ~CFS_PAGE_MASK;
1850
1851         for (;;) {
1852                 pages--;
1853                 if (pages == 0)         /* that's all */
1854                         return count;
1855
1856                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1857                         return count;   /* doesn't end on page boundary */
1858
1859                 i++;
1860                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1861                 if (offset != 0)        /* doesn't start on page boundary */
1862                         return count;
1863
1864                 count++;
1865         }
1866 }
1867
1868 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1869 {
1870         struct brw_page **ppga;
1871         int i;
1872
1873         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1874         if (ppga == NULL)
1875                 return NULL;
1876
1877         for (i = 0; i < count; i++)
1878                 ppga[i] = pga + i;
1879         return ppga;
1880 }
1881
1882 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1883 {
1884         LASSERT(ppga != NULL);
1885         OBD_FREE(ppga, sizeof(*ppga) * count);
1886 }
1887
1888 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1889                    obd_count page_count, struct brw_page *pga,
1890                    struct obd_trans_info *oti)
1891 {
1892         struct obdo *saved_oa = NULL;
1893         struct brw_page **ppga, **orig;
1894         struct obd_import *imp = class_exp2cliimp(exp);
1895         struct client_obd *cli;
1896         int rc, page_count_orig;
1897         ENTRY;
1898
1899         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1900         cli = &imp->imp_obd->u.cli;
1901
1902         if (cmd & OBD_BRW_CHECK) {
1903                 /* The caller just wants to know if there's a chance that this
1904                  * I/O can succeed */
1905
1906                 if (imp->imp_invalid)
1907                         RETURN(-EIO);
1908                 RETURN(0);
1909         }
1910
1911         /* test_brw with a failed create can trip this, maybe others. */
1912         LASSERT(cli->cl_max_pages_per_rpc);
1913
1914         rc = 0;
1915
1916         orig = ppga = osc_build_ppga(pga, page_count);
1917         if (ppga == NULL)
1918                 RETURN(-ENOMEM);
1919         page_count_orig = page_count;
1920
1921         sort_brw_pages(ppga, page_count);
1922         while (page_count) {
1923                 obd_count pages_per_brw;
1924
1925                 if (page_count > cli->cl_max_pages_per_rpc)
1926                         pages_per_brw = cli->cl_max_pages_per_rpc;
1927                 else
1928                         pages_per_brw = page_count;
1929
1930                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1931
1932                 if (saved_oa != NULL) {
1933                         /* restore previously saved oa */
1934                         *oinfo->oi_oa = *saved_oa;
1935                 } else if (page_count > pages_per_brw) {
1936                         /* save a copy of oa (brw will clobber it) */
1937                         OBDO_ALLOC(saved_oa);
1938                         if (saved_oa == NULL)
1939                                 GOTO(out, rc = -ENOMEM);
1940                         *saved_oa = *oinfo->oi_oa;
1941                 }
1942
1943                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1944                                       pages_per_brw, ppga, oinfo->oi_capa);
1945
1946                 if (rc != 0)
1947                         break;
1948
1949                 page_count -= pages_per_brw;
1950                 ppga += pages_per_brw;
1951         }
1952
1953 out:
1954         osc_release_ppga(orig, page_count_orig);
1955
1956         if (saved_oa != NULL)
1957                 OBDO_FREE(saved_oa);
1958
1959         RETURN(rc);
1960 }
1961
1962 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1963  * the dirty accounting.  Writeback completes or truncate happens before
1964  * writing starts.  Must be called with the loi lock held. */
1965 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1966                            int sent)
1967 {
1968         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1969 }
1970
1971
1972 /* This maintains the lists of pending pages to read/write for a given object
1973  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1974  * to quickly find objects that are ready to send an RPC. */
1975 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1976                          int cmd)
1977 {
1978         int optimal;
1979         ENTRY;
1980
1981         if (lop->lop_num_pending == 0)
1982                 RETURN(0);
1983
1984         /* if we have an invalid import we want to drain the queued pages
1985          * by forcing them through rpcs that immediately fail and complete
1986          * the pages.  recovery relies on this to empty the queued pages
1987          * before canceling the locks and evicting down the llite pages */
1988         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1989                 RETURN(1);
1990
1991         /* stream rpcs in queue order as long as as there is an urgent page
1992          * queued.  this is our cheap solution for good batching in the case
1993          * where writepage marks some random page in the middle of the file
1994          * as urgent because of, say, memory pressure */
1995         if (!cfs_list_empty(&lop->lop_urgent)) {
1996                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1997                 RETURN(1);
1998         }
1999         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
2000         optimal = cli->cl_max_pages_per_rpc;
2001         if (cmd & OBD_BRW_WRITE) {
2002                 /* trigger a write rpc stream as long as there are dirtiers
2003                  * waiting for space.  as they're waiting, they're not going to
2004                  * create more pages to coalesce with what's waiting.. */
2005                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
2006                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
2007                         RETURN(1);
2008                 }
2009                 /* +16 to avoid triggering rpcs that would want to include pages
2010                  * that are being queued but which can't be made ready until
2011                  * the queuer finishes with the page. this is a wart for
2012                  * llite::commit_write() */
2013                 optimal += 16;
2014         }
2015         if (lop->lop_num_pending >= optimal)
2016                 RETURN(1);
2017
2018         RETURN(0);
2019 }
2020
2021 static int lop_makes_hprpc(struct loi_oap_pages *lop)
2022 {
2023         struct osc_async_page *oap;
2024         ENTRY;
2025
2026         if (cfs_list_empty(&lop->lop_urgent))
2027                 RETURN(0);
2028
2029         oap = cfs_list_entry(lop->lop_urgent.next,
2030                          struct osc_async_page, oap_urgent_item);
2031
2032         if (oap->oap_async_flags & ASYNC_HP) {
2033                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
2034                 RETURN(1);
2035         }
2036
2037         RETURN(0);
2038 }
2039
2040 static void on_list(cfs_list_t *item, cfs_list_t *list,
2041                     int should_be_on)
2042 {
2043         if (cfs_list_empty(item) && should_be_on)
2044                 cfs_list_add_tail(item, list);
2045         else if (!cfs_list_empty(item) && !should_be_on)
2046                 cfs_list_del_init(item);
2047 }
2048
2049 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
2050  * can find pages to build into rpcs quickly */
2051 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
2052 {
2053         if (lop_makes_hprpc(&loi->loi_write_lop) ||
2054             lop_makes_hprpc(&loi->loi_read_lop)) {
2055                 /* HP rpc */
2056                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2057                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2058         } else {
2059                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2060                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2061                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2062                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2063         }
2064
2065         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2066                 loi->loi_write_lop.lop_num_pending);
2067
2068         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2069                 loi->loi_read_lop.lop_num_pending);
2070 }
2071
2072 static void lop_update_pending(struct client_obd *cli,
2073                                struct loi_oap_pages *lop, int cmd, int delta)
2074 {
2075         lop->lop_num_pending += delta;
2076         if (cmd & OBD_BRW_WRITE)
2077                 cli->cl_pending_w_pages += delta;
2078         else
2079                 cli->cl_pending_r_pages += delta;
2080 }
2081
2082 /**
2083  * this is called when a sync waiter receives an interruption.  Its job is to
2084  * get the caller woken as soon as possible.  If its page hasn't been put in an
2085  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2086  * desiring interruption which will forcefully complete the rpc once the rpc
2087  * has timed out.
2088  */
2089 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2090 {
2091         struct loi_oap_pages *lop;
2092         struct lov_oinfo *loi;
2093         int rc = -EBUSY;
2094         ENTRY;
2095
2096         LASSERT(!oap->oap_interrupted);
2097         oap->oap_interrupted = 1;
2098
2099         /* ok, it's been put in an rpc. only one oap gets a request reference */
2100         if (oap->oap_request != NULL) {
2101                 ptlrpc_mark_interrupted(oap->oap_request);
2102                 ptlrpcd_wake(oap->oap_request);
2103                 ptlrpc_req_finished(oap->oap_request);
2104                 oap->oap_request = NULL;
2105         }
2106
2107         /*
2108          * page completion may be called only if ->cpo_prep() method was
2109          * executed by osc_io_submit(), that also adds page the to pending list
2110          */
2111         if (!cfs_list_empty(&oap->oap_pending_item)) {
2112                 cfs_list_del_init(&oap->oap_pending_item);
2113                 cfs_list_del_init(&oap->oap_urgent_item);
2114
2115                 loi = oap->oap_loi;
2116                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2117                         &loi->loi_write_lop : &loi->loi_read_lop;
2118                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2119                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2120                 rc = oap->oap_caller_ops->ap_completion(env,
2121                                           oap->oap_caller_data,
2122                                           oap->oap_cmd, NULL, -EINTR);
2123         }
2124
2125         RETURN(rc);
2126 }
2127
2128 /* this is trying to propogate async writeback errors back up to the
2129  * application.  As an async write fails we record the error code for later if
2130  * the app does an fsync.  As long as errors persist we force future rpcs to be
2131  * sync so that the app can get a sync error and break the cycle of queueing
2132  * pages for which writeback will fail. */
2133 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2134                            int rc)
2135 {
2136         if (rc) {
2137                 if (!ar->ar_rc)
2138                         ar->ar_rc = rc;
2139
2140                 ar->ar_force_sync = 1;
2141                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2142                 return;
2143
2144         }
2145
2146         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2147                 ar->ar_force_sync = 0;
2148 }
2149
2150 void osc_oap_to_pending(struct osc_async_page *oap)
2151 {
2152         struct loi_oap_pages *lop;
2153
2154         if (oap->oap_cmd & OBD_BRW_WRITE)
2155                 lop = &oap->oap_loi->loi_write_lop;
2156         else
2157                 lop = &oap->oap_loi->loi_read_lop;
2158
2159         if (oap->oap_async_flags & ASYNC_HP)
2160                 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2161         else if (oap->oap_async_flags & ASYNC_URGENT)
2162                 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2163         cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2164         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2165 }
2166
2167 /* this must be called holding the loi list lock to give coverage to exit_cache,
2168  * async_flag maintenance, and oap_request */
2169 static void osc_ap_completion(const struct lu_env *env,
2170                               struct client_obd *cli, struct obdo *oa,
2171                               struct osc_async_page *oap, int sent, int rc)
2172 {
2173         __u64 xid = 0;
2174
2175         ENTRY;
2176         if (oap->oap_request != NULL) {
2177                 xid = ptlrpc_req_xid(oap->oap_request);
2178                 ptlrpc_req_finished(oap->oap_request);
2179                 oap->oap_request = NULL;
2180         }
2181
2182         cfs_spin_lock(&oap->oap_lock);
2183         oap->oap_async_flags = 0;
2184         cfs_spin_unlock(&oap->oap_lock);
2185         oap->oap_interrupted = 0;
2186
2187         if (oap->oap_cmd & OBD_BRW_WRITE) {
2188                 osc_process_ar(&cli->cl_ar, xid, rc);
2189                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2190         }
2191
2192         if (rc == 0 && oa != NULL) {
2193                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2194                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2195                 if (oa->o_valid & OBD_MD_FLMTIME)
2196                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2197                 if (oa->o_valid & OBD_MD_FLATIME)
2198                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2199                 if (oa->o_valid & OBD_MD_FLCTIME)
2200                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2201         }
2202
2203         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2204                                                 oap->oap_cmd, oa, rc);
2205
2206         /* ll_ap_completion (from llite) drops PG_locked. so, a new
2207          * I/O on the page could start, but OSC calls it under lock
2208          * and thus we can add oap back to pending safely */
2209         if (rc)
2210                 /* upper layer wants to leave the page on pending queue */
2211                 osc_oap_to_pending(oap);
2212         else
2213                 osc_exit_cache(cli, oap, sent);
2214         EXIT;
2215 }
2216
2217 static int brw_interpret(const struct lu_env *env,
2218                          struct ptlrpc_request *req, void *data, int rc)
2219 {
2220         struct osc_brw_async_args *aa = data;
2221         struct client_obd *cli;
2222         int async;
2223         ENTRY;
2224
2225         rc = osc_brw_fini_request(req, rc);
2226         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2227         if (osc_recoverable_error(rc)) {
2228                 /* Only retry once for mmaped files since the mmaped page
2229                  * might be modified at anytime. We have to retry at least
2230                  * once in case there WAS really a corruption of the page
2231                  * on the network, that was not caused by mmap() modifying
2232                  * the page. Bug11742 */
2233                 if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
2234                     aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
2235                     aa->aa_oa->o_flags & OBD_FL_MMAP) {
2236                         rc = 0;
2237                 } else {
2238                         rc = osc_brw_redo_request(req, aa);
2239                         if (rc == 0)
2240                                 RETURN(0);
2241                 }
2242         }
2243
2244         if (aa->aa_ocapa) {
2245                 capa_put(aa->aa_ocapa);
2246                 aa->aa_ocapa = NULL;
2247         }
2248
2249         cli = aa->aa_cli;
2250
2251         client_obd_list_lock(&cli->cl_loi_list_lock);
2252
2253         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2254          * is called so we know whether to go to sync BRWs or wait for more
2255          * RPCs to complete */
2256         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2257                 cli->cl_w_in_flight--;
2258         else
2259                 cli->cl_r_in_flight--;
2260
2261         async = cfs_list_empty(&aa->aa_oaps);
2262         if (!async) { /* from osc_send_oap_rpc() */
2263                 struct osc_async_page *oap, *tmp;
2264                 /* the caller may re-use the oap after the completion call so
2265                  * we need to clean it up a little */
2266                 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2267                                              oap_rpc_item) {
2268                         cfs_list_del_init(&oap->oap_rpc_item);
2269                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2270                 }
2271                 OBDO_FREE(aa->aa_oa);
2272         } else { /* from async_internal() */
2273                 obd_count i;
2274                 for (i = 0; i < aa->aa_page_count; i++)
2275                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2276         }
2277         osc_wake_cache_waiters(cli);
2278         osc_check_rpcs(env, cli);
2279         client_obd_list_unlock(&cli->cl_loi_list_lock);
2280         if (!async)
2281                 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2282                                   req->rq_bulk->bd_nob_transferred);
2283         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2284         ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
2285
2286         RETURN(rc);
2287 }
2288
2289 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2290                                             struct client_obd *cli,
2291                                             cfs_list_t *rpc_list,
2292                                             int page_count, int cmd)
2293 {
2294         struct ptlrpc_request *req;
2295         struct brw_page **pga = NULL;
2296         struct osc_brw_async_args *aa;
2297         struct obdo *oa = NULL;
2298         const struct obd_async_page_ops *ops = NULL;
2299         void *caller_data = NULL;
2300         struct osc_async_page *oap;
2301         struct osc_async_page *tmp;
2302         struct ost_body *body;
2303         struct cl_req *clerq = NULL;
2304         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2305         struct ldlm_lock *lock = NULL;
2306         struct cl_req_attr crattr;
2307         int i, rc, mpflag = 0;
2308
2309         ENTRY;
2310         LASSERT(!cfs_list_empty(rpc_list));
2311
2312         if (cmd & OBD_BRW_MEMALLOC)
2313                 mpflag = cfs_memory_pressure_get_and_set();
2314
2315         memset(&crattr, 0, sizeof crattr);
2316         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2317         if (pga == NULL)
2318                 GOTO(out, req = ERR_PTR(-ENOMEM));
2319
2320         OBDO_ALLOC(oa);
2321         if (oa == NULL)
2322                 GOTO(out, req = ERR_PTR(-ENOMEM));
2323
2324         i = 0;
2325         cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2326                 struct cl_page *page = osc_oap2cl_page(oap);
2327                 if (ops == NULL) {
2328                         ops = oap->oap_caller_ops;
2329                         caller_data = oap->oap_caller_data;
2330
2331                         clerq = cl_req_alloc(env, page, crt,
2332                                              1 /* only 1-object rpcs for
2333                                                 * now */);
2334                         if (IS_ERR(clerq))
2335                                 GOTO(out, req = (void *)clerq);
2336                         lock = oap->oap_ldlm_lock;
2337                 }
2338                 pga[i] = &oap->oap_brw_page;
2339                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2340                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2341                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2342                 i++;
2343                 cl_req_page_add(env, clerq, page);
2344         }
2345
2346         /* always get the data for the obdo for the rpc */
2347         LASSERT(ops != NULL);
2348         crattr.cra_oa = oa;
2349         crattr.cra_capa = NULL;
2350         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2351         if (lock) {
2352                 oa->o_handle = lock->l_remote_handle;
2353                 oa->o_valid |= OBD_MD_FLHANDLE;
2354         }
2355
2356         rc = cl_req_prep(env, clerq);
2357         if (rc != 0) {
2358                 CERROR("cl_req_prep failed: %d\n", rc);
2359                 GOTO(out, req = ERR_PTR(rc));
2360         }
2361
2362         sort_brw_pages(pga, page_count);
2363         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2364                                   pga, &req, crattr.cra_capa, 1, 0);
2365         if (rc != 0) {
2366                 CERROR("prep_req failed: %d\n", rc);
2367                 GOTO(out, req = ERR_PTR(rc));
2368         }
2369
2370         if (cmd & OBD_BRW_MEMALLOC)
2371                 req->rq_memalloc = 1;
2372
2373         /* Need to update the timestamps after the request is built in case
2374          * we race with setattr (locally or in queue at OST).  If OST gets
2375          * later setattr before earlier BRW (as determined by the request xid),
2376          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2377          * way to do this in a single call.  bug 10150 */
2378         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2379         cl_req_attr_set(env, clerq, &crattr,
2380                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2381
2382         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2383         aa = ptlrpc_req_async_args(req);
2384         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2385         cfs_list_splice(rpc_list, &aa->aa_oaps);
2386         CFS_INIT_LIST_HEAD(rpc_list);
2387         aa->aa_clerq = clerq;
2388 out:
2389         if (cmd & OBD_BRW_MEMALLOC)
2390                 cfs_memory_pressure_restore(mpflag);
2391
2392         capa_put(crattr.cra_capa);
2393         if (IS_ERR(req)) {
2394                 if (oa)
2395                         OBDO_FREE(oa);
2396                 if (pga)
2397                         OBD_FREE(pga, sizeof(*pga) * page_count);
2398                 /* this should happen rarely and is pretty bad, it makes the
2399                  * pending list not follow the dirty order */
2400                 client_obd_list_lock(&cli->cl_loi_list_lock);
2401                 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2402                         cfs_list_del_init(&oap->oap_rpc_item);
2403
2404                         /* queued sync pages can be torn down while the pages
2405                          * were between the pending list and the rpc */
2406                         if (oap->oap_interrupted) {
2407                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2408                                 osc_ap_completion(env, cli, NULL, oap, 0,
2409                                                   oap->oap_count);
2410                                 continue;
2411                         }
2412                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2413                 }
2414                 if (clerq && !IS_ERR(clerq))
2415                         cl_req_completion(env, clerq, PTR_ERR(req));
2416         }
2417         RETURN(req);
2418 }
2419
2420 /**
2421  * prepare pages for ASYNC io and put pages in send queue.
2422  *
2423  * \param cmd OBD_BRW_* macroses
2424  * \param lop pending pages
2425  *
2426  * \return zero if no page added to send queue.
2427  * \return 1 if pages successfully added to send queue.
2428  * \return negative on errors.
2429  */
2430 static int
2431 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2432                  struct lov_oinfo *loi,
2433                  int cmd, struct loi_oap_pages *lop)
2434 {
2435         struct ptlrpc_request *req;
2436         obd_count page_count = 0;
2437         struct osc_async_page *oap = NULL, *tmp;
2438         struct osc_brw_async_args *aa;
2439         const struct obd_async_page_ops *ops;
2440         CFS_LIST_HEAD(rpc_list);
2441         int srvlock = 0, mem_tight = 0;
2442         struct cl_object *clob = NULL;
2443         obd_off starting_offset = OBD_OBJECT_EOF;
2444         unsigned int ending_offset;
2445         int starting_page_off = 0;
2446         ENTRY;
2447
2448         /* ASYNC_HP pages first. At present, when the lock the pages is
2449          * to be canceled, the pages covered by the lock will be sent out
2450          * with ASYNC_HP. We have to send out them as soon as possible. */
2451         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2452                 if (oap->oap_async_flags & ASYNC_HP)
2453                         cfs_list_move(&oap->oap_pending_item, &lop->lop_pending);
2454                 if (++page_count >= cli->cl_max_pages_per_rpc)
2455                         break;
2456         }
2457         page_count = 0;
2458
2459         /* first we find the pages we're allowed to work with */
2460         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2461                                      oap_pending_item) {
2462                 ops = oap->oap_caller_ops;
2463
2464                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2465                          "magic 0x%x\n", oap, oap->oap_magic);
2466
2467                 if (clob == NULL) {
2468                         /* pin object in memory, so that completion call-backs
2469                          * can be safely called under client_obd_list lock. */
2470                         clob = osc_oap2cl_page(oap)->cp_obj;
2471                         cl_object_get(clob);
2472                 }
2473
2474                 if (page_count != 0 &&
2475                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2476                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2477                                " oap %p, page %p, srvlock %u\n",
2478                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2479                         break;
2480                 }
2481
2482                 /* If there is a gap at the start of this page, it can't merge
2483                  * with any previous page, so we'll hand the network a
2484                  * "fragmented" page array that it can't transfer in 1 RDMA */
2485                 if (oap->oap_obj_off < starting_offset) {
2486                         if (starting_page_off != 0)
2487                                 break;
2488
2489                         starting_page_off = oap->oap_page_off;
2490                         starting_offset = oap->oap_obj_off + starting_page_off;
2491                 } else if (oap->oap_page_off != 0)
2492                         break;
2493
2494                 /* in llite being 'ready' equates to the page being locked
2495                  * until completion unlocks it.  commit_write submits a page
2496                  * as not ready because its unlock will happen unconditionally
2497                  * as the call returns.  if we race with commit_write giving
2498                  * us that page we don't want to create a hole in the page
2499                  * stream, so we stop and leave the rpc to be fired by
2500                  * another dirtier or kupdated interval (the not ready page
2501                  * will still be on the dirty list).  we could call in
2502                  * at the end of ll_file_write to process the queue again. */
2503                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2504                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2505                                                     cmd);
2506                         if (rc < 0)
2507                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2508                                                 "instead of ready\n", oap,
2509                                                 oap->oap_page, rc);
2510                         switch (rc) {
2511                         case -EAGAIN:
2512                                 /* llite is telling us that the page is still
2513                                  * in commit_write and that we should try
2514                                  * and put it in an rpc again later.  we
2515                                  * break out of the loop so we don't create
2516                                  * a hole in the sequence of pages in the rpc
2517                                  * stream.*/
2518                                 oap = NULL;
2519                                 break;
2520                         case -EINTR:
2521                                 /* the io isn't needed.. tell the checks
2522                                  * below to complete the rpc with EINTR */
2523                                 cfs_spin_lock(&oap->oap_lock);
2524                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2525                                 cfs_spin_unlock(&oap->oap_lock);
2526                                 oap->oap_count = -EINTR;
2527                                 break;
2528                         case 0:
2529                                 cfs_spin_lock(&oap->oap_lock);
2530                                 oap->oap_async_flags |= ASYNC_READY;
2531                                 cfs_spin_unlock(&oap->oap_lock);
2532                                 break;
2533                         default:
2534                                 LASSERTF(0, "oap %p page %p returned %d "
2535                                             "from make_ready\n", oap,
2536                                             oap->oap_page, rc);
2537                                 break;
2538                         }
2539                 }
2540                 if (oap == NULL)
2541                         break;
2542                 /*
2543                  * Page submitted for IO has to be locked. Either by
2544                  * ->ap_make_ready() or by higher layers.
2545                  */
2546 #if defined(__KERNEL__) && defined(__linux__)
2547                 {
2548                         struct cl_page *page;
2549
2550                         page = osc_oap2cl_page(oap);
2551
2552                         if (page->cp_type == CPT_CACHEABLE &&
2553                             !(PageLocked(oap->oap_page) &&
2554                               (CheckWriteback(oap->oap_page, cmd)))) {
2555                                 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2556                                        oap->oap_page,
2557                                        (long)oap->oap_page->flags,
2558                                        oap->oap_async_flags);
2559                                 LBUG();
2560                         }
2561                 }
2562 #endif
2563
2564                 /* take the page out of our book-keeping */
2565                 cfs_list_del_init(&oap->oap_pending_item);
2566                 lop_update_pending(cli, lop, cmd, -1);
2567                 cfs_list_del_init(&oap->oap_urgent_item);
2568
2569                 /* ask the caller for the size of the io as the rpc leaves. */
2570                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2571                         oap->oap_count =
2572                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2573                                                       cmd);
2574                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2575                 }
2576                 if (oap->oap_count <= 0) {
2577                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2578                                oap->oap_count);
2579                         osc_ap_completion(env, cli, NULL,
2580                                           oap, 0, oap->oap_count);
2581                         continue;
2582                 }
2583
2584                 /* now put the page back in our accounting */
2585                 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2586                 if (page_count++ == 0)
2587                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2588
2589                 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2590                         mem_tight = 1;
2591
2592                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2593                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2594                  * have the same alignment as the initial writes that allocated
2595                  * extents on the server. */
2596                 ending_offset = oap->oap_obj_off + oap->oap_page_off +
2597                                 oap->oap_count;
2598                 if (!(ending_offset & (PTLRPC_MAX_BRW_SIZE - 1)))
2599                         break;
2600
2601                 if (page_count >= cli->cl_max_pages_per_rpc)
2602                         break;
2603
2604                 /* If there is a gap at the end of this page, it can't merge
2605                  * with any subsequent pages, so we'll hand the network a
2606                  * "fragmented" page array that it can't transfer in 1 RDMA */
2607                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2608                         break;
2609         }
2610
2611         osc_wake_cache_waiters(cli);
2612
2613         loi_list_maint(cli, loi);
2614
2615         client_obd_list_unlock(&cli->cl_loi_list_lock);
2616
2617         if (clob != NULL)
2618                 cl_object_put(env, clob);
2619
2620         if (page_count == 0) {
2621                 client_obd_list_lock(&cli->cl_loi_list_lock);
2622                 RETURN(0);
2623         }
2624
2625         req = osc_build_req(env, cli, &rpc_list, page_count,
2626                             mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2627         if (IS_ERR(req)) {
2628                 LASSERT(cfs_list_empty(&rpc_list));
2629                 loi_list_maint(cli, loi);
2630                 RETURN(PTR_ERR(req));
2631         }
2632
2633         aa = ptlrpc_req_async_args(req);
2634
2635         starting_offset &= PTLRPC_MAX_BRW_SIZE - 1;
2636         if (cmd == OBD_BRW_READ) {
2637                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2638                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2639                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2640                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2641         } else {
2642                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2643                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2644                                  cli->cl_w_in_flight);
2645                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2646                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2647         }
2648
2649         client_obd_list_lock(&cli->cl_loi_list_lock);
2650
2651         if (cmd == OBD_BRW_READ)
2652                 cli->cl_r_in_flight++;
2653         else
2654                 cli->cl_w_in_flight++;
2655
2656         /* queued sync pages can be torn down while the pages
2657          * were between the pending list and the rpc */
2658         tmp = NULL;
2659         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2660                 /* only one oap gets a request reference */
2661                 if (tmp == NULL)
2662                         tmp = oap;
2663                 if (oap->oap_interrupted && !req->rq_intr) {
2664                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2665                                oap, req);
2666                         ptlrpc_mark_interrupted(req);
2667                 }
2668         }
2669         if (tmp != NULL)
2670                 tmp->oap_request = ptlrpc_request_addref(req);
2671
2672         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2673                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2674
2675         req->rq_interpret_reply = brw_interpret;
2676
2677         /* XXX: Maybe the caller can check the RPC bulk descriptor to see which
2678          *      CPU/NUMA node the majority of pages were allocated on, and try
2679          *      to assign the async RPC to the CPU core (PDL_POLICY_PREFERRED)
2680          *      to reduce cross-CPU memory traffic.
2681          *
2682          *      But on the other hand, we expect that multiple ptlrpcd threads
2683          *      and the initial write sponsor can run in parallel, especially
2684          *      when data checksum is enabled, which is CPU-bound operation and
2685          *      single ptlrpcd thread cannot process in time. So more ptlrpcd
2686          *      threads sharing BRW load (with PDL_POLICY_ROUND) seems better.
2687          */
2688         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2689         RETURN(1);
2690 }
2691
2692 #define LOI_DEBUG(LOI, STR, args...)                                     \
2693         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2694                !cfs_list_empty(&(LOI)->loi_ready_item) ||                \
2695                !cfs_list_empty(&(LOI)->loi_hp_ready_item),               \
2696                (LOI)->loi_write_lop.lop_num_pending,                     \
2697                !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent),        \
2698                (LOI)->loi_read_lop.lop_num_pending,                      \
2699                !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent),         \
2700                args)                                                     \
2701
2702 /* This is called by osc_check_rpcs() to find which objects have pages that
2703  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2704 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2705 {
2706         ENTRY;
2707
2708         /* First return objects that have blocked locks so that they
2709          * will be flushed quickly and other clients can get the lock,
2710          * then objects which have pages ready to be stuffed into RPCs */
2711         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2712                 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2713                                       struct lov_oinfo, loi_hp_ready_item));
2714         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2715                 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2716                                       struct lov_oinfo, loi_ready_item));
2717
2718         /* then if we have cache waiters, return all objects with queued
2719          * writes.  This is especially important when many small files
2720          * have filled up the cache and not been fired into rpcs because
2721          * they don't pass the nr_pending/object threshhold */
2722         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2723             !cfs_list_empty(&cli->cl_loi_write_list))
2724                 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2725                                       struct lov_oinfo, loi_write_item));
2726
2727         /* then return all queued objects when we have an invalid import
2728          * so that they get flushed */
2729         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2730                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2731                         RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2732                                               struct lov_oinfo,
2733                                               loi_write_item));
2734                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2735                         RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2736                                               struct lov_oinfo, loi_read_item));
2737         }
2738         RETURN(NULL);
2739 }
2740
2741 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2742 {
2743         struct osc_async_page *oap;
2744         int hprpc = 0;
2745
2746         if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2747                 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2748                                      struct osc_async_page, oap_urgent_item);
2749                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2750         }
2751
2752         if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2753                 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2754                                      struct osc_async_page, oap_urgent_item);
2755                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2756         }
2757
2758         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2759 }
2760
2761 /* called with the loi list lock held */
2762 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2763 {
2764         struct lov_oinfo *loi;
2765         int rc = 0, race_counter = 0;
2766         ENTRY;
2767
2768         while ((loi = osc_next_loi(cli)) != NULL) {
2769                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2770
2771                 if (osc_max_rpc_in_flight(cli, loi))
2772                         break;
2773
2774                 /* attempt some read/write balancing by alternating between
2775                  * reads and writes in an object.  The makes_rpc checks here
2776                  * would be redundant if we were getting read/write work items
2777                  * instead of objects.  we don't want send_oap_rpc to drain a
2778                  * partial read pending queue when we're given this object to
2779                  * do io on writes while there are cache waiters */
2780                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2781                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2782                                               &loi->loi_write_lop);
2783                         if (rc < 0) {
2784                                 CERROR("Write request failed with %d\n", rc);
2785
2786                                 /* osc_send_oap_rpc failed, mostly because of
2787                                  * memory pressure.
2788                                  *
2789                                  * It can't break here, because if:
2790                                  *  - a page was submitted by osc_io_submit, so
2791                                  *    page locked;
2792                                  *  - no request in flight
2793                                  *  - no subsequent request
2794                                  * The system will be in live-lock state,
2795                                  * because there is no chance to call
2796                                  * osc_io_unplug() and osc_check_rpcs() any
2797                                  * more. pdflush can't help in this case,
2798                                  * because it might be blocked at grabbing
2799                                  * the page lock as we mentioned.
2800                                  *
2801                                  * Anyway, continue to drain pages. */
2802                                 /* break; */
2803                         }
2804
2805                         if (rc > 0)
2806                                 race_counter = 0;
2807                         else if (rc == 0)
2808                                 race_counter++;
2809                 }
2810                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2811                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2812                                               &loi->loi_read_lop);
2813                         if (rc < 0)
2814                                 CERROR("Read request failed with %d\n", rc);
2815
2816                         if (rc > 0)
2817                                 race_counter = 0;
2818                         else if (rc == 0)
2819                                 race_counter++;
2820                 }
2821
2822                 /* attempt some inter-object balancing by issuing rpcs
2823                  * for each object in turn */
2824                 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2825                         cfs_list_del_init(&loi->loi_hp_ready_item);
2826                 if (!cfs_list_empty(&loi->loi_ready_item))
2827                         cfs_list_del_init(&loi->loi_ready_item);
2828                 if (!cfs_list_empty(&loi->loi_write_item))
2829                         cfs_list_del_init(&loi->loi_write_item);
2830                 if (!cfs_list_empty(&loi->loi_read_item))
2831                         cfs_list_del_init(&loi->loi_read_item);
2832
2833                 loi_list_maint(cli, loi);
2834
2835                 /* send_oap_rpc fails with 0 when make_ready tells it to
2836                  * back off.  llite's make_ready does this when it tries
2837                  * to lock a page queued for write that is already locked.
2838                  * we want to try sending rpcs from many objects, but we
2839                  * don't want to spin failing with 0.  */
2840                 if (race_counter == 10)
2841                         break;
2842         }
2843         EXIT;
2844 }
2845
2846 /* we're trying to queue a page in the osc so we're subject to the
2847  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2848  * If the osc's queued pages are already at that limit, then we want to sleep
2849  * until there is space in the osc's queue for us.  We also may be waiting for
2850  * write credits from the OST if there are RPCs in flight that may return some
2851  * before we fall back to sync writes.
2852  *
2853  * We need this know our allocation was granted in the presence of signals */
2854 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2855 {
2856         int rc;
2857         ENTRY;
2858         client_obd_list_lock(&cli->cl_loi_list_lock);
2859         rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2860         client_obd_list_unlock(&cli->cl_loi_list_lock);
2861         RETURN(rc);
2862 };
2863
2864 /**
2865  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2866  * is available.
2867  */
2868 int osc_enter_cache_try(const struct lu_env *env,
2869                         struct client_obd *cli, struct lov_oinfo *loi,
2870                         struct osc_async_page *oap, int transient)
2871 {
2872         int has_grant;
2873
2874         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2875         if (has_grant) {
2876                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2877                 if (transient) {
2878                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2879                         cfs_atomic_inc(&obd_dirty_transit_pages);
2880                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2881                 }
2882         }
2883         return has_grant;
2884 }
2885
2886 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2887  * grant or cache space. */
2888 static int osc_enter_cache(const struct lu_env *env,
2889                            struct client_obd *cli, struct lov_oinfo *loi,
2890                            struct osc_async_page *oap)
2891 {
2892         struct osc_cache_waiter ocw;
2893         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
2894
2895         ENTRY;
2896
2897         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2898                "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2899                cli->cl_dirty_max, obd_max_dirty_pages,
2900                cli->cl_lost_grant, cli->cl_avail_grant);
2901
2902         /* force the caller to try sync io.  this can jump the list
2903          * of queued writes and create a discontiguous rpc stream */
2904         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
2905             cli->cl_dirty_max < CFS_PAGE_SIZE     ||
2906             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
2907                 RETURN(-EDQUOT);
2908
2909         /* Hopefully normal case - cache space and write credits available */
2910         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2911             cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2912             osc_enter_cache_try(env, cli, loi, oap, 0))
2913                 RETURN(0);
2914
2915         /* It is safe to block as a cache waiter as long as there is grant
2916          * space available or the hope of additional grant being returned
2917          * when an in flight write completes.  Using the write back cache
2918          * if possible is preferable to sending the data synchronously
2919          * because write pages can then be merged in to large requests.
2920          * The addition of this cache waiter will causing pending write
2921          * pages to be sent immediately. */
2922         if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2923                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2924                 cfs_waitq_init(&ocw.ocw_waitq);
2925                 ocw.ocw_oap = oap;
2926                 ocw.ocw_rc = 0;
2927
2928                 loi_list_maint(cli, loi);
2929                 osc_check_rpcs(env, cli);
2930                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2931
2932                 CDEBUG(D_CACHE, "sleeping for cache space\n");
2933                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2934
2935                 client_obd_list_lock(&cli->cl_loi_list_lock);
2936                 if (!cfs_list_empty(&ocw.ocw_entry)) {
2937                         cfs_list_del(&ocw.ocw_entry);
2938                         RETURN(-EINTR);
2939                 }
2940                 RETURN(ocw.ocw_rc);
2941         }
2942
2943         RETURN(-EDQUOT);
2944 }
2945
2946
2947 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2948                         struct lov_oinfo *loi, cfs_page_t *page,
2949                         obd_off offset, const struct obd_async_page_ops *ops,
2950                         void *data, void **res, int nocache,
2951                         struct lustre_handle *lockh)
2952 {
2953         struct osc_async_page *oap;
2954
2955         ENTRY;
2956
2957         if (!page)
2958                 return cfs_size_round(sizeof(*oap));
2959
2960         oap = *res;
2961         oap->oap_magic = OAP_MAGIC;
2962         oap->oap_cli = &exp->exp_obd->u.cli;
2963         oap->oap_loi = loi;
2964
2965         oap->oap_caller_ops = ops;
2966         oap->oap_caller_data = data;
2967
2968         oap->oap_page = page;
2969         oap->oap_obj_off = offset;
2970         if (!client_is_remote(exp) &&
2971             cfs_capable(CFS_CAP_SYS_RESOURCE))
2972                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2973
2974         LASSERT(!(offset & ~CFS_PAGE_MASK));
2975
2976         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2977         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2978         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2979         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2980
2981         cfs_spin_lock_init(&oap->oap_lock);
2982         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2983         RETURN(0);
2984 }
2985
2986 int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
2987                        struct lov_stripe_md *lsm, struct lov_oinfo *loi,
2988                        struct osc_async_page *oap, int cmd, int off,
2989                        int count, obd_flag brw_flags, enum async_flags async_flags)
2990 {
2991         struct client_obd *cli = &exp->exp_obd->u.cli;
2992         int rc = 0;
2993         ENTRY;
2994
2995         if (oap->oap_magic != OAP_MAGIC)
2996                 RETURN(-EINVAL);
2997
2998         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2999                 RETURN(-EIO);
3000
3001         if (!cfs_list_empty(&oap->oap_pending_item) ||
3002             !cfs_list_empty(&oap->oap_urgent_item) ||
3003             !cfs_list_empty(&oap->oap_rpc_item))
3004                 RETURN(-EBUSY);
3005
3006         /* check if the file's owner/group is over quota */
3007         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
3008                 struct cl_object *obj;
3009                 struct cl_attr    attr; /* XXX put attr into thread info */
3010                 unsigned int qid[MAXQUOTAS];
3011
3012                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
3013
3014                 cl_object_attr_lock(obj);
3015                 rc = cl_object_attr_get(env, obj, &attr);
3016                 cl_object_attr_unlock(obj);
3017
3018                 qid[USRQUOTA] = attr.cat_uid;
3019                 qid[GRPQUOTA] = attr.cat_gid;
3020                 if (rc == 0 &&
3021                     lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
3022                         rc = -EDQUOT;
3023                 if (rc)
3024                         RETURN(rc);
3025         }
3026
3027         if (loi == NULL)
3028                 loi = lsm->lsm_oinfo[0];
3029
3030         client_obd_list_lock(&cli->cl_loi_list_lock);
3031
3032         LASSERT(off + count <= CFS_PAGE_SIZE);
3033         oap->oap_cmd = cmd;
3034         oap->oap_page_off = off;
3035         oap->oap_count = count;
3036         oap->oap_brw_flags = brw_flags;
3037         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
3038         if (cfs_memory_pressure_get())
3039                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
3040         cfs_spin_lock(&oap->oap_lock);
3041         oap->oap_async_flags = async_flags;
3042         cfs_spin_unlock(&oap->oap_lock);
3043
3044         if (cmd & OBD_BRW_WRITE) {
3045                 rc = osc_enter_cache(env, cli, loi, oap);
3046                 if (rc) {
3047                         client_obd_list_unlock(&cli->cl_loi_list_lock);
3048                         RETURN(rc);
3049                 }
3050         }
3051
3052         osc_oap_to_pending(oap);
3053         loi_list_maint(cli, loi);
3054
3055         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3056                   cmd);
3057
3058         osc_check_rpcs(env, cli);
3059         client_obd_list_unlock(&cli->cl_loi_list_lock);
3060
3061         RETURN(0);
3062 }
3063
3064 /* aka (~was & now & flag), but this is more clear :) */
3065 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3066
3067 int osc_set_async_flags_base(struct client_obd *cli,
3068                              struct lov_oinfo *loi, struct osc_async_page *oap,
3069                              obd_flag async_flags)
3070 {
3071         struct loi_oap_pages *lop;
3072         int flags = 0;
3073         ENTRY;
3074
3075         LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3076
3077         if (oap->oap_cmd & OBD_BRW_WRITE) {
3078                 lop = &loi->loi_write_lop;
3079         } else {
3080                 lop = &loi->loi_read_lop;
3081         }
3082
3083         if ((oap->oap_async_flags & async_flags) == async_flags)
3084                 RETURN(0);
3085
3086         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3087                 flags |= ASYNC_READY;
3088
3089         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3090             cfs_list_empty(&oap->oap_rpc_item)) {
3091                 if (oap->oap_async_flags & ASYNC_HP)
3092                         cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3093                 else
3094                         cfs_list_add_tail(&oap->oap_urgent_item,
3095                                           &lop->lop_urgent);
3096                 flags |= ASYNC_URGENT;
3097                 loi_list_maint(cli, loi);
3098         }
3099         cfs_spin_lock(&oap->oap_lock);
3100         oap->oap_async_flags |= flags;
3101         cfs_spin_unlock(&oap->oap_lock);
3102
3103         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3104                         oap->oap_async_flags);
3105         RETURN(0);
3106 }
3107
3108 int osc_teardown_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
3109                             struct lov_oinfo *loi, struct osc_async_page *oap)
3110 {
3111         struct client_obd *cli = &exp->exp_obd->u.cli;
3112         struct loi_oap_pages *lop;
3113         int rc = 0;
3114         ENTRY;
3115
3116         if (oap->oap_magic != OAP_MAGIC)
3117                 RETURN(-EINVAL);
3118
3119         if (loi == NULL)
3120                 loi = lsm->lsm_oinfo[0];
3121
3122         if (oap->oap_cmd & OBD_BRW_WRITE) {
3123                 lop = &loi->loi_write_lop;
3124         } else {
3125                 lop = &loi->loi_read_lop;
3126         }
3127
3128         client_obd_list_lock(&cli->cl_loi_list_lock);
3129
3130         if (!cfs_list_empty(&oap->oap_rpc_item))
3131                 GOTO(out, rc = -EBUSY);
3132
3133         osc_exit_cache(cli, oap, 0);
3134         osc_wake_cache_waiters(cli);
3135
3136         if (!cfs_list_empty(&oap->oap_urgent_item)) {
3137                 cfs_list_del_init(&oap->oap_urgent_item);
3138                 cfs_spin_lock(&oap->oap_lock);
3139                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3140                 cfs_spin_unlock(&oap->oap_lock);
3141         }
3142         if (!cfs_list_empty(&oap->oap_pending_item)) {
3143                 cfs_list_del_init(&oap->oap_pending_item);
3144                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3145         }
3146         loi_list_maint(cli, loi);
3147         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3148 out:
3149         client_obd_list_unlock(&cli->cl_loi_list_lock);
3150         RETURN(rc);
3151 }
3152
3153 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
3154                                         struct ldlm_enqueue_info *einfo)
3155 {
3156         void *data = einfo->ei_cbdata;
3157         int set = 0;
3158
3159         LASSERT(lock != NULL);