Whamcloud - gitweb
b=21656 Quiet dirty_max CERROR
[fs/lustre-release.git] / lustre / osc / osc_request.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 #ifndef EXPORT_SYMTAB
38 # define EXPORT_SYMTAB
39 #endif
40 #define DEBUG_SUBSYSTEM S_OSC
41
42 #include <libcfs/libcfs.h>
43
44 #ifndef __KERNEL__
45 # include <liblustre.h>
46 #endif
47
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre/lustre_user.h>
51 #include <obd_cksum.h>
52 #include <obd_ost.h>
53 #include <obd_lov.h>
54
55 #ifdef  __CYGWIN__
56 # include <ctype.h>
57 #endif
58
59 #include <lustre_ha.h>
60 #include <lprocfs_status.h>
61 #include <lustre_log.h>
62 #include <lustre_debug.h>
63 #include <lustre_param.h>
64 #include "osc_internal.h"
65
66 static quota_interface_t *quota_interface = NULL;
67 extern quota_interface_t osc_quota_interface;
68
69 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
70 static int brw_interpret(const struct lu_env *env,
71                          struct ptlrpc_request *req, void *data, int rc);
72 int osc_cleanup(struct obd_device *obd);
73
74 /* Pack OSC object metadata for disk storage (LE byte order). */
75 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
76                       struct lov_stripe_md *lsm)
77 {
78         int lmm_size;
79         ENTRY;
80
81         lmm_size = sizeof(**lmmp);
82         if (!lmmp)
83                 RETURN(lmm_size);
84
85         if (*lmmp && !lsm) {
86                 OBD_FREE(*lmmp, lmm_size);
87                 *lmmp = NULL;
88                 RETURN(0);
89         }
90
91         if (!*lmmp) {
92                 OBD_ALLOC(*lmmp, lmm_size);
93                 if (!*lmmp)
94                         RETURN(-ENOMEM);
95         }
96
97         if (lsm) {
98                 LASSERT(lsm->lsm_object_id);
99                 LASSERT_MDS_GROUP(lsm->lsm_object_gr);
100                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
101                 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
102         }
103
104         RETURN(lmm_size);
105 }
106
107 /* Unpack OSC object metadata from disk storage (LE byte order). */
108 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
109                         struct lov_mds_md *lmm, int lmm_bytes)
110 {
111         int lsm_size;
112         ENTRY;
113
114         if (lmm != NULL) {
115                 if (lmm_bytes < sizeof (*lmm)) {
116                         CERROR("lov_mds_md too small: %d, need %d\n",
117                                lmm_bytes, (int)sizeof(*lmm));
118                         RETURN(-EINVAL);
119                 }
120                 /* XXX LOV_MAGIC etc check? */
121
122                 if (lmm->lmm_object_id == 0) {
123                         CERROR("lov_mds_md: zero lmm_object_id\n");
124                         RETURN(-EINVAL);
125                 }
126         }
127
128         lsm_size = lov_stripe_md_size(1);
129         if (lsmp == NULL)
130                 RETURN(lsm_size);
131
132         if (*lsmp != NULL && lmm == NULL) {
133                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
134                 OBD_FREE(*lsmp, lsm_size);
135                 *lsmp = NULL;
136                 RETURN(0);
137         }
138
139         if (*lsmp == NULL) {
140                 OBD_ALLOC(*lsmp, lsm_size);
141                 if (*lsmp == NULL)
142                         RETURN(-ENOMEM);
143                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
144                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
145                         OBD_FREE(*lsmp, lsm_size);
146                         RETURN(-ENOMEM);
147                 }
148                 loi_init((*lsmp)->lsm_oinfo[0]);
149         }
150
151         if (lmm != NULL) {
152                 /* XXX zero *lsmp? */
153                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
154                 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
155                 LASSERT((*lsmp)->lsm_object_id);
156                 LASSERT_MDS_GROUP((*lsmp)->lsm_object_gr);
157         }
158
159         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
160
161         RETURN(lsm_size);
162 }
163
164 static inline void osc_pack_capa(struct ptlrpc_request *req,
165                                  struct ost_body *body, void *capa)
166 {
167         struct obd_capa *oc = (struct obd_capa *)capa;
168         struct lustre_capa *c;
169
170         if (!capa)
171                 return;
172
173         c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
174         LASSERT(c);
175         capa_cpy(c, oc);
176         body->oa.o_valid |= OBD_MD_FLOSSCAPA;
177         DEBUG_CAPA(D_SEC, c, "pack");
178 }
179
180 static inline void osc_pack_req_body(struct ptlrpc_request *req,
181                                      struct obd_info *oinfo)
182 {
183         struct ost_body *body;
184
185         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
186         LASSERT(body);
187
188         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
189         osc_pack_capa(req, body, oinfo->oi_capa);
190 }
191
192 static inline void osc_set_capa_size(struct ptlrpc_request *req,
193                                      const struct req_msg_field *field,
194                                      struct obd_capa *oc)
195 {
196         if (oc == NULL)
197                 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
198         else
199                 /* it is already calculated as sizeof struct obd_capa */
200                 ;
201 }
202
203 static int osc_getattr_interpret(const struct lu_env *env,
204                                  struct ptlrpc_request *req,
205                                  struct osc_async_args *aa, int rc)
206 {
207         struct ost_body *body;
208         ENTRY;
209
210         if (rc != 0)
211                 GOTO(out, rc);
212
213         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
214         if (body) {
215                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
216                 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
217
218                 /* This should really be sent by the OST */
219                 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
220                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
221         } else {
222                 CDEBUG(D_INFO, "can't unpack ost_body\n");
223                 rc = -EPROTO;
224                 aa->aa_oi->oi_oa->o_valid = 0;
225         }
226 out:
227         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
228         RETURN(rc);
229 }
230
231 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
232                              struct ptlrpc_request_set *set)
233 {
234         struct ptlrpc_request *req;
235         struct osc_async_args *aa;
236         int                    rc;
237         ENTRY;
238
239         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
240         if (req == NULL)
241                 RETURN(-ENOMEM);
242
243         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
244         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
245         if (rc) {
246                 ptlrpc_request_free(req);
247                 RETURN(rc);
248         }
249
250         osc_pack_req_body(req, oinfo);
251
252         ptlrpc_request_set_replen(req);
253         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
254
255         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
256         aa = ptlrpc_req_async_args(req);
257         aa->aa_oi = oinfo;
258
259         ptlrpc_set_add_req(set, req);
260         RETURN(0);
261 }
262
263 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
264 {
265         struct ptlrpc_request *req;
266         struct ost_body       *body;
267         int                    rc;
268         ENTRY;
269
270         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
271         if (req == NULL)
272                 RETURN(-ENOMEM);
273
274         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
275         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
276         if (rc) {
277                 ptlrpc_request_free(req);
278                 RETURN(rc);
279         }
280
281         osc_pack_req_body(req, oinfo);
282
283         ptlrpc_request_set_replen(req);
284
285         rc = ptlrpc_queue_wait(req);
286         if (rc)
287                 GOTO(out, rc);
288
289         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
290         if (body == NULL)
291                 GOTO(out, rc = -EPROTO);
292
293         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
294         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
295
296         /* This should really be sent by the OST */
297         oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
298         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
299
300         EXIT;
301  out:
302         ptlrpc_req_finished(req);
303         return rc;
304 }
305
306 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
307                        struct obd_trans_info *oti)
308 {
309         struct ptlrpc_request *req;
310         struct ost_body       *body;
311         int                    rc;
312         ENTRY;
313
314         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
315
316         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
317         if (req == NULL)
318                 RETURN(-ENOMEM);
319
320         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
321         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
322         if (rc) {
323                 ptlrpc_request_free(req);
324                 RETURN(rc);
325         }
326
327         osc_pack_req_body(req, oinfo);
328
329         ptlrpc_request_set_replen(req);
330
331         rc = ptlrpc_queue_wait(req);
332         if (rc)
333                 GOTO(out, rc);
334
335         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
336         if (body == NULL)
337                 GOTO(out, rc = -EPROTO);
338
339         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
340
341         EXIT;
342 out:
343         ptlrpc_req_finished(req);
344         RETURN(rc);
345 }
346
347 static int osc_setattr_interpret(const struct lu_env *env,
348                                  struct ptlrpc_request *req,
349                                  struct osc_async_args *aa, int rc)
350 {
351         struct ost_body *body;
352         ENTRY;
353
354         if (rc != 0)
355                 GOTO(out, rc);
356
357         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
358         if (body == NULL)
359                 GOTO(out, rc = -EPROTO);
360
361         lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
362 out:
363         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
364         RETURN(rc);
365 }
366
367 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
368                              struct obd_trans_info *oti,
369                              struct ptlrpc_request_set *rqset)
370 {
371         struct ptlrpc_request *req;
372         struct osc_async_args *aa;
373         int                    rc;
374         ENTRY;
375
376         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
377         if (req == NULL)
378                 RETURN(-ENOMEM);
379
380         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
381         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
382         if (rc) {
383                 ptlrpc_request_free(req);
384                 RETURN(rc);
385         }
386
387         if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
388                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
389
390         osc_pack_req_body(req, oinfo);
391
392         ptlrpc_request_set_replen(req);
393
394         /* do mds to ost setattr asynchronously */
395         if (!rqset) {
396                 /* Do not wait for response. */
397                 ptlrpcd_add_req(req, PSCOPE_OTHER);
398         } else {
399                 req->rq_interpret_reply =
400                         (ptlrpc_interpterer_t)osc_setattr_interpret;
401
402                 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
403                 aa = ptlrpc_req_async_args(req);
404                 aa->aa_oi = oinfo;
405
406                 ptlrpc_set_add_req(rqset, req);
407         }
408
409         RETURN(0);
410 }
411
412 int osc_real_create(struct obd_export *exp, struct obdo *oa,
413                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
414 {
415         struct ptlrpc_request *req;
416         struct ost_body       *body;
417         struct lov_stripe_md  *lsm;
418         int                    rc;
419         ENTRY;
420
421         LASSERT(oa);
422         LASSERT(ea);
423
424         lsm = *ea;
425         if (!lsm) {
426                 rc = obd_alloc_memmd(exp, &lsm);
427                 if (rc < 0)
428                         RETURN(rc);
429         }
430
431         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
432         if (req == NULL)
433                 GOTO(out, rc = -ENOMEM);
434
435         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
436         if (rc) {
437                 ptlrpc_request_free(req);
438                 GOTO(out, rc);
439         }
440
441         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
442         LASSERT(body);
443         lustre_set_wire_obdo(&body->oa, oa);
444
445         ptlrpc_request_set_replen(req);
446
447         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
448             oa->o_flags == OBD_FL_DELORPHAN) {
449                 DEBUG_REQ(D_HA, req,
450                           "delorphan from OST integration");
451                 /* Don't resend the delorphan req */
452                 req->rq_no_resend = req->rq_no_delay = 1;
453         }
454
455         rc = ptlrpc_queue_wait(req);
456         if (rc)
457                 GOTO(out_req, rc);
458
459         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
460         if (body == NULL)
461                 GOTO(out_req, rc = -EPROTO);
462
463         lustre_get_wire_obdo(oa, &body->oa);
464
465         /* This should really be sent by the OST */
466         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
467         oa->o_valid |= OBD_MD_FLBLKSZ;
468
469         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
470          * have valid lsm_oinfo data structs, so don't go touching that.
471          * This needs to be fixed in a big way.
472          */
473         lsm->lsm_object_id = oa->o_id;
474         lsm->lsm_object_gr = oa->o_gr;
475         *ea = lsm;
476
477         if (oti != NULL) {
478                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
479
480                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
481                         if (!oti->oti_logcookies)
482                                 oti_alloc_cookies(oti, 1);
483                         *oti->oti_logcookies = oa->o_lcookie;
484                 }
485         }
486
487         CDEBUG(D_HA, "transno: "LPD64"\n",
488                lustre_msg_get_transno(req->rq_repmsg));
489 out_req:
490         ptlrpc_req_finished(req);
491 out:
492         if (rc && !*ea)
493                 obd_free_memmd(exp, &lsm);
494         RETURN(rc);
495 }
496
497 static int osc_punch_interpret(const struct lu_env *env,
498                                struct ptlrpc_request *req,
499                                struct osc_punch_args *aa, int rc)
500 {
501         struct ost_body *body;
502         ENTRY;
503
504         if (rc != 0)
505                 GOTO(out, rc);
506
507         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
508         if (body == NULL)
509                 GOTO(out, rc = -EPROTO);
510
511         lustre_get_wire_obdo(aa->pa_oa, &body->oa);
512 out:
513         rc = aa->pa_upcall(aa->pa_cookie, rc);
514         RETURN(rc);
515 }
516
517 int osc_punch_base(struct obd_export *exp, struct obdo *oa,
518                    struct obd_capa *capa,
519                    obd_enqueue_update_f upcall, void *cookie,
520                    struct ptlrpc_request_set *rqset)
521 {
522         struct ptlrpc_request *req;
523         struct osc_punch_args *aa;
524         struct ost_body       *body;
525         int                    rc;
526         ENTRY;
527
528         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
529         if (req == NULL)
530                 RETURN(-ENOMEM);
531
532         osc_set_capa_size(req, &RMF_CAPA1, capa);
533         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
534         if (rc) {
535                 ptlrpc_request_free(req);
536                 RETURN(rc);
537         }
538         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
539         ptlrpc_at_set_req_timeout(req);
540
541         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
542         LASSERT(body);
543         lustre_set_wire_obdo(&body->oa, oa);
544         osc_pack_capa(req, body, capa);
545
546         ptlrpc_request_set_replen(req);
547
548
549         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_punch_interpret;
550         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
551         aa = ptlrpc_req_async_args(req);
552         aa->pa_oa     = oa;
553         aa->pa_upcall = upcall;
554         aa->pa_cookie = cookie;
555         if (rqset == PTLRPCD_SET)
556                 ptlrpcd_add_req(req, PSCOPE_OTHER);
557         else
558                 ptlrpc_set_add_req(rqset, req);
559
560         RETURN(0);
561 }
562
563 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
564                      struct obd_trans_info *oti,
565                      struct ptlrpc_request_set *rqset)
566 {
567         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
568         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
569         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
570         return osc_punch_base(exp, oinfo->oi_oa, oinfo->oi_capa,
571                               oinfo->oi_cb_up, oinfo, rqset);
572 }
573
574 static int osc_sync(struct obd_export *exp, struct obdo *oa,
575                     struct lov_stripe_md *md, obd_size start, obd_size end,
576                     void *capa)
577 {
578         struct ptlrpc_request *req;
579         struct ost_body       *body;
580         int                    rc;
581         ENTRY;
582
583         if (!oa) {
584                 CDEBUG(D_INFO, "oa NULL\n");
585                 RETURN(-EINVAL);
586         }
587
588         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
589         if (req == NULL)
590                 RETURN(-ENOMEM);
591
592         osc_set_capa_size(req, &RMF_CAPA1, capa);
593         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
594         if (rc) {
595                 ptlrpc_request_free(req);
596                 RETURN(rc);
597         }
598
599         /* overload the size and blocks fields in the oa with start/end */
600         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
601         LASSERT(body);
602         lustre_set_wire_obdo(&body->oa, oa);
603         body->oa.o_size = start;
604         body->oa.o_blocks = end;
605         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
606         osc_pack_capa(req, body, capa);
607
608         ptlrpc_request_set_replen(req);
609
610         rc = ptlrpc_queue_wait(req);
611         if (rc)
612                 GOTO(out, rc);
613
614         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
615         if (body == NULL)
616                 GOTO(out, rc = -EPROTO);
617
618         lustre_get_wire_obdo(oa, &body->oa);
619
620         EXIT;
621  out:
622         ptlrpc_req_finished(req);
623         return rc;
624 }
625
626 /* Find and cancel locally locks matched by @mode in the resource found by
627  * @objid. Found locks are added into @cancel list. Returns the amount of
628  * locks added to @cancels list. */
629 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
630                                    struct list_head *cancels, ldlm_mode_t mode,
631                                    int lock_flags)
632 {
633         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
634         struct ldlm_res_id res_id;
635         struct ldlm_resource *res;
636         int count;
637         ENTRY;
638
639         osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
640         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
641         if (res == NULL)
642                 RETURN(0);
643
644         LDLM_RESOURCE_ADDREF(res);
645         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
646                                            lock_flags, 0, NULL);
647         LDLM_RESOURCE_DELREF(res);
648         ldlm_resource_putref(res);
649         RETURN(count);
650 }
651
652 static int osc_destroy_interpret(const struct lu_env *env,
653                                  struct ptlrpc_request *req, void *data,
654                                  int rc)
655 {
656         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
657
658         atomic_dec(&cli->cl_destroy_in_flight);
659         cfs_waitq_signal(&cli->cl_destroy_waitq);
660         return 0;
661 }
662
663 static int osc_can_send_destroy(struct client_obd *cli)
664 {
665         if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
666             cli->cl_max_rpcs_in_flight) {
667                 /* The destroy request can be sent */
668                 return 1;
669         }
670         if (atomic_dec_return(&cli->cl_destroy_in_flight) <
671             cli->cl_max_rpcs_in_flight) {
672                 /*
673                  * The counter has been modified between the two atomic
674                  * operations.
675                  */
676                 cfs_waitq_signal(&cli->cl_destroy_waitq);
677         }
678         return 0;
679 }
680
681 /* Destroy requests can be async always on the client, and we don't even really
682  * care about the return code since the client cannot do anything at all about
683  * a destroy failure.
684  * When the MDS is unlinking a filename, it saves the file objects into a
685  * recovery llog, and these object records are cancelled when the OST reports
686  * they were destroyed and sync'd to disk (i.e. transaction committed).
687  * If the client dies, or the OST is down when the object should be destroyed,
688  * the records are not cancelled, and when the OST reconnects to the MDS next,
689  * it will retrieve the llog unlink logs and then sends the log cancellation
690  * cookies to the MDS after committing destroy transactions. */
691 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
692                        struct lov_stripe_md *ea, struct obd_trans_info *oti,
693                        struct obd_export *md_export, void *capa)
694 {
695         struct client_obd     *cli = &exp->exp_obd->u.cli;
696         struct ptlrpc_request *req;
697         struct ost_body       *body;
698         CFS_LIST_HEAD(cancels);
699         int rc, count;
700         ENTRY;
701
702         if (!oa) {
703                 CDEBUG(D_INFO, "oa NULL\n");
704                 RETURN(-EINVAL);
705         }
706
707         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
708                                         LDLM_FL_DISCARD_DATA);
709
710         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
711         if (req == NULL) {
712                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
713                 RETURN(-ENOMEM);
714         }
715
716         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
717         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
718                                0, &cancels, count);
719         if (rc) {
720                 ptlrpc_request_free(req);
721                 RETURN(rc);
722         }
723
724         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
725         ptlrpc_at_set_req_timeout(req);
726
727         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
728                 oa->o_lcookie = *oti->oti_logcookies;
729         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
730         LASSERT(body);
731         lustre_set_wire_obdo(&body->oa, oa);
732
733         osc_pack_capa(req, body, (struct obd_capa *)capa);
734         ptlrpc_request_set_replen(req);
735
736         /* don't throttle destroy RPCs for the MDT */
737         if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
738                 req->rq_interpret_reply = osc_destroy_interpret;
739                 if (!osc_can_send_destroy(cli)) {
740                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
741                                                           NULL);
742
743                         /*
744                          * Wait until the number of on-going destroy RPCs drops
745                          * under max_rpc_in_flight
746                          */
747                         l_wait_event_exclusive(cli->cl_destroy_waitq,
748                                                osc_can_send_destroy(cli), &lwi);
749                 }
750         }
751
752         /* Do not wait for response */
753         ptlrpcd_add_req(req, PSCOPE_OTHER);
754         RETURN(0);
755 }
756
757 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
758                                 long writing_bytes)
759 {
760         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
761
762         LASSERT(!(oa->o_valid & bits));
763
764         oa->o_valid |= bits;
765         client_obd_list_lock(&cli->cl_loi_list_lock);
766         oa->o_dirty = cli->cl_dirty;
767         if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
768                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
769                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
770                 oa->o_undirty = 0;
771         } else if (atomic_read(&obd_dirty_pages) -
772                    atomic_read(&obd_dirty_transit_pages) > obd_max_dirty_pages + 1){
773                 /* The atomic_read() allowing the atomic_inc() are not covered
774                  * by a lock thus they may safely race and trip this CERROR()
775                  * unless we add in a small fudge factor (+1). */
776                 CERROR("dirty %d - %d > system dirty_max %d\n",
777                        atomic_read(&obd_dirty_pages),
778                        atomic_read(&obd_dirty_transit_pages),
779                        obd_max_dirty_pages);
780                 oa->o_undirty = 0;
781         } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
782                 CERROR("dirty %lu - dirty_max %lu too big???\n",
783                        cli->cl_dirty, cli->cl_dirty_max);
784                 oa->o_undirty = 0;
785         } else {
786                 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
787                                 (cli->cl_max_rpcs_in_flight + 1);
788                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
789         }
790         oa->o_grant = cli->cl_avail_grant;
791         oa->o_dropped = cli->cl_lost_grant;
792         cli->cl_lost_grant = 0;
793         client_obd_list_unlock(&cli->cl_loi_list_lock);
794         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
795                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
796
797 }
798
799 static void osc_update_next_shrink(struct client_obd *cli)
800 {
801         cli->cl_next_shrink_grant =
802                 cfs_time_shift(cli->cl_grant_shrink_interval);
803         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
804                cli->cl_next_shrink_grant);
805 }
806
807 /* caller must hold loi_list_lock */
808 static void osc_consume_write_grant(struct client_obd *cli,
809                                     struct brw_page *pga)
810 {
811         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
812         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
813         atomic_inc(&obd_dirty_pages);
814         cli->cl_dirty += CFS_PAGE_SIZE;
815         cli->cl_avail_grant -= CFS_PAGE_SIZE;
816         pga->flag |= OBD_BRW_FROM_GRANT;
817         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
818                CFS_PAGE_SIZE, pga, pga->pg);
819         LASSERT(cli->cl_avail_grant >= 0);
820         osc_update_next_shrink(cli);
821 }
822
823 /* the companion to osc_consume_write_grant, called when a brw has completed.
824  * must be called with the loi lock held. */
825 static void osc_release_write_grant(struct client_obd *cli,
826                                     struct brw_page *pga, int sent)
827 {
828         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
829         ENTRY;
830
831         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
832         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
833                 EXIT;
834                 return;
835         }
836
837         pga->flag &= ~OBD_BRW_FROM_GRANT;
838         atomic_dec(&obd_dirty_pages);
839         cli->cl_dirty -= CFS_PAGE_SIZE;
840         if (pga->flag & OBD_BRW_NOCACHE) {
841                 pga->flag &= ~OBD_BRW_NOCACHE;
842                 atomic_dec(&obd_dirty_transit_pages);
843                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
844         }
845         if (!sent) {
846                 cli->cl_lost_grant += CFS_PAGE_SIZE;
847                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
848                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
849         } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
850                 /* For short writes we shouldn't count parts of pages that
851                  * span a whole block on the OST side, or our accounting goes
852                  * wrong.  Should match the code in filter_grant_check. */
853                 int offset = pga->off & ~CFS_PAGE_MASK;
854                 int count = pga->count + (offset & (blocksize - 1));
855                 int end = (offset + pga->count) & (blocksize - 1);
856                 if (end)
857                         count += blocksize - end;
858
859                 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
860                 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
861                        CFS_PAGE_SIZE - count, cli->cl_lost_grant,
862                        cli->cl_avail_grant, cli->cl_dirty);
863         }
864
865         EXIT;
866 }
867
868 static unsigned long rpcs_in_flight(struct client_obd *cli)
869 {
870         return cli->cl_r_in_flight + cli->cl_w_in_flight;
871 }
872
873 /* caller must hold loi_list_lock */
874 void osc_wake_cache_waiters(struct client_obd *cli)
875 {
876         struct list_head *l, *tmp;
877         struct osc_cache_waiter *ocw;
878
879         ENTRY;
880         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
881                 /* if we can't dirty more, we must wait until some is written */
882                 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
883                    (atomic_read(&obd_dirty_pages) + 1 > obd_max_dirty_pages)) {
884                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
885                                "osc max %ld, sys max %d\n", cli->cl_dirty,
886                                cli->cl_dirty_max, obd_max_dirty_pages);
887                         return;
888                 }
889
890                 /* if still dirty cache but no grant wait for pending RPCs that
891                  * may yet return us some grant before doing sync writes */
892                 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
893                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
894                                cli->cl_w_in_flight);
895                         return;
896                 }
897
898                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
899                 list_del_init(&ocw->ocw_entry);
900                 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
901                         /* no more RPCs in flight to return grant, do sync IO */
902                         ocw->ocw_rc = -EDQUOT;
903                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
904                 } else {
905                         osc_consume_write_grant(cli,
906                                                 &ocw->ocw_oap->oap_brw_page);
907                 }
908
909                 cfs_waitq_signal(&ocw->ocw_waitq);
910         }
911
912         EXIT;
913 }
914
915 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
916 {
917         client_obd_list_lock(&cli->cl_loi_list_lock);
918         cli->cl_avail_grant += grant;
919         client_obd_list_unlock(&cli->cl_loi_list_lock);
920 }
921
922 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
923 {
924         if (body->oa.o_valid & OBD_MD_FLGRANT) {
925                 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
926                 __osc_update_grant(cli, body->oa.o_grant);
927         }
928 }
929
930 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
931                               void *key, obd_count vallen, void *val,
932                               struct ptlrpc_request_set *set);
933
934 static int osc_shrink_grant_interpret(const struct lu_env *env,
935                                       struct ptlrpc_request *req,
936                                       void *aa, int rc)
937 {
938         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
939         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
940         struct ost_body *body;
941
942         if (rc != 0) {
943                 __osc_update_grant(cli, oa->o_grant);
944                 GOTO(out, rc);
945         }
946
947         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
948         LASSERT(body);
949         osc_update_grant(cli, body);
950 out:
951         OBD_FREE_PTR(oa);
952         return rc;
953 }
954
955 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
956 {
957         client_obd_list_lock(&cli->cl_loi_list_lock);
958         oa->o_grant = cli->cl_avail_grant / 4;
959         cli->cl_avail_grant -= oa->o_grant;
960         client_obd_list_unlock(&cli->cl_loi_list_lock);
961         oa->o_flags |= OBD_FL_SHRINK_GRANT;
962         osc_update_next_shrink(cli);
963 }
964
965 /* Shrink the current grant, either from some large amount to enough for a
966  * full set of in-flight RPCs, or if we have already shrunk to that limit
967  * then to enough for a single RPC.  This avoids keeping more grant than
968  * needed, and avoids shrinking the grant piecemeal. */
969 static int osc_shrink_grant(struct client_obd *cli)
970 {
971         long target = (cli->cl_max_rpcs_in_flight + 1) *
972                       cli->cl_max_pages_per_rpc;
973
974         client_obd_list_lock(&cli->cl_loi_list_lock);
975         if (cli->cl_avail_grant <= target)
976                 target = cli->cl_max_pages_per_rpc;
977         client_obd_list_unlock(&cli->cl_loi_list_lock);
978
979         return osc_shrink_grant_to_target(cli, target);
980 }
981
982 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
983 {
984         int    rc = 0;
985         struct ost_body     *body;
986         ENTRY;
987
988         client_obd_list_lock(&cli->cl_loi_list_lock);
989         /* Don't shrink if we are already above or below the desired limit
990          * We don't want to shrink below a single RPC, as that will negatively
991          * impact block allocation and long-term performance. */
992         if (target < cli->cl_max_pages_per_rpc)
993                 target = cli->cl_max_pages_per_rpc;
994
995         if (target >= cli->cl_avail_grant) {
996                 client_obd_list_unlock(&cli->cl_loi_list_lock);
997                 RETURN(0);
998         }
999         client_obd_list_unlock(&cli->cl_loi_list_lock);
1000
1001         OBD_ALLOC_PTR(body);
1002         if (!body)
1003                 RETURN(-ENOMEM);
1004
1005         osc_announce_cached(cli, &body->oa, 0);
1006
1007         client_obd_list_lock(&cli->cl_loi_list_lock);
1008         body->oa.o_grant = cli->cl_avail_grant - target;
1009         cli->cl_avail_grant = target;
1010         client_obd_list_unlock(&cli->cl_loi_list_lock);
1011         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1012         osc_update_next_shrink(cli);
1013
1014         rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1015                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1016                                 sizeof(*body), body, NULL);
1017         if (rc != 0)
1018                 __osc_update_grant(cli, body->oa.o_grant);
1019         OBD_FREE_PTR(body);
1020         RETURN(rc);
1021 }
1022
1023 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1024 static int osc_should_shrink_grant(struct client_obd *client)
1025 {
1026         cfs_time_t time = cfs_time_current();
1027         cfs_time_t next_shrink = client->cl_next_shrink_grant;
1028         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1029                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1030                     client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1031                         return 1;
1032                 else
1033                         osc_update_next_shrink(client);
1034         }
1035         return 0;
1036 }
1037
1038 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1039 {
1040         struct client_obd *client;
1041
1042         list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
1043                 if (osc_should_shrink_grant(client))
1044                         osc_shrink_grant(client);
1045         }
1046         return 0;
1047 }
1048
1049 static int osc_add_shrink_grant(struct client_obd *client)
1050 {
1051         int rc;
1052
1053         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1054                                        TIMEOUT_GRANT,
1055                                        osc_grant_shrink_grant_cb, NULL,
1056                                        &client->cl_grant_shrink_list);
1057         if (rc) {
1058                 CERROR("add grant client %s error %d\n",
1059                         client->cl_import->imp_obd->obd_name, rc);
1060                 return rc;
1061         }
1062         CDEBUG(D_CACHE, "add grant client %s \n",
1063                client->cl_import->imp_obd->obd_name);
1064         osc_update_next_shrink(client);
1065         return 0;
1066 }
1067
1068 static int osc_del_shrink_grant(struct client_obd *client)
1069 {
1070         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1071                                          TIMEOUT_GRANT);
1072 }
1073
1074 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1075 {
1076         /*
1077          * ocd_grant is the total grant amount we're expect to hold: if we've
1078          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1079          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1080          *
1081          * race is tolerable here: if we're evicted, but imp_state already
1082          * left EVICTED state, then cl_dirty must be 0 already.
1083          */
1084         client_obd_list_lock(&cli->cl_loi_list_lock);
1085         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1086                 cli->cl_avail_grant = ocd->ocd_grant;
1087         else
1088                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1089         client_obd_list_unlock(&cli->cl_loi_list_lock);
1090
1091         CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1092                cli->cl_avail_grant, cli->cl_lost_grant);
1093         LASSERT(cli->cl_avail_grant >= 0);
1094
1095         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1096             list_empty(&cli->cl_grant_shrink_list))
1097                 osc_add_shrink_grant(cli);
1098 }
1099
1100 /* We assume that the reason this OSC got a short read is because it read
1101  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1102  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1103  * this stripe never got written at or beyond this stripe offset yet. */
1104 static void handle_short_read(int nob_read, obd_count page_count,
1105                               struct brw_page **pga)
1106 {
1107         char *ptr;
1108         int i = 0;
1109
1110         /* skip bytes read OK */
1111         while (nob_read > 0) {
1112                 LASSERT (page_count > 0);
1113
1114                 if (pga[i]->count > nob_read) {
1115                         /* EOF inside this page */
1116                         ptr = cfs_kmap(pga[i]->pg) +
1117                                 (pga[i]->off & ~CFS_PAGE_MASK);
1118                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1119                         cfs_kunmap(pga[i]->pg);
1120                         page_count--;
1121                         i++;
1122                         break;
1123                 }
1124
1125                 nob_read -= pga[i]->count;
1126                 page_count--;
1127                 i++;
1128         }
1129
1130         /* zero remaining pages */
1131         while (page_count-- > 0) {
1132                 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1133                 memset(ptr, 0, pga[i]->count);
1134                 cfs_kunmap(pga[i]->pg);
1135                 i++;
1136         }
1137 }
1138
1139 static int check_write_rcs(struct ptlrpc_request *req,
1140                            int requested_nob, int niocount,
1141                            obd_count page_count, struct brw_page **pga)
1142 {
1143         int     i;
1144         __u32   *remote_rcs;
1145
1146         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1147                                                   sizeof(*remote_rcs) *
1148                                                   niocount);
1149         if (remote_rcs == NULL) {
1150                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1151                 return(-EPROTO);
1152         }
1153
1154         /* return error if any niobuf was in error */
1155         for (i = 0; i < niocount; i++) {
1156                 if (remote_rcs[i] < 0)
1157                         return(remote_rcs[i]);
1158
1159                 if (remote_rcs[i] != 0) {
1160                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1161                                 i, remote_rcs[i], req);
1162                         return(-EPROTO);
1163                 }
1164         }
1165
1166         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1167                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1168                        req->rq_bulk->bd_nob_transferred, requested_nob);
1169                 return(-EPROTO);
1170         }
1171
1172         return (0);
1173 }
1174
1175 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1176 {
1177         if (p1->flag != p2->flag) {
1178                 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1179                                   OBD_BRW_NOCACHE|OBD_BRW_SYNC);
1180
1181                 /* warn if we try to combine flags that we don't know to be
1182                  * safe to combine */
1183                 if ((p1->flag & mask) != (p2->flag & mask))
1184                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
1185                                "same brw?\n", p1->flag, p2->flag);
1186                 return 0;
1187         }
1188
1189         return (p1->off + p1->count == p2->off);
1190 }
1191
1192 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1193                                    struct brw_page **pga, int opc,
1194                                    cksum_type_t cksum_type)
1195 {
1196         __u32 cksum;
1197         int i = 0;
1198
1199         LASSERT (pg_count > 0);
1200         cksum = init_checksum(cksum_type);
1201         while (nob > 0 && pg_count > 0) {
1202                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1203                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1204                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1205
1206                 /* corrupt the data before we compute the checksum, to
1207                  * simulate an OST->client data error */
1208                 if (i == 0 && opc == OST_READ &&
1209                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1210                         memcpy(ptr + off, "bad1", min(4, nob));
1211                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1212                 cfs_kunmap(pga[i]->pg);
1213                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1214                                off, cksum);
1215
1216                 nob -= pga[i]->count;
1217                 pg_count--;
1218                 i++;
1219         }
1220         /* For sending we only compute the wrong checksum instead
1221          * of corrupting the data so it is still correct on a redo */
1222         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1223                 cksum++;
1224
1225         return cksum;
1226 }
1227
1228 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1229                                 struct lov_stripe_md *lsm, obd_count page_count,
1230                                 struct brw_page **pga,
1231                                 struct ptlrpc_request **reqp,
1232                                 struct obd_capa *ocapa, int reserve)
1233 {
1234         struct ptlrpc_request   *req;
1235         struct ptlrpc_bulk_desc *desc;
1236         struct ost_body         *body;
1237         struct obd_ioobj        *ioobj;
1238         struct niobuf_remote    *niobuf;
1239         int niocount, i, requested_nob, opc, rc;
1240         struct osc_brw_async_args *aa;
1241         struct req_capsule      *pill;
1242         struct brw_page *pg_prev;
1243
1244         ENTRY;
1245         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1246                 RETURN(-ENOMEM); /* Recoverable */
1247         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1248                 RETURN(-EINVAL); /* Fatal */
1249
1250         if ((cmd & OBD_BRW_WRITE) != 0) {
1251                 opc = OST_WRITE;
1252                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1253                                                 cli->cl_import->imp_rq_pool,
1254                                                 &RQF_OST_BRW);
1255         } else {
1256                 opc = OST_READ;
1257                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1258         }
1259         if (req == NULL)
1260                 RETURN(-ENOMEM);
1261
1262         for (niocount = i = 1; i < page_count; i++) {
1263                 if (!can_merge_pages(pga[i - 1], pga[i]))
1264                         niocount++;
1265         }
1266
1267         pill = &req->rq_pill;
1268         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1269                              sizeof(*ioobj));
1270         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1271                              niocount * sizeof(*niobuf));
1272         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1273
1274         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1275         if (rc) {
1276                 ptlrpc_request_free(req);
1277                 RETURN(rc);
1278         }
1279         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1280         ptlrpc_at_set_req_timeout(req);
1281
1282         if (opc == OST_WRITE)
1283                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1284                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1285         else
1286                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1287                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1288
1289         if (desc == NULL)
1290                 GOTO(out, rc = -ENOMEM);
1291         /* NB request now owns desc and will free it when it gets freed */
1292
1293         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1294         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1295         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1296         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1297
1298         lustre_set_wire_obdo(&body->oa, oa);
1299
1300         obdo_to_ioobj(oa, ioobj);
1301         ioobj->ioo_bufcnt = niocount;
1302         osc_pack_capa(req, body, ocapa);
1303         LASSERT (page_count > 0);
1304         pg_prev = pga[0];
1305         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1306                 struct brw_page *pg = pga[i];
1307
1308                 LASSERT(pg->count > 0);
1309                 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1310                          "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1311                          pg->off, pg->count);
1312 #ifdef __linux__
1313                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1314                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1315                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1316                          i, page_count,
1317                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1318                          pg_prev->pg, page_private(pg_prev->pg),
1319                          pg_prev->pg->index, pg_prev->off);
1320 #else
1321                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1322                          "i %d p_c %u\n", i, page_count);
1323 #endif
1324                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1325                         (pg->flag & OBD_BRW_SRVLOCK));
1326
1327                 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1328                                       pg->count);
1329                 requested_nob += pg->count;
1330
1331                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1332                         niobuf--;
1333                         niobuf->len += pg->count;
1334                 } else {
1335                         niobuf->offset = pg->off;
1336                         niobuf->len    = pg->count;
1337                         niobuf->flags  = pg->flag;
1338                 }
1339                 pg_prev = pg;
1340         }
1341
1342         LASSERTF((void *)(niobuf - niocount) ==
1343                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1344                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1345                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1346
1347         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1348         if (osc_should_shrink_grant(cli))
1349                 osc_shrink_grant_local(cli, &body->oa);
1350
1351         /* size[REQ_REC_OFF] still sizeof (*body) */
1352         if (opc == OST_WRITE) {
1353                 if (unlikely(cli->cl_checksum) &&
1354                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1355                         /* store cl_cksum_type in a local variable since
1356                          * it can be changed via lprocfs */
1357                         cksum_type_t cksum_type = cli->cl_cksum_type;
1358
1359                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1360                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1361                                 body->oa.o_flags = 0;
1362                         }
1363                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1364                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1365                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1366                                                              page_count, pga,
1367                                                              OST_WRITE,
1368                                                              cksum_type);
1369                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1370                                body->oa.o_cksum);
1371                         /* save this in 'oa', too, for later checking */
1372                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1373                         oa->o_flags |= cksum_type_pack(cksum_type);
1374                 } else {
1375                         /* clear out the checksum flag, in case this is a
1376                          * resend but cl_checksum is no longer set. b=11238 */
1377                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1378                 }
1379                 oa->o_cksum = body->oa.o_cksum;
1380                 /* 1 RC per niobuf */
1381                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1382                                      sizeof(__u32) * niocount);
1383         } else {
1384                 if (unlikely(cli->cl_checksum) &&
1385                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1386                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1387                                 body->oa.o_flags = 0;
1388                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1389                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1390                 }
1391                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER, 0);
1392                 /* 1 RC for the whole I/O */
1393         }
1394         ptlrpc_request_set_replen(req);
1395
1396         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1397         aa = ptlrpc_req_async_args(req);
1398         aa->aa_oa = oa;
1399         aa->aa_requested_nob = requested_nob;
1400         aa->aa_nio_count = niocount;
1401         aa->aa_page_count = page_count;
1402         aa->aa_resends = 0;
1403         aa->aa_ppga = pga;
1404         aa->aa_cli = cli;
1405         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1406         if (ocapa && reserve)
1407                 aa->aa_ocapa = capa_get(ocapa);
1408
1409         *reqp = req;
1410         RETURN(0);
1411
1412  out:
1413         ptlrpc_req_finished(req);
1414         RETURN(rc);
1415 }
1416
1417 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1418                                 __u32 client_cksum, __u32 server_cksum, int nob,
1419                                 obd_count page_count, struct brw_page **pga,
1420                                 cksum_type_t client_cksum_type)
1421 {
1422         __u32 new_cksum;
1423         char *msg;
1424         cksum_type_t cksum_type;
1425
1426         if (server_cksum == client_cksum) {
1427                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1428                 return 0;
1429         }
1430
1431         if (oa->o_valid & OBD_MD_FLFLAGS)
1432                 cksum_type = cksum_type_unpack(oa->o_flags);
1433         else
1434                 cksum_type = OBD_CKSUM_CRC32;
1435
1436         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1437                                       cksum_type);
1438
1439         if (cksum_type != client_cksum_type)
1440                 msg = "the server did not use the checksum type specified in "
1441                       "the original request - likely a protocol problem";
1442         else if (new_cksum == server_cksum)
1443                 msg = "changed on the client after we checksummed it - "
1444                       "likely false positive due to mmap IO (bug 11742)";
1445         else if (new_cksum == client_cksum)
1446                 msg = "changed in transit before arrival at OST";
1447         else
1448                 msg = "changed in transit AND doesn't match the original - "
1449                       "likely false positive due to mmap IO (bug 11742)";
1450
1451         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1452                            LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1453                            "["LPU64"-"LPU64"]\n",
1454                            msg, libcfs_nid2str(peer->nid),
1455                            oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1456                            oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1457                                                         (__u64)0,
1458                            oa->o_id,
1459                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1460                            pga[0]->off,
1461                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1462         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1463                "client csum now %x\n", client_cksum, client_cksum_type,
1464                server_cksum, cksum_type, new_cksum);
1465         return 1;
1466 }
1467
1468 /* Note rc enters this function as number of bytes transferred */
1469 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1470 {
1471         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1472         const lnet_process_id_t *peer =
1473                         &req->rq_import->imp_connection->c_peer;
1474         struct client_obd *cli = aa->aa_cli;
1475         struct ost_body *body;
1476         __u32 client_cksum = 0;
1477         ENTRY;
1478
1479         if (rc < 0 && rc != -EDQUOT)
1480                 RETURN(rc);
1481
1482         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1483         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1484         if (body == NULL) {
1485                 CDEBUG(D_INFO, "Can't unpack body\n");
1486                 RETURN(-EPROTO);
1487         }
1488
1489         /* set/clear over quota flag for a uid/gid */
1490         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1491             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1492                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1493
1494                 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1495                              body->oa.o_flags);
1496         }
1497
1498         if (rc < 0)
1499                 RETURN(rc);
1500
1501         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1502                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1503
1504         osc_update_grant(cli, body);
1505
1506         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1507                 if (rc > 0) {
1508                         CERROR("Unexpected +ve rc %d\n", rc);
1509                         RETURN(-EPROTO);
1510                 }
1511                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1512
1513                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1514                         RETURN(-EAGAIN);
1515
1516                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1517                     check_write_checksum(&body->oa, peer, client_cksum,
1518                                          body->oa.o_cksum, aa->aa_requested_nob,
1519                                          aa->aa_page_count, aa->aa_ppga,
1520                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1521                         RETURN(-EAGAIN);
1522
1523                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1524                                      aa->aa_page_count, aa->aa_ppga);
1525                 GOTO(out, rc);
1526         }
1527
1528         /* The rest of this function executes only for OST_READs */
1529
1530         /* if unwrap_bulk failed, return -EAGAIN to retry */
1531         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1532         if (rc < 0)
1533                 GOTO(out, rc = -EAGAIN);
1534
1535         if (rc > aa->aa_requested_nob) {
1536                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1537                        aa->aa_requested_nob);
1538                 RETURN(-EPROTO);
1539         }
1540
1541         if (rc != req->rq_bulk->bd_nob_transferred) {
1542                 CERROR ("Unexpected rc %d (%d transferred)\n",
1543                         rc, req->rq_bulk->bd_nob_transferred);
1544                 return (-EPROTO);
1545         }
1546
1547         if (rc < aa->aa_requested_nob)
1548                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1549
1550         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1551                 static int cksum_counter;
1552                 __u32      server_cksum = body->oa.o_cksum;
1553                 char      *via;
1554                 char      *router;
1555                 cksum_type_t cksum_type;
1556
1557                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1558                         cksum_type = cksum_type_unpack(body->oa.o_flags);
1559                 else
1560                         cksum_type = OBD_CKSUM_CRC32;
1561                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1562                                                  aa->aa_ppga, OST_READ,
1563                                                  cksum_type);
1564
1565                 if (peer->nid == req->rq_bulk->bd_sender) {
1566                         via = router = "";
1567                 } else {
1568                         via = " via ";
1569                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1570                 }
1571
1572                 if (server_cksum == ~0 && rc > 0) {
1573                         CERROR("Protocol error: server %s set the 'checksum' "
1574                                "bit, but didn't send a checksum.  Not fatal, "
1575                                "but please notify on http://bugzilla.lustre.org/\n",
1576                                libcfs_nid2str(peer->nid));
1577                 } else if (server_cksum != client_cksum) {
1578                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1579                                            "%s%s%s inum "LPU64"/"LPU64" object "
1580                                            LPU64"/"LPU64" extent "
1581                                            "["LPU64"-"LPU64"]\n",
1582                                            req->rq_import->imp_obd->obd_name,
1583                                            libcfs_nid2str(peer->nid),
1584                                            via, router,
1585                                            body->oa.o_valid & OBD_MD_FLFID ?
1586                                                 body->oa.o_fid : (__u64)0,
1587                                            body->oa.o_valid & OBD_MD_FLFID ?
1588                                                 body->oa.o_generation :(__u64)0,
1589                                            body->oa.o_id,
1590                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1591                                                 body->oa.o_gr : (__u64)0,
1592                                            aa->aa_ppga[0]->off,
1593                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1594                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1595                                                                         1);
1596                         CERROR("client %x, server %x, cksum_type %x\n",
1597                                client_cksum, server_cksum, cksum_type);
1598                         cksum_counter = 0;
1599                         aa->aa_oa->o_cksum = client_cksum;
1600                         rc = -EAGAIN;
1601                 } else {
1602                         cksum_counter++;
1603                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1604                         rc = 0;
1605                 }
1606         } else if (unlikely(client_cksum)) {
1607                 static int cksum_missed;
1608
1609                 cksum_missed++;
1610                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1611                         CERROR("Checksum %u requested from %s but not sent\n",
1612                                cksum_missed, libcfs_nid2str(peer->nid));
1613         } else {
1614                 rc = 0;
1615         }
1616 out:
1617         if (rc >= 0)
1618                 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1619
1620         RETURN(rc);
1621 }
1622
1623 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1624                             struct lov_stripe_md *lsm,
1625                             obd_count page_count, struct brw_page **pga,
1626                             struct obd_capa *ocapa)
1627 {
1628         struct ptlrpc_request *req;
1629         int                    rc;
1630         cfs_waitq_t            waitq;
1631         int                    resends = 0;
1632         struct l_wait_info     lwi;
1633
1634         ENTRY;
1635
1636         cfs_waitq_init(&waitq);
1637
1638 restart_bulk:
1639         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1640                                   page_count, pga, &req, ocapa, 0);
1641         if (rc != 0)
1642                 return (rc);
1643
1644         rc = ptlrpc_queue_wait(req);
1645
1646         if (rc == -ETIMEDOUT && req->rq_resend) {
1647                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1648                 ptlrpc_req_finished(req);
1649                 goto restart_bulk;
1650         }
1651
1652         rc = osc_brw_fini_request(req, rc);
1653
1654         ptlrpc_req_finished(req);
1655         if (osc_recoverable_error(rc)) {
1656                 resends++;
1657                 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1658                         CERROR("too many resend retries, returning error\n");
1659                         RETURN(-EIO);
1660                 }
1661
1662                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1663                 l_wait_event(waitq, 0, &lwi);
1664
1665                 goto restart_bulk;
1666         }
1667
1668         RETURN (rc);
1669 }
1670
1671 int osc_brw_redo_request(struct ptlrpc_request *request,
1672                          struct osc_brw_async_args *aa)
1673 {
1674         struct ptlrpc_request *new_req;
1675         struct ptlrpc_request_set *set = request->rq_set;
1676         struct osc_brw_async_args *new_aa;
1677         struct osc_async_page *oap;
1678         int rc = 0;
1679         ENTRY;
1680
1681         if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1682                 CERROR("too many resend retries, returning error\n");
1683                 RETURN(-EIO);
1684         }
1685
1686         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1687
1688         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1689                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1690                                   aa->aa_cli, aa->aa_oa,
1691                                   NULL /* lsm unused by osc currently */,
1692                                   aa->aa_page_count, aa->aa_ppga,
1693                                   &new_req, aa->aa_ocapa, 0);
1694         if (rc)
1695                 RETURN(rc);
1696
1697         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1698
1699         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1700                 if (oap->oap_request != NULL) {
1701                         LASSERTF(request == oap->oap_request,
1702                                  "request %p != oap_request %p\n",
1703                                  request, oap->oap_request);
1704                         if (oap->oap_interrupted) {
1705                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1706                                 ptlrpc_req_finished(new_req);
1707                                 RETURN(-EINTR);
1708                         }
1709                 }
1710         }
1711         /* New request takes over pga and oaps from old request.
1712          * Note that copying a list_head doesn't work, need to move it... */
1713         aa->aa_resends++;
1714         new_req->rq_interpret_reply = request->rq_interpret_reply;
1715         new_req->rq_async_args = request->rq_async_args;
1716         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1717
1718         new_aa = ptlrpc_req_async_args(new_req);
1719
1720         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1721         list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1722         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1723
1724         list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1725                 if (oap->oap_request) {
1726                         ptlrpc_req_finished(oap->oap_request);
1727                         oap->oap_request = ptlrpc_request_addref(new_req);
1728                 }
1729         }
1730
1731         new_aa->aa_ocapa = aa->aa_ocapa;
1732         aa->aa_ocapa = NULL;
1733
1734         /* use ptlrpc_set_add_req is safe because interpret functions work
1735          * in check_set context. only one way exist with access to request
1736          * from different thread got -EINTR - this way protected with
1737          * cl_loi_list_lock */
1738         ptlrpc_set_add_req(set, new_req);
1739
1740         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1741
1742         DEBUG_REQ(D_INFO, new_req, "new request");
1743         RETURN(0);
1744 }
1745
1746 /*
1747  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1748  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1749  * fine for our small page arrays and doesn't require allocation.  its an
1750  * insertion sort that swaps elements that are strides apart, shrinking the
1751  * stride down until its '1' and the array is sorted.
1752  */
1753 static void sort_brw_pages(struct brw_page **array, int num)
1754 {
1755         int stride, i, j;
1756         struct brw_page *tmp;
1757
1758         if (num == 1)
1759                 return;
1760         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1761                 ;
1762
1763         do {
1764                 stride /= 3;
1765                 for (i = stride ; i < num ; i++) {
1766                         tmp = array[i];
1767                         j = i;
1768                         while (j >= stride && array[j - stride]->off > tmp->off) {
1769                                 array[j] = array[j - stride];
1770                                 j -= stride;
1771                         }
1772                         array[j] = tmp;
1773                 }
1774         } while (stride > 1);
1775 }
1776
1777 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1778 {
1779         int count = 1;
1780         int offset;
1781         int i = 0;
1782
1783         LASSERT (pages > 0);
1784         offset = pg[i]->off & ~CFS_PAGE_MASK;
1785
1786         for (;;) {
1787                 pages--;
1788                 if (pages == 0)         /* that's all */
1789                         return count;
1790
1791                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1792                         return count;   /* doesn't end on page boundary */
1793
1794                 i++;
1795                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1796                 if (offset != 0)        /* doesn't start on page boundary */
1797                         return count;
1798
1799                 count++;
1800         }
1801 }
1802
1803 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1804 {
1805         struct brw_page **ppga;
1806         int i;
1807
1808         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1809         if (ppga == NULL)
1810                 return NULL;
1811
1812         for (i = 0; i < count; i++)
1813                 ppga[i] = pga + i;
1814         return ppga;
1815 }
1816
1817 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1818 {
1819         LASSERT(ppga != NULL);
1820         OBD_FREE(ppga, sizeof(*ppga) * count);
1821 }
1822
1823 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1824                    obd_count page_count, struct brw_page *pga,
1825                    struct obd_trans_info *oti)
1826 {
1827         struct obdo *saved_oa = NULL;
1828         struct brw_page **ppga, **orig;
1829         struct obd_import *imp = class_exp2cliimp(exp);
1830         struct client_obd *cli;
1831         int rc, page_count_orig;
1832         ENTRY;
1833
1834         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1835         cli = &imp->imp_obd->u.cli;
1836
1837         if (cmd & OBD_BRW_CHECK) {
1838                 /* The caller just wants to know if there's a chance that this
1839                  * I/O can succeed */
1840
1841                 if (imp->imp_invalid)
1842                         RETURN(-EIO);
1843                 RETURN(0);
1844         }
1845
1846         /* test_brw with a failed create can trip this, maybe others. */
1847         LASSERT(cli->cl_max_pages_per_rpc);
1848
1849         rc = 0;
1850
1851         orig = ppga = osc_build_ppga(pga, page_count);
1852         if (ppga == NULL)
1853                 RETURN(-ENOMEM);
1854         page_count_orig = page_count;
1855
1856         sort_brw_pages(ppga, page_count);
1857         while (page_count) {
1858                 obd_count pages_per_brw;
1859
1860                 if (page_count > cli->cl_max_pages_per_rpc)
1861                         pages_per_brw = cli->cl_max_pages_per_rpc;
1862                 else
1863                         pages_per_brw = page_count;
1864
1865                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1866
1867                 if (saved_oa != NULL) {
1868                         /* restore previously saved oa */
1869                         *oinfo->oi_oa = *saved_oa;
1870                 } else if (page_count > pages_per_brw) {
1871                         /* save a copy of oa (brw will clobber it) */
1872                         OBDO_ALLOC(saved_oa);
1873                         if (saved_oa == NULL)
1874                                 GOTO(out, rc = -ENOMEM);
1875                         *saved_oa = *oinfo->oi_oa;
1876                 }
1877
1878                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1879                                       pages_per_brw, ppga, oinfo->oi_capa);
1880
1881                 if (rc != 0)
1882                         break;
1883
1884                 page_count -= pages_per_brw;
1885                 ppga += pages_per_brw;
1886         }
1887
1888 out:
1889         osc_release_ppga(orig, page_count_orig);
1890
1891         if (saved_oa != NULL)
1892                 OBDO_FREE(saved_oa);
1893
1894         RETURN(rc);
1895 }
1896
1897 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1898  * the dirty accounting.  Writeback completes or truncate happens before
1899  * writing starts.  Must be called with the loi lock held. */
1900 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1901                            int sent)
1902 {
1903         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1904 }
1905
1906
1907 /* This maintains the lists of pending pages to read/write for a given object
1908  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1909  * to quickly find objects that are ready to send an RPC. */
1910 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1911                          int cmd)
1912 {
1913         int optimal;
1914         ENTRY;
1915
1916         if (lop->lop_num_pending == 0)
1917                 RETURN(0);
1918
1919         /* if we have an invalid import we want to drain the queued pages
1920          * by forcing them through rpcs that immediately fail and complete
1921          * the pages.  recovery relies on this to empty the queued pages
1922          * before canceling the locks and evicting down the llite pages */
1923         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1924                 RETURN(1);
1925
1926         /* stream rpcs in queue order as long as as there is an urgent page
1927          * queued.  this is our cheap solution for good batching in the case
1928          * where writepage marks some random page in the middle of the file
1929          * as urgent because of, say, memory pressure */
1930         if (!list_empty(&lop->lop_urgent)) {
1931                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1932                 RETURN(1);
1933         }
1934         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1935         optimal = cli->cl_max_pages_per_rpc;
1936         if (cmd & OBD_BRW_WRITE) {
1937                 /* trigger a write rpc stream as long as there are dirtiers
1938                  * waiting for space.  as they're waiting, they're not going to
1939                  * create more pages to coallesce with what's waiting.. */
1940                 if (!list_empty(&cli->cl_cache_waiters)) {
1941                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1942                         RETURN(1);
1943                 }
1944                 /* +16 to avoid triggering rpcs that would want to include pages
1945                  * that are being queued but which can't be made ready until
1946                  * the queuer finishes with the page. this is a wart for
1947                  * llite::commit_write() */
1948                 optimal += 16;
1949         }
1950         if (lop->lop_num_pending >= optimal)
1951                 RETURN(1);
1952
1953         RETURN(0);
1954 }
1955
1956 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1957 {
1958         struct osc_async_page *oap;
1959         ENTRY;
1960
1961         if (list_empty(&lop->lop_urgent))
1962                 RETURN(0);
1963
1964         oap = list_entry(lop->lop_urgent.next,
1965                          struct osc_async_page, oap_urgent_item);
1966
1967         if (oap->oap_async_flags & ASYNC_HP) {
1968                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1969                 RETURN(1);
1970         }
1971
1972         RETURN(0);
1973 }
1974
1975 static void on_list(struct list_head *item, struct list_head *list,
1976                     int should_be_on)
1977 {
1978         if (list_empty(item) && should_be_on)
1979                 list_add_tail(item, list);
1980         else if (!list_empty(item) && !should_be_on)
1981                 list_del_init(item);
1982 }
1983
1984 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1985  * can find pages to build into rpcs quickly */
1986 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1987 {
1988         if (lop_makes_hprpc(&loi->loi_write_lop) ||
1989             lop_makes_hprpc(&loi->loi_read_lop)) {
1990                 /* HP rpc */
1991                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1992                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1993         } else {
1994                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1995                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
1996                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
1997                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1998         }
1999
2000         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2001                 loi->loi_write_lop.lop_num_pending);
2002
2003         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2004                 loi->loi_read_lop.lop_num_pending);
2005 }
2006
2007 static void lop_update_pending(struct client_obd *cli,
2008                                struct loi_oap_pages *lop, int cmd, int delta)
2009 {
2010         lop->lop_num_pending += delta;
2011         if (cmd & OBD_BRW_WRITE)
2012                 cli->cl_pending_w_pages += delta;
2013         else
2014                 cli->cl_pending_r_pages += delta;
2015 }
2016
2017 /**
2018  * this is called when a sync waiter receives an interruption.  Its job is to
2019  * get the caller woken as soon as possible.  If its page hasn't been put in an
2020  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2021  * desiring interruption which will forcefully complete the rpc once the rpc
2022  * has timed out.
2023  */
2024 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2025 {
2026         struct loi_oap_pages *lop;
2027         struct lov_oinfo *loi;
2028         int rc = -EBUSY;
2029         ENTRY;
2030
2031         LASSERT(!oap->oap_interrupted);
2032         oap->oap_interrupted = 1;
2033
2034         /* ok, it's been put in an rpc. only one oap gets a request reference */
2035         if (oap->oap_request != NULL) {
2036                 ptlrpc_mark_interrupted(oap->oap_request);
2037                 ptlrpcd_wake(oap->oap_request);
2038                 ptlrpc_req_finished(oap->oap_request);
2039                 oap->oap_request = NULL;
2040         }
2041
2042         /*
2043          * page completion may be called only if ->cpo_prep() method was
2044          * executed by osc_io_submit(), that also adds page the to pending list
2045          */
2046         if (!list_empty(&oap->oap_pending_item)) {
2047                 list_del_init(&oap->oap_pending_item);
2048                 list_del_init(&oap->oap_urgent_item);
2049
2050                 loi = oap->oap_loi;
2051                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2052                         &loi->loi_write_lop : &loi->loi_read_lop;
2053                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2054                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2055                 rc = oap->oap_caller_ops->ap_completion(env,
2056                                           oap->oap_caller_data,
2057                                           oap->oap_cmd, NULL, -EINTR);
2058         }
2059
2060         RETURN(rc);
2061 }
2062
2063 /* this is trying to propogate async writeback errors back up to the
2064  * application.  As an async write fails we record the error code for later if
2065  * the app does an fsync.  As long as errors persist we force future rpcs to be
2066  * sync so that the app can get a sync error and break the cycle of queueing
2067  * pages for which writeback will fail. */
2068 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2069                            int rc)
2070 {
2071         if (rc) {
2072                 if (!ar->ar_rc)
2073                         ar->ar_rc = rc;
2074
2075                 ar->ar_force_sync = 1;
2076                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2077                 return;
2078
2079         }
2080
2081         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2082                 ar->ar_force_sync = 0;
2083 }
2084
2085 void osc_oap_to_pending(struct osc_async_page *oap)
2086 {
2087         struct loi_oap_pages *lop;
2088
2089         if (oap->oap_cmd & OBD_BRW_WRITE)
2090                 lop = &oap->oap_loi->loi_write_lop;
2091         else
2092                 lop = &oap->oap_loi->loi_read_lop;
2093
2094         if (oap->oap_async_flags & ASYNC_HP)
2095                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2096         else if (oap->oap_async_flags & ASYNC_URGENT)
2097                 list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2098         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2099         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2100 }
2101
2102 /* this must be called holding the loi list lock to give coverage to exit_cache,
2103  * async_flag maintenance, and oap_request */
2104 static void osc_ap_completion(const struct lu_env *env,
2105                               struct client_obd *cli, struct obdo *oa,
2106                               struct osc_async_page *oap, int sent, int rc)
2107 {
2108         __u64 xid = 0;
2109
2110         ENTRY;
2111         if (oap->oap_request != NULL) {
2112                 xid = ptlrpc_req_xid(oap->oap_request);
2113                 ptlrpc_req_finished(oap->oap_request);
2114                 oap->oap_request = NULL;
2115         }
2116
2117         spin_lock(&oap->oap_lock);
2118         oap->oap_async_flags = 0;
2119         spin_unlock(&oap->oap_lock);
2120         oap->oap_interrupted = 0;
2121
2122         if (oap->oap_cmd & OBD_BRW_WRITE) {
2123                 osc_process_ar(&cli->cl_ar, xid, rc);
2124                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2125         }
2126
2127         if (rc == 0 && oa != NULL) {
2128                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2129                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2130                 if (oa->o_valid & OBD_MD_FLMTIME)
2131                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2132                 if (oa->o_valid & OBD_MD_FLATIME)
2133                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2134                 if (oa->o_valid & OBD_MD_FLCTIME)
2135                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2136         }
2137
2138         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2139                                                 oap->oap_cmd, oa, rc);
2140
2141         /* ll_ap_completion (from llite) drops PG_locked. so, a new
2142          * I/O on the page could start, but OSC calls it under lock
2143          * and thus we can add oap back to pending safely */
2144         if (rc)
2145                 /* upper layer wants to leave the page on pending queue */
2146                 osc_oap_to_pending(oap);
2147         else
2148                 osc_exit_cache(cli, oap, sent);
2149         EXIT;
2150 }
2151
2152 static int brw_interpret(const struct lu_env *env,
2153                          struct ptlrpc_request *req, void *data, int rc)
2154 {
2155         struct osc_brw_async_args *aa = data;
2156         struct client_obd *cli;
2157         int async;
2158         ENTRY;
2159
2160         rc = osc_brw_fini_request(req, rc);
2161         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2162         if (osc_recoverable_error(rc)) {
2163                 rc = osc_brw_redo_request(req, aa);
2164                 if (rc == 0)
2165                         RETURN(0);
2166         }
2167
2168         if (aa->aa_ocapa) {
2169                 capa_put(aa->aa_ocapa);
2170                 aa->aa_ocapa = NULL;
2171         }
2172
2173         cli = aa->aa_cli;
2174
2175         client_obd_list_lock(&cli->cl_loi_list_lock);
2176
2177         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2178          * is called so we know whether to go to sync BRWs or wait for more
2179          * RPCs to complete */
2180         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2181                 cli->cl_w_in_flight--;
2182         else
2183                 cli->cl_r_in_flight--;
2184
2185         async = list_empty(&aa->aa_oaps);
2186         if (!async) { /* from osc_send_oap_rpc() */
2187                 struct osc_async_page *oap, *tmp;
2188                 /* the caller may re-use the oap after the completion call so
2189                  * we need to clean it up a little */
2190                 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
2191                         list_del_init(&oap->oap_rpc_item);
2192                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2193                 }
2194                 OBDO_FREE(aa->aa_oa);
2195         } else { /* from async_internal() */
2196                 int i;
2197                 for (i = 0; i < aa->aa_page_count; i++)
2198                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2199
2200                 if (aa->aa_oa->o_flags & OBD_FL_TEMPORARY)
2201                         OBDO_FREE(aa->aa_oa);
2202         }
2203         osc_wake_cache_waiters(cli);
2204         osc_check_rpcs(env, cli);
2205         client_obd_list_unlock(&cli->cl_loi_list_lock);
2206         if (!async)
2207                 cl_req_completion(env, aa->aa_clerq, rc);
2208         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2209         RETURN(rc);
2210 }
2211
2212 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2213                                             struct client_obd *cli,
2214                                             struct list_head *rpc_list,
2215                                             int page_count, int cmd)
2216 {
2217         struct ptlrpc_request *req;
2218         struct brw_page **pga = NULL;
2219         struct osc_brw_async_args *aa;
2220         struct obdo *oa = NULL;
2221         const struct obd_async_page_ops *ops = NULL;
2222         void *caller_data = NULL;
2223         struct osc_async_page *oap;
2224         struct osc_async_page *tmp;
2225         struct ost_body *body;
2226         struct cl_req *clerq = NULL;
2227         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2228         struct ldlm_lock *lock = NULL;
2229         struct cl_req_attr crattr;
2230         int i, rc;
2231
2232         ENTRY;
2233         LASSERT(!list_empty(rpc_list));
2234
2235         memset(&crattr, 0, sizeof crattr);
2236         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2237         if (pga == NULL)
2238                 GOTO(out, req = ERR_PTR(-ENOMEM));
2239
2240         OBDO_ALLOC(oa);
2241         if (oa == NULL)
2242                 GOTO(out, req = ERR_PTR(-ENOMEM));
2243
2244         i = 0;
2245         list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2246                 struct cl_page *page = osc_oap2cl_page(oap);
2247                 if (ops == NULL) {
2248                         ops = oap->oap_caller_ops;
2249                         caller_data = oap->oap_caller_data;
2250
2251                         clerq = cl_req_alloc(env, page, crt,
2252                                              1 /* only 1-object rpcs for
2253                                                 * now */);
2254                         if (IS_ERR(clerq))
2255                                 GOTO(out, req = (void *)clerq);
2256                         lock = oap->oap_ldlm_lock;
2257                 }
2258                 pga[i] = &oap->oap_brw_page;
2259                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2260                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2261                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2262                 i++;
2263                 cl_req_page_add(env, clerq, page);
2264         }
2265
2266         /* always get the data for the obdo for the rpc */
2267         LASSERT(ops != NULL);
2268         crattr.cra_oa = oa;
2269         crattr.cra_capa = NULL;
2270         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2271         if (lock) {
2272                 oa->o_handle = lock->l_remote_handle;
2273                 oa->o_valid |= OBD_MD_FLHANDLE;
2274         }
2275
2276         rc = cl_req_prep(env, clerq);
2277         if (rc != 0) {
2278                 CERROR("cl_req_prep failed: %d\n", rc);
2279                 GOTO(out, req = ERR_PTR(rc));
2280         }
2281
2282         sort_brw_pages(pga, page_count);
2283         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2284                                   pga, &req, crattr.cra_capa, 1);
2285         if (rc != 0) {
2286                 CERROR("prep_req failed: %d\n", rc);
2287                 GOTO(out, req = ERR_PTR(rc));
2288         }
2289
2290         /* Need to update the timestamps after the request is built in case
2291          * we race with setattr (locally or in queue at OST).  If OST gets
2292          * later setattr before earlier BRW (as determined by the request xid),
2293          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2294          * way to do this in a single call.  bug 10150 */
2295         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2296         cl_req_attr_set(env, clerq, &crattr,
2297                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2298
2299         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2300         aa = ptlrpc_req_async_args(req);
2301         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2302         list_splice(rpc_list, &aa->aa_oaps);
2303         CFS_INIT_LIST_HEAD(rpc_list);
2304         aa->aa_clerq = clerq;
2305 out:
2306         capa_put(crattr.cra_capa);
2307         if (IS_ERR(req)) {
2308                 if (oa)
2309                         OBDO_FREE(oa);
2310                 if (pga)
2311                         OBD_FREE(pga, sizeof(*pga) * page_count);
2312                 /* this should happen rarely and is pretty bad, it makes the
2313                  * pending list not follow the dirty order */
2314                 client_obd_list_lock(&cli->cl_loi_list_lock);
2315                 list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2316                         list_del_init(&oap->oap_rpc_item);
2317
2318                         /* queued sync pages can be torn down while the pages
2319                          * were between the pending list and the rpc */
2320                         if (oap->oap_interrupted) {
2321                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2322                                 osc_ap_completion(env, cli, NULL, oap, 0,
2323                                                   oap->oap_count);
2324                                 continue;
2325                         }
2326                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2327                 }
2328                 if (clerq && !IS_ERR(clerq))
2329                         cl_req_completion(env, clerq, PTR_ERR(req));
2330         }
2331         RETURN(req);
2332 }
2333
2334 /**
2335  * prepare pages for ASYNC io and put pages in send queue.
2336  *
2337  * \param cmd OBD_BRW_* macroses
2338  * \param lop pending pages
2339  *
2340  * \return zero if pages successfully add to send queue.
2341  * \return not zere if error occurring.
2342  */
2343 static int
2344 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2345                  struct lov_oinfo *loi,
2346                  int cmd, struct loi_oap_pages *lop)
2347 {
2348         struct ptlrpc_request *req;
2349         obd_count page_count = 0;
2350         struct osc_async_page *oap = NULL, *tmp;
2351         struct osc_brw_async_args *aa;
2352         const struct obd_async_page_ops *ops;
2353         CFS_LIST_HEAD(rpc_list);
2354         CFS_LIST_HEAD(tmp_list);
2355         unsigned int ending_offset;
2356         unsigned  starting_offset = 0;
2357         int srvlock = 0;
2358         struct cl_object *clob = NULL;
2359         ENTRY;
2360
2361         /* ASYNC_HP pages first. At present, when the lock the pages is
2362          * to be canceled, the pages covered by the lock will be sent out
2363          * with ASYNC_HP. We have to send out them as soon as possible. */
2364         list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2365                 if (oap->oap_async_flags & ASYNC_HP) 
2366                         list_move(&oap->oap_pending_item, &tmp_list);
2367                 else
2368                         list_move_tail(&oap->oap_pending_item, &tmp_list);
2369                 if (++page_count >= cli->cl_max_pages_per_rpc)
2370                         break;
2371         }
2372
2373         list_splice(&tmp_list, &lop->lop_pending);
2374         page_count = 0;
2375
2376         /* first we find the pages we're allowed to work with */
2377         list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2378                                  oap_pending_item) {
2379                 ops = oap->oap_caller_ops;
2380
2381                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2382                          "magic 0x%x\n", oap, oap->oap_magic);
2383
2384                 if (clob == NULL) {
2385                         /* pin object in memory, so that completion call-backs
2386                          * can be safely called under client_obd_list lock. */
2387                         clob = osc_oap2cl_page(oap)->cp_obj;
2388                         cl_object_get(clob);
2389                 }
2390
2391                 if (page_count != 0 &&
2392                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2393                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2394                                " oap %p, page %p, srvlock %u\n",
2395                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2396                         break;
2397                 }
2398
2399                 /* If there is a gap at the start of this page, it can't merge
2400                  * with any previous page, so we'll hand the network a
2401                  * "fragmented" page array that it can't transfer in 1 RDMA */
2402                 if (page_count != 0 && oap->oap_page_off != 0)
2403                         break;
2404
2405                 /* in llite being 'ready' equates to the page being locked
2406                  * until completion unlocks it.  commit_write submits a page
2407                  * as not ready because its unlock will happen unconditionally
2408                  * as the call returns.  if we race with commit_write giving
2409                  * us that page we dont' want to create a hole in the page
2410                  * stream, so we stop and leave the rpc to be fired by
2411                  * another dirtier or kupdated interval (the not ready page
2412                  * will still be on the dirty list).  we could call in
2413                  * at the end of ll_file_write to process the queue again. */
2414                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2415                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2416                                                     cmd);
2417                         if (rc < 0)
2418                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2419                                                 "instead of ready\n", oap,
2420                                                 oap->oap_page, rc);
2421                         switch (rc) {
2422                         case -EAGAIN:
2423                                 /* llite is telling us that the page is still
2424                                  * in commit_write and that we should try
2425                                  * and put it in an rpc again later.  we
2426                                  * break out of the loop so we don't create
2427                                  * a hole in the sequence of pages in the rpc
2428                                  * stream.*/
2429                                 oap = NULL;
2430                                 break;
2431                         case -EINTR:
2432                                 /* the io isn't needed.. tell the checks
2433                                  * below to complete the rpc with EINTR */
2434                                 spin_lock(&oap->oap_lock);
2435                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2436                                 spin_unlock(&oap->oap_lock);
2437                                 oap->oap_count = -EINTR;
2438                                 break;
2439                         case 0:
2440                                 spin_lock(&oap->oap_lock);
2441                                 oap->oap_async_flags |= ASYNC_READY;
2442                                 spin_unlock(&oap->oap_lock);
2443                                 break;
2444                         default:
2445                                 LASSERTF(0, "oap %p page %p returned %d "
2446                                             "from make_ready\n", oap,
2447                                             oap->oap_page, rc);
2448                                 break;
2449                         }
2450                 }
2451                 if (oap == NULL)
2452                         break;
2453                 /*
2454                  * Page submitted for IO has to be locked. Either by
2455                  * ->ap_make_ready() or by higher layers.
2456                  */
2457 #if defined(__KERNEL__) && defined(__linux__)
2458                 {
2459                         struct cl_page *page;
2460
2461                         page = osc_oap2cl_page(oap);
2462
2463                         if (page->cp_type == CPT_CACHEABLE &&
2464                             !(PageLocked(oap->oap_page) &&
2465                               (CheckWriteback(oap->oap_page, cmd)))) {
2466                                 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2467                                        oap->oap_page,
2468                                        (long)oap->oap_page->flags,
2469                                        oap->oap_async_flags);
2470                                 LBUG();
2471                         }
2472                 }
2473 #endif
2474
2475                 /* take the page out of our book-keeping */
2476                 list_del_init(&oap->oap_pending_item);
2477                 lop_update_pending(cli, lop, cmd, -1);
2478                 list_del_init(&oap->oap_urgent_item);
2479
2480                 if (page_count == 0)
2481                         starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2482                                           (PTLRPC_MAX_BRW_SIZE - 1);
2483
2484                 /* ask the caller for the size of the io as the rpc leaves. */
2485                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2486                         oap->oap_count =
2487                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2488                                                       cmd);
2489                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2490                 }
2491                 if (oap->oap_count <= 0) {
2492                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2493                                oap->oap_count);
2494                         osc_ap_completion(env, cli, NULL,
2495                                           oap, 0, oap->oap_count);
2496                         continue;
2497                 }
2498
2499                 /* now put the page back in our accounting */
2500                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2501                 if (page_count == 0)
2502                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2503                 if (++page_count >= cli->cl_max_pages_per_rpc)
2504                         break;
2505
2506                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2507                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2508                  * have the same alignment as the initial writes that allocated
2509                  * extents on the server. */
2510                 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2511                                  oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2512                 if (ending_offset == 0)
2513                         break;
2514
2515                 /* If there is a gap at the end of this page, it can't merge
2516                  * with any subsequent pages, so we'll hand the network a
2517                  * "fragmented" page array that it can't transfer in 1 RDMA */
2518                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2519                         break;
2520         }
2521
2522         osc_wake_cache_waiters(cli);
2523
2524         loi_list_maint(cli, loi);
2525
2526         client_obd_list_unlock(&cli->cl_loi_list_lock);
2527
2528         if (clob != NULL)
2529                 cl_object_put(env, clob);
2530
2531         if (page_count == 0) {
2532                 client_obd_list_lock(&cli->cl_loi_list_lock);
2533                 RETURN(0);
2534         }
2535
2536         req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2537         if (IS_ERR(req)) {
2538                 LASSERT(list_empty(&rpc_list));
2539                 loi_list_maint(cli, loi);
2540                 RETURN(PTR_ERR(req));
2541         }
2542
2543         aa = ptlrpc_req_async_args(req);
2544
2545         if (cmd == OBD_BRW_READ) {
2546                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2547                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2548                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2549                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2550         } else {
2551                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2552                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2553                                  cli->cl_w_in_flight);
2554                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2555                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2556         }
2557         ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2558
2559         client_obd_list_lock(&cli->cl_loi_list_lock);
2560
2561         if (cmd == OBD_BRW_READ)
2562                 cli->cl_r_in_flight++;
2563         else
2564                 cli->cl_w_in_flight++;
2565
2566         /* queued sync pages can be torn down while the pages
2567          * were between the pending list and the rpc */
2568         tmp = NULL;
2569         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2570                 /* only one oap gets a request reference */
2571                 if (tmp == NULL)
2572                         tmp = oap;
2573                 if (oap->oap_interrupted && !req->rq_intr) {
2574                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2575                                oap, req);
2576                         ptlrpc_mark_interrupted(req);
2577                 }
2578         }
2579         if (tmp != NULL)
2580                 tmp->oap_request = ptlrpc_request_addref(req);
2581
2582         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2583                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2584
2585         req->rq_interpret_reply = brw_interpret;
2586         ptlrpcd_add_req(req, PSCOPE_BRW);
2587         RETURN(1);
2588 }
2589
2590 #define LOI_DEBUG(LOI, STR, args...)                                     \
2591         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2592                !list_empty(&(LOI)->loi_ready_item) ||                    \
2593                !list_empty(&(LOI)->loi_hp_ready_item),                   \
2594                (LOI)->loi_write_lop.lop_num_pending,                     \
2595                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
2596                (LOI)->loi_read_lop.lop_num_pending,                      \
2597                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
2598                args)                                                     \
2599
2600 /* This is called by osc_check_rpcs() to find which objects have pages that
2601  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2602 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2603 {
2604         ENTRY;
2605
2606         /* First return objects that have blocked locks so that they
2607          * will be flushed quickly and other clients can get the lock,
2608          * then objects which have pages ready to be stuffed into RPCs */
2609         if (!list_empty(&cli->cl_loi_hp_ready_list))
2610                 RETURN(list_entry(cli->cl_loi_hp_ready_list.next,
2611                                   struct lov_oinfo, loi_hp_ready_item));
2612         if (!list_empty(&cli->cl_loi_ready_list))
2613                 RETURN(list_entry(cli->cl_loi_ready_list.next,
2614                                   struct lov_oinfo, loi_ready_item));
2615
2616         /* then if we have cache waiters, return all objects with queued
2617          * writes.  This is especially important when many small files
2618          * have filled up the cache and not been fired into rpcs because
2619          * they don't pass the nr_pending/object threshhold */
2620         if (!list_empty(&cli->cl_cache_waiters) &&
2621             !list_empty(&cli->cl_loi_write_list))
2622                 RETURN(list_entry(cli->cl_loi_write_list.next,
2623                                   struct lov_oinfo, loi_write_item));
2624
2625         /* then return all queued objects when we have an invalid import
2626          * so that they get flushed */
2627         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2628                 if (!list_empty(&cli->cl_loi_write_list))
2629                         RETURN(list_entry(cli->cl_loi_write_list.next,
2630                                           struct lov_oinfo, loi_write_item));
2631                 if (!list_empty(&cli->cl_loi_read_list))
2632                         RETURN(list_entry(cli->cl_loi_read_list.next,
2633                                           struct lov_oinfo, loi_read_item));
2634         }
2635         RETURN(NULL);
2636 }
2637
2638 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2639 {
2640         struct osc_async_page *oap;
2641         int hprpc = 0;
2642
2643         if (!list_empty(&loi->loi_write_lop.lop_urgent)) {
2644                 oap = list_entry(loi->loi_write_lop.lop_urgent.next,
2645                                  struct osc_async_page, oap_urgent_item);
2646                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2647         }
2648
2649         if (!hprpc && !list_empty(&loi->loi_read_lop.lop_urgent)) {
2650                 oap = list_entry(loi->loi_read_lop.lop_urgent.next,
2651                                  struct osc_async_page, oap_urgent_item);
2652                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2653         }
2654
2655         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2656 }
2657
2658 /* called with the loi list lock held */
2659 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2660 {
2661         struct lov_oinfo *loi;
2662         int rc = 0, race_counter = 0;
2663         ENTRY;
2664
2665         while ((loi = osc_next_loi(cli)) != NULL) {
2666                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2667
2668                 if (osc_max_rpc_in_flight(cli, loi))
2669                         break;
2670
2671                 /* attempt some read/write balancing by alternating between
2672                  * reads and writes in an object.  The makes_rpc checks here
2673                  * would be redundant if we were getting read/write work items
2674                  * instead of objects.  we don't want send_oap_rpc to drain a
2675                  * partial read pending queue when we're given this object to
2676                  * do io on writes while there are cache waiters */
2677                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2678                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2679                                               &loi->loi_write_lop);
2680                         if (rc < 0) {
2681                                 CERROR("Write request failed with %d\n", rc);
2682
2683                                 /* osc_send_oap_rpc failed, mostly because of
2684                                  * memory pressure.
2685                                  *
2686                                  * It can't break here, because if:
2687                                  *  - a page was submitted by osc_io_submit, so
2688                                  *    page locked;
2689                                  *  - no request in flight
2690                                  *  - no subsequent request
2691                                  * The system will be in live-lock state,
2692                                  * because there is no chance to call
2693                                  * osc_io_unplug() and osc_check_rpcs() any
2694                                  * more. pdflush can't help in this case,
2695                                  * because it might be blocked at grabbing
2696                                  * the page lock as we mentioned.
2697                                  *
2698                                  * Anyway, continue to drain pages. */
2699                                 /* break; */
2700                         }
2701
2702                         if (rc > 0)
2703                                 race_counter = 0;
2704                         else
2705                                 race_counter++;
2706                 }
2707                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2708                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2709                                               &loi->loi_read_lop);
2710                         if (rc < 0)
2711                                 CERROR("Read request failed with %d\n", rc);
2712
2713                         if (rc > 0)
2714                                 race_counter = 0;
2715                         else
2716                                 race_counter++;
2717                 }
2718
2719                 /* attempt some inter-object balancing by issueing rpcs
2720                  * for each object in turn */
2721                 if (!list_empty(&loi->loi_hp_ready_item))
2722                         list_del_init(&loi->loi_hp_ready_item);
2723                 if (!list_empty(&loi->loi_ready_item))
2724                         list_del_init(&loi->loi_ready_item);
2725                 if (!list_empty(&loi->loi_write_item))
2726                         list_del_init(&loi->loi_write_item);
2727                 if (!list_empty(&loi->loi_read_item))
2728                         list_del_init(&loi->loi_read_item);
2729
2730                 loi_list_maint(cli, loi);
2731
2732                 /* send_oap_rpc fails with 0 when make_ready tells it to
2733                  * back off.  llite's make_ready does this when it tries
2734                  * to lock a page queued for write that is already locked.
2735                  * we want to try sending rpcs from many objects, but we
2736                  * don't want to spin failing with 0.  */
2737                 if (race_counter == 10)
2738                         break;
2739         }
2740         EXIT;
2741 }
2742
2743 /* we're trying to queue a page in the osc so we're subject to the
2744  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2745  * If the osc's queued pages are already at that limit, then we want to sleep
2746  * until there is space in the osc's queue for us.  We also may be waiting for
2747  * write credits from the OST if there are RPCs in flight that may return some
2748  * before we fall back to sync writes.
2749  *
2750  * We need this know our allocation was granted in the presence of signals */
2751 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2752 {
2753         int rc;
2754         ENTRY;
2755         client_obd_list_lock(&cli->cl_loi_list_lock);
2756         rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2757         client_obd_list_unlock(&cli->cl_loi_list_lock);
2758         RETURN(rc);
2759 };
2760
2761 /**
2762  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2763  * is available.
2764  */
2765 int osc_enter_cache_try(const struct lu_env *env,
2766                         struct client_obd *cli, struct lov_oinfo *loi,
2767                         struct osc_async_page *oap, int transient)
2768 {
2769         int has_grant;
2770
2771         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2772         if (has_grant) {
2773                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2774                 if (transient) {
2775                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2776                         atomic_inc(&obd_dirty_transit_pages);
2777                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2778                 }
2779         }
2780         return has_grant;
2781 }
2782
2783 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2784  * grant or cache space. */
2785 static int osc_enter_cache(const struct lu_env *env,
2786                            struct client_obd *cli, struct lov_oinfo *loi,
2787                            struct osc_async_page *oap)
2788 {
2789         struct osc_cache_waiter ocw;
2790         struct l_wait_info lwi = { 0 };
2791
2792         ENTRY;
2793
2794         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2795                "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2796                cli->cl_dirty_max, obd_max_dirty_pages,
2797                cli->cl_lost_grant, cli->cl_avail_grant);
2798
2799         /* force the caller to try sync io.  this can jump the list
2800          * of queued writes and create a discontiguous rpc stream */
2801         if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2802             loi->loi_ar.ar_force_sync)
2803                 RETURN(-EDQUOT);
2804
2805         /* Hopefully normal case - cache space and write credits available */
2806         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2807             atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2808             osc_enter_cache_try(env, cli, loi, oap, 0))
2809                 RETURN(0);
2810
2811         /* Make sure that there are write rpcs in flight to wait for.  This
2812          * is a little silly as this object may not have any pending but
2813          * other objects sure might. */
2814         if (cli->cl_w_in_flight) {
2815                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2816                 cfs_waitq_init(&ocw.ocw_waitq);
2817                 ocw.ocw_oap = oap;
2818                 ocw.ocw_rc = 0;
2819
2820                 loi_list_maint(cli, loi);
2821                 osc_check_rpcs(env, cli);
2822                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2823
2824                 CDEBUG(D_CACHE, "sleeping for cache space\n");
2825                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2826
2827                 client_obd_list_lock(&cli->cl_loi_list_lock);
2828                 if (!list_empty(&ocw.ocw_entry)) {
2829                         list_del(&ocw.ocw_entry);
2830                         RETURN(-EINTR);
2831                 }
2832                 RETURN(ocw.ocw_rc);
2833         }
2834
2835         RETURN(-EDQUOT);
2836 }
2837
2838
2839 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2840                         struct lov_oinfo *loi, cfs_page_t *page,
2841                         obd_off offset, const struct obd_async_page_ops *ops,
2842                         void *data, void **res, int nocache,
2843                         struct lustre_handle *lockh)
2844 {
2845         struct osc_async_page *oap;
2846
2847         ENTRY;
2848
2849         if (!page)
2850                 return size_round(sizeof(*oap));
2851
2852         oap = *res;
2853         oap->oap_magic = OAP_MAGIC;
2854         oap->oap_cli = &exp->exp_obd->u.cli;
2855         oap->oap_loi = loi;
2856
2857         oap->oap_caller_ops = ops;
2858         oap->oap_caller_data = data;
2859
2860         oap->oap_page = page;
2861         oap->oap_obj_off = offset;
2862         if (!client_is_remote(exp) &&
2863             cfs_capable(CFS_CAP_SYS_RESOURCE))
2864                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2865
2866         LASSERT(!(offset & ~CFS_PAGE_MASK));
2867
2868         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2869         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2870         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2871         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2872
2873         spin_lock_init(&oap->oap_lock);
2874         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2875         RETURN(0);
2876 }
2877
2878 struct osc_async_page *oap_from_cookie(void *cookie)
2879 {
2880         struct osc_async_page *oap = cookie;
2881         if (oap->oap_magic != OAP_MAGIC)
2882                 return ERR_PTR(-EINVAL);
2883         return oap;
2884 };
2885
2886 int osc_queue_async_io(const struct lu_env *env,
2887                        struct obd_export *exp, struct lov_stripe_md *lsm,
2888                        struct lov_oinfo *loi, void *cookie,
2889                        int cmd, obd_off off, int count,
2890                        obd_flag brw_flags, enum async_flags async_flags)
2891 {
2892         struct client_obd *cli = &exp->exp_obd->u.cli;
2893         struct osc_async_page *oap;
2894         int rc = 0;
2895         ENTRY;
2896
2897         oap = oap_from_cookie(cookie);
2898         if (IS_ERR(oap))
2899                 RETURN(PTR_ERR(oap));
2900
2901         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2902                 RETURN(-EIO);
2903
2904         if (!list_empty(&oap->oap_pending_item) ||
2905             !list_empty(&oap->oap_urgent_item) ||
2906             !list_empty(&oap->oap_rpc_item))
2907                 RETURN(-EBUSY);
2908
2909         /* check if the file's owner/group is over quota */
2910         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2911                 struct cl_object *obj;
2912                 struct cl_attr    attr; /* XXX put attr into thread info */
2913                 unsigned int qid[MAXQUOTAS];
2914
2915                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2916
2917                 cl_object_attr_lock(obj);
2918                 rc = cl_object_attr_get(env, obj, &attr);
2919                 cl_object_attr_unlock(obj);
2920
2921                 qid[USRQUOTA] = attr.cat_uid;
2922                 qid[GRPQUOTA] = attr.cat_gid;
2923                 if (rc == 0 &&
2924                     lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
2925                         rc = -EDQUOT;
2926                 if (rc)
2927                         RETURN(rc);
2928         }
2929
2930         if (loi == NULL)
2931                 loi = lsm->lsm_oinfo[0];
2932
2933         client_obd_list_lock(&cli->cl_loi_list_lock);
2934
2935         LASSERT(off + count <= CFS_PAGE_SIZE);
2936         oap->oap_cmd = cmd;
2937         oap->oap_page_off = off;
2938         oap->oap_count = count;
2939         oap->oap_brw_flags = brw_flags;
2940         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
2941         if (libcfs_memory_pressure_get())
2942                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2943         spin_lock(&oap->oap_lock);
2944         oap->oap_async_flags = async_flags;
2945         spin_unlock(&oap->oap_lock);
2946
2947         if (cmd & OBD_BRW_WRITE) {
2948                 rc = osc_enter_cache(env, cli, loi, oap);
2949                 if (rc) {
2950                         client_obd_list_unlock(&cli->cl_loi_list_lock);
2951                         RETURN(rc);
2952                 }
2953         }
2954
2955         osc_oap_to_pending(oap);
2956         loi_list_maint(cli, loi);
2957
2958         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2959                   cmd);
2960
2961         osc_check_rpcs(env, cli);
2962         client_obd_list_unlock(&cli->cl_loi_list_lock);
2963
2964         RETURN(0);
2965 }
2966
2967 /* aka (~was & now & flag), but this is more clear :) */
2968 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2969
2970 int osc_set_async_flags_base(struct client_obd *cli,
2971                              struct lov_oinfo *loi, struct osc_async_page *oap,
2972                              obd_flag async_flags)
2973 {
2974         struct loi_oap_pages *lop;
2975         int flags = 0;
2976         ENTRY;
2977
2978         LASSERT(!list_empty(&oap->oap_pending_item));
2979
2980         if (oap->oap_cmd & OBD_BRW_WRITE) {
2981                 lop = &loi->loi_write_lop;
2982         } else {
2983                 lop = &loi->loi_read_lop;
2984         }
2985
2986         if ((oap->oap_async_flags & async_flags) == async_flags)
2987                 RETURN(0);
2988
2989         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2990                 flags |= ASYNC_READY;
2991
2992         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
2993             list_empty(&oap->oap_rpc_item)) {
2994                 if (oap->oap_async_flags & ASYNC_HP)
2995                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2996                 else
2997                         list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2998                 flags |= ASYNC_URGENT;
2999                 loi_list_maint(cli, loi);
3000         }
3001         spin_lock(&oap->oap_lock);
3002         oap->oap_async_flags |= flags;
3003         spin_unlock(&oap->oap_lock);
3004
3005         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3006                         oap->oap_async_flags);
3007         RETURN(0);
3008 }
3009
3010 int osc_teardown_async_page(struct obd_export *exp,
3011                             struct lov_stripe_md *lsm,
3012                             struct lov_oinfo *loi, void *cookie)
3013 {
3014         struct client_obd *cli = &exp->exp_obd->u.cli;
3015         struct loi_oap_pages *lop;
3016         struct osc_async_page *oap;
3017         int rc = 0;
3018         ENTRY;
3019
3020         oap = oap_from_cookie(cookie);
3021         if (IS_ERR(oap))
3022                 RETURN(PTR_ERR(oap));
3023
3024         if (loi == NULL)
3025                 loi = lsm->lsm_oinfo[0];
3026
3027         if (oap->oap_cmd & OBD_BRW_WRITE) {
3028                 lop = &loi->loi_write_lop;
3029         } else {
3030                 lop = &loi->loi_read_lop;
3031         }
3032
3033         client_obd_list_lock(&cli->cl_loi_list_lock);
3034
3035         if (!list_empty(&oap->oap_rpc_item))
3036                 GOTO(out, rc = -EBUSY);
3037
3038         osc_exit_cache(cli, oap, 0);
3039         osc_wake_cache_waiters(cli);
3040
3041         if (!list_empty(&oap->oap_urgent_item)) {
3042                 list_del_init(&oap->oap_urgent_item);
3043                 spin_lock(&oap->oap_lock);
3044                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3045                 spin_unlock(&oap->oap_lock);
3046         }
3047         if (!list_empty(&oap->oap_pending_item)) {
3048                 list_del_init(&oap->oap_pending_item);
3049                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3050         }
3051         loi_list_maint(cli, loi);
3052         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3053 out:
3054         client_obd_list_unlock(&cli->cl_loi_list_lock);
3055         RETURN(rc);
3056 }
3057
3058 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
3059                                          struct ldlm_enqueue_info *einfo,
3060                                          int flags)
3061 {
3062         void *data = einfo->ei_cbdata;
3063
3064         LASSERT(lock != NULL);
3065         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3066         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3067         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3068         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3069
3070         lock_res_and_lock(lock);
3071         spin_lock(&osc_ast_guard);
3072         LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
3073         lock->l_ast_data = data;
3074         spin_unlock(&osc_ast_guard);
3075         unlock_res_and_lock(lock);
3076 }
3077
3078 static void osc_set_data_with_check(struct lustre_handle *lockh,
3079                                     struct ldlm_enqueue_info *einfo,
3080                                     int flags)
3081 {
3082         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3083
3084         if (lock != NULL) {
3085                 osc_set_lock_data_with_check(lock, einfo, flags);
3086                 LDLM_LOCK_PUT(lock);
3087         } else
3088                 CERROR("lockh %p, data %p - client evicted?\n",
3089                        lockh, einfo->ei_cbdata);
3090 }
3091
3092 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3093                              ldlm_iterator_t replace, void *data)
3094 {
3095         struct ldlm_res_id res_id;
3096         struct obd_device *obd = class_exp2obd(exp);
3097
3098         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3099         ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3100         return 0;
3101 }
3102
3103 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3104                             obd_enqueue_update_f upcall, void *cookie,
3105                             int *flags, int rc)
3106 {
3107         int intent = *flags & LDLM_FL_HAS_INTENT;
3108         ENTRY;
3109
3110         if (intent) {
3111                 /* The request was created before ldlm_cli_enqueue call. */
3112                 if (rc == ELDLM_LOCK_ABORTED) {
3113                         struct ldlm_reply *rep;
3114                         rep = req_capsule_server_get(&req->rq_pill,
3115                                                      &RMF_DLM_REP);
3116
3117                         LASSERT(rep != NULL);
3118                         if (rep->lock_policy_res1)
3119                                 rc = rep->lock_policy_res1;
3120                 }
3121         }
3122
3123         if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3124                 *flags |= LDLM_FL_LVB_READY;
3125                 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3126                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3127         }
3128
3129         /* Call the update callback. */
3130         rc = (*upcall)(cookie, rc);
3131         RETURN(rc);
3132 }
3133
3134 static int osc_enqueue_interpret(const struct lu_env *env,
3135                                  struct ptlrpc_request *req,
3136                                  struct osc_enqueue_args *aa, int rc)
3137 {
3138         struct ldlm_lock *lock;
3139         struct lustre_handle handle;
3140         __u32 mode;
3141
3142         /* Make a local copy of a lock handle and a mode, because aa->oa_*
3143          * might be freed anytime after lock upcall has been called. */
3144         lustre_handle_copy(&handle, aa->oa_lockh);
3145         mode = aa->oa_ei->ei_mode;
3146
3147         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3148          * be valid. */
3149         lock = ldlm_handle2lock(&handle);
3150
3151         /* Take an additional reference so that a blocking AST that
3152          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3153          * to arrive after an upcall has been executed by
3154          * osc_enqueue_fini(). */
3155         ldlm_lock_addref(&handle, mode);
3156
3157         /* Complete obtaining the lock procedure. */
3158         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3159                                    mode, aa->oa_flags, aa->oa_lvb,
3160                                    sizeof(*aa->oa_lvb), &handle, rc);
3161         /* Complete osc stuff. */
3162         rc = osc_enqueue_fini(req, aa->oa_lvb,
3163                               aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3164
3165         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3166
3167         /* Release the lock for async request. */
3168         if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3169                 /*
3170                  * Releases a reference taken by ldlm_cli_enqueue(), if it is
3171                  * not already released by
3172                  * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3173                  */
3174                 ldlm_lock_decref(&handle, mode);
3175
3176         LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",