Whamcloud - gitweb
Branch: HEAD
[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         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 = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
214                                   lustre_swab_ost_body);
215         if (body) {
216                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
217                 memcpy(aa->aa_oi->oi_oa, &body->oa, sizeof(*aa->aa_oi->oi_oa));
218
219                 /* This should really be sent by the OST */
220                 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
221                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
222         } else {
223                 CDEBUG(D_INFO, "can't unpack ost_body\n");
224                 rc = -EPROTO;
225                 aa->aa_oi->oi_oa->o_valid = 0;
226         }
227 out:
228         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
229         RETURN(rc);
230 }
231
232 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
233                              struct ptlrpc_request_set *set)
234 {
235         struct ptlrpc_request *req;
236         struct osc_async_args *aa;
237         int                    rc;
238         ENTRY;
239
240         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
241         if (req == NULL)
242                 RETURN(-ENOMEM);
243
244         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
245         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
246         if (rc) {
247                 ptlrpc_request_free(req);
248                 RETURN(rc);
249         }
250
251         osc_pack_req_body(req, oinfo);
252
253         ptlrpc_request_set_replen(req);
254         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
255
256         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
257         aa = ptlrpc_req_async_args(req);
258         aa->aa_oi = oinfo;
259
260         ptlrpc_set_add_req(set, req);
261         RETURN(0);
262 }
263
264 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
265 {
266         struct ptlrpc_request *req;
267         struct ost_body       *body;
268         int                    rc;
269         ENTRY;
270
271         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
272         if (req == NULL)
273                 RETURN(-ENOMEM);
274
275         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
276         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
277         if (rc) {
278                 ptlrpc_request_free(req);
279                 RETURN(rc);
280         }
281
282         osc_pack_req_body(req, oinfo);
283
284         ptlrpc_request_set_replen(req);
285
286         rc = ptlrpc_queue_wait(req);
287         if (rc)
288                 GOTO(out, rc);
289
290         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
291         if (body == NULL)
292                 GOTO(out, rc = -EPROTO);
293
294         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
295         *oinfo->oi_oa = body->oa;
296
297         /* This should really be sent by the OST */
298         oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
299         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
300
301         EXIT;
302  out:
303         ptlrpc_req_finished(req);
304         return rc;
305 }
306
307 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
308                        struct obd_trans_info *oti)
309 {
310         struct ptlrpc_request *req;
311         struct ost_body       *body;
312         int                    rc;
313         ENTRY;
314
315         LASSERTF(!(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP) ||
316                  CHECK_MDS_GROUP(oinfo->oi_oa->o_gr),
317                  "oinfo->oi_oa->o_valid="LPU64" oinfo->oi_oa->o_gr="LPU64"\n",
318                  oinfo->oi_oa->o_valid, oinfo->oi_oa->o_gr);
319
320         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
321         if (req == NULL)
322                 RETURN(-ENOMEM);
323
324         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
325         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
326         if (rc) {
327                 ptlrpc_request_free(req);
328                 RETURN(rc);
329         }
330
331         osc_pack_req_body(req, oinfo);
332
333         ptlrpc_request_set_replen(req);
334
335         rc = ptlrpc_queue_wait(req);
336         if (rc)
337                 GOTO(out, rc);
338
339         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
340         if (body == NULL)
341                 GOTO(out, rc = -EPROTO);
342
343         *oinfo->oi_oa = body->oa;
344
345         EXIT;
346 out:
347         ptlrpc_req_finished(req);
348         RETURN(rc);
349 }
350
351 static int osc_setattr_interpret(const struct lu_env *env,
352                                  struct ptlrpc_request *req,
353                                  struct osc_async_args *aa, int rc)
354 {
355         struct ost_body *body;
356         ENTRY;
357
358         if (rc != 0)
359                 GOTO(out, rc);
360
361         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
362         if (body == NULL)
363                 GOTO(out, rc = -EPROTO);
364
365         *aa->aa_oi->oi_oa = body->oa;
366 out:
367         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
368         RETURN(rc);
369 }
370
371 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
372                              struct obd_trans_info *oti,
373                              struct ptlrpc_request_set *rqset)
374 {
375         struct ptlrpc_request *req;
376         struct osc_async_args *aa;
377         int                    rc;
378         ENTRY;
379
380         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
381         if (req == NULL)
382                 RETURN(-ENOMEM);
383
384         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
385         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
386         if (rc) {
387                 ptlrpc_request_free(req);
388                 RETURN(rc);
389         }
390
391         osc_pack_req_body(req, oinfo);
392
393         ptlrpc_request_set_replen(req);
394
395         if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
396                 LASSERT(oti);
397                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
398         }
399
400         /* do mds to ost setattr asynchronously */
401         if (!rqset) {
402                 /* Do not wait for response. */
403                 ptlrpcd_add_req(req, PSCOPE_OTHER);
404         } else {
405                 req->rq_interpret_reply =
406                         (ptlrpc_interpterer_t)osc_setattr_interpret;
407
408                 CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
409                 aa = ptlrpc_req_async_args(req);
410                 aa->aa_oi = oinfo;
411
412                 ptlrpc_set_add_req(rqset, req);
413         }
414
415         RETURN(0);
416 }
417
418 int osc_real_create(struct obd_export *exp, struct obdo *oa,
419                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
420 {
421         struct ptlrpc_request *req;
422         struct ost_body       *body;
423         struct lov_stripe_md  *lsm;
424         int                    rc;
425         ENTRY;
426
427         LASSERT(oa);
428         LASSERT(ea);
429
430         lsm = *ea;
431         if (!lsm) {
432                 rc = obd_alloc_memmd(exp, &lsm);
433                 if (rc < 0)
434                         RETURN(rc);
435         }
436
437         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
438         if (req == NULL)
439                 GOTO(out, rc = -ENOMEM);
440
441         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
442         if (rc) {
443                 ptlrpc_request_free(req);
444                 GOTO(out, rc);
445         }
446
447         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
448         LASSERT(body);
449         body->oa = *oa;
450
451         ptlrpc_request_set_replen(req);
452
453         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
454             oa->o_flags == OBD_FL_DELORPHAN) {
455                 DEBUG_REQ(D_HA, req,
456                           "delorphan from OST integration");
457                 /* Don't resend the delorphan req */
458                 req->rq_no_resend = req->rq_no_delay = 1;
459         }
460
461         rc = ptlrpc_queue_wait(req);
462         if (rc)
463                 GOTO(out_req, rc);
464
465         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
466         if (body == NULL)
467                 GOTO(out_req, rc = -EPROTO);
468
469         *oa = body->oa;
470
471         /* This should really be sent by the OST */
472         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
473         oa->o_valid |= OBD_MD_FLBLKSZ;
474
475         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
476          * have valid lsm_oinfo data structs, so don't go touching that.
477          * This needs to be fixed in a big way.
478          */
479         lsm->lsm_object_id = oa->o_id;
480         lsm->lsm_object_gr = oa->o_gr;
481         *ea = lsm;
482
483         if (oti != NULL) {
484                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
485
486                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
487                         if (!oti->oti_logcookies)
488                                 oti_alloc_cookies(oti, 1);
489                         *oti->oti_logcookies = oa->o_lcookie;
490                 }
491         }
492
493         CDEBUG(D_HA, "transno: "LPD64"\n",
494                lustre_msg_get_transno(req->rq_repmsg));
495 out_req:
496         ptlrpc_req_finished(req);
497 out:
498         if (rc && !*ea)
499                 obd_free_memmd(exp, &lsm);
500         RETURN(rc);
501 }
502
503 static int osc_punch_interpret(const struct lu_env *env,
504                                struct ptlrpc_request *req,
505                                struct osc_punch_args *aa, int rc)
506 {
507         struct ost_body *body;
508         ENTRY;
509
510         if (rc != 0)
511                 GOTO(out, rc);
512
513         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
514         if (body == NULL)
515                 GOTO(out, rc = -EPROTO);
516
517         *aa->pa_oa = body->oa;
518 out:
519         rc = aa->pa_upcall(aa->pa_cookie, rc);
520         RETURN(rc);
521 }
522
523 int osc_punch_base(struct obd_export *exp, struct obdo *oa,
524                    struct obd_capa *capa,
525                    obd_enqueue_update_f upcall, void *cookie,
526                    struct ptlrpc_request_set *rqset)
527 {
528         struct ptlrpc_request *req;
529         struct osc_punch_args *aa;
530         struct ost_body       *body;
531         int                    rc;
532         ENTRY;
533
534         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
535         if (req == NULL)
536                 RETURN(-ENOMEM);
537
538         osc_set_capa_size(req, &RMF_CAPA1, capa);
539         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
540         if (rc) {
541                 ptlrpc_request_free(req);
542                 RETURN(rc);
543         }
544         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
545         ptlrpc_at_set_req_timeout(req);
546
547         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
548         LASSERT(body);
549         body->oa = *oa;
550         osc_pack_capa(req, body, capa);
551
552         ptlrpc_request_set_replen(req);
553
554
555         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_punch_interpret;
556         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
557         aa = ptlrpc_req_async_args(req);
558         aa->pa_oa     = oa;
559         aa->pa_upcall = upcall;
560         aa->pa_cookie = cookie;
561         if (rqset == PTLRPCD_SET)
562                 ptlrpcd_add_req(req, PSCOPE_OTHER);
563         else
564                 ptlrpc_set_add_req(rqset, req);
565
566         RETURN(0);
567 }
568
569 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
570                      struct obd_trans_info *oti,
571                      struct ptlrpc_request_set *rqset)
572 {
573         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
574         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
575         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
576         return osc_punch_base(exp, oinfo->oi_oa, oinfo->oi_capa,
577                               oinfo->oi_cb_up, oinfo, rqset);
578 }
579
580 static int osc_sync(struct obd_export *exp, struct obdo *oa,
581                     struct lov_stripe_md *md, obd_size start, obd_size end,
582                     void *capa)
583 {
584         struct ptlrpc_request *req;
585         struct ost_body       *body;
586         int                    rc;
587         ENTRY;
588
589         if (!oa) {
590                 CDEBUG(D_INFO, "oa NULL\n");
591                 RETURN(-EINVAL);
592         }
593
594         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
595         if (req == NULL)
596                 RETURN(-ENOMEM);
597
598         osc_set_capa_size(req, &RMF_CAPA1, capa);
599         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
600         if (rc) {
601                 ptlrpc_request_free(req);
602                 RETURN(rc);
603         }
604
605         /* overload the size and blocks fields in the oa with start/end */
606         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
607         LASSERT(body);
608         body->oa = *oa;
609         body->oa.o_size = start;
610         body->oa.o_blocks = end;
611         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
612         osc_pack_capa(req, body, capa);
613
614         ptlrpc_request_set_replen(req);
615
616         rc = ptlrpc_queue_wait(req);
617         if (rc)
618                 GOTO(out, rc);
619
620         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
621         if (body == NULL)
622                 GOTO(out, rc = -EPROTO);
623
624         *oa = body->oa;
625
626         EXIT;
627  out:
628         ptlrpc_req_finished(req);
629         return rc;
630 }
631
632 /* Find and cancel locally locks matched by @mode in the resource found by
633  * @objid. Found locks are added into @cancel list. Returns the amount of
634  * locks added to @cancels list. */
635 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
636                                    struct list_head *cancels, ldlm_mode_t mode,
637                                    int lock_flags)
638 {
639         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
640         struct ldlm_res_id res_id;
641         struct ldlm_resource *res;
642         int count;
643         ENTRY;
644
645         osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
646         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
647         if (res == NULL)
648                 RETURN(0);
649
650         LDLM_RESOURCE_ADDREF(res);
651         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
652                                            lock_flags, 0, NULL);
653         LDLM_RESOURCE_DELREF(res);
654         ldlm_resource_putref(res);
655         RETURN(count);
656 }
657
658 static int osc_destroy_interpret(const struct lu_env *env,
659                                  struct ptlrpc_request *req, void *data,
660                                  int rc)
661 {
662         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
663
664         atomic_dec(&cli->cl_destroy_in_flight);
665         cfs_waitq_signal(&cli->cl_destroy_waitq);
666         return 0;
667 }
668
669 static int osc_can_send_destroy(struct client_obd *cli)
670 {
671         if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
672             cli->cl_max_rpcs_in_flight) {
673                 /* The destroy request can be sent */
674                 return 1;
675         }
676         if (atomic_dec_return(&cli->cl_destroy_in_flight) <
677             cli->cl_max_rpcs_in_flight) {
678                 /*
679                  * The counter has been modified between the two atomic
680                  * operations.
681                  */
682                 cfs_waitq_signal(&cli->cl_destroy_waitq);
683         }
684         return 0;
685 }
686
687 /* Destroy requests can be async always on the client, and we don't even really
688  * care about the return code since the client cannot do anything at all about
689  * a destroy failure.
690  * When the MDS is unlinking a filename, it saves the file objects into a
691  * recovery llog, and these object records are cancelled when the OST reports
692  * they were destroyed and sync'd to disk (i.e. transaction committed).
693  * If the client dies, or the OST is down when the object should be destroyed,
694  * the records are not cancelled, and when the OST reconnects to the MDS next,
695  * it will retrieve the llog unlink logs and then sends the log cancellation
696  * cookies to the MDS after committing destroy transactions. */
697 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
698                        struct lov_stripe_md *ea, struct obd_trans_info *oti,
699                        struct obd_export *md_export, void *capa)
700 {
701         struct client_obd     *cli = &exp->exp_obd->u.cli;
702         struct ptlrpc_request *req;
703         struct ost_body       *body;
704         CFS_LIST_HEAD(cancels);
705         int rc, count;
706         ENTRY;
707
708         if (!oa) {
709                 CDEBUG(D_INFO, "oa NULL\n");
710                 RETURN(-EINVAL);
711         }
712
713         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
714                                         LDLM_FL_DISCARD_DATA);
715
716         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
717         if (req == NULL) {
718                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
719                 RETURN(-ENOMEM);
720         }
721
722         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
723         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
724                                0, &cancels, count);
725         if (rc) {
726                 ptlrpc_request_free(req);
727                 RETURN(rc);
728         }
729
730         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
731         ptlrpc_at_set_req_timeout(req);
732
733         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
734                 oa->o_lcookie = *oti->oti_logcookies;
735         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
736         LASSERT(body);
737         body->oa = *oa;
738
739         osc_pack_capa(req, body, (struct obd_capa *)capa);
740         ptlrpc_request_set_replen(req);
741
742         /* don't throttle destroy RPCs for the MDT */
743         if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
744                 req->rq_interpret_reply = osc_destroy_interpret;
745                 if (!osc_can_send_destroy(cli)) {
746                         struct l_wait_info lwi = { 0 };
747
748                         /*
749                          * Wait until the number of on-going destroy RPCs drops
750                          * under max_rpc_in_flight
751                          */
752                         l_wait_event_exclusive(cli->cl_destroy_waitq,
753                                                osc_can_send_destroy(cli), &lwi);
754                 }
755         }
756
757         /* Do not wait for response */
758         ptlrpcd_add_req(req, PSCOPE_OTHER);
759         RETURN(0);
760 }
761
762 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
763                                 long writing_bytes)
764 {
765         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
766
767         LASSERT(!(oa->o_valid & bits));
768
769         oa->o_valid |= bits;
770         client_obd_list_lock(&cli->cl_loi_list_lock);
771         oa->o_dirty = cli->cl_dirty;
772         if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
773                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
774                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
775                 oa->o_undirty = 0;
776         } else if (atomic_read(&obd_dirty_pages) -
777                    atomic_read(&obd_dirty_transit_pages) > obd_max_dirty_pages){
778                 CERROR("dirty %d - %d > system dirty_max %d\n",
779                        atomic_read(&obd_dirty_pages),
780                        atomic_read(&obd_dirty_transit_pages),
781                        obd_max_dirty_pages);
782                 oa->o_undirty = 0;
783         } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
784                 CERROR("dirty %lu - dirty_max %lu too big???\n",
785                        cli->cl_dirty, cli->cl_dirty_max);
786                 oa->o_undirty = 0;
787         } else {
788                 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
789                                 (cli->cl_max_rpcs_in_flight + 1);
790                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
791         }
792         oa->o_grant = cli->cl_avail_grant;
793         oa->o_dropped = cli->cl_lost_grant;
794         cli->cl_lost_grant = 0;
795         client_obd_list_unlock(&cli->cl_loi_list_lock);
796         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
797                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
798
799 }
800
801 static void osc_update_next_shrink(struct client_obd *cli)
802 {
803         int time = GRANT_SHRINK_INTERVAL;
804         cli->cl_next_shrink_grant = cfs_time_shift(time);
805         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
806                cli->cl_next_shrink_grant);
807 }
808
809 /* caller must hold loi_list_lock */
810 static void osc_consume_write_grant(struct client_obd *cli,
811                                     struct brw_page *pga)
812 {
813         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
814         atomic_inc(&obd_dirty_pages);
815         cli->cl_dirty += CFS_PAGE_SIZE;
816         cli->cl_avail_grant -= CFS_PAGE_SIZE;
817         pga->flag |= OBD_BRW_FROM_GRANT;
818         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
819                CFS_PAGE_SIZE, pga, pga->pg);
820         LASSERT(cli->cl_avail_grant >= 0);
821         osc_update_next_shrink(cli);
822 }
823
824 /* the companion to osc_consume_write_grant, called when a brw has completed.
825  * must be called with the loi lock held. */
826 static void osc_release_write_grant(struct client_obd *cli,
827                                     struct brw_page *pga, int sent)
828 {
829         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
830         ENTRY;
831
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, struct ost_body *body)
916 {
917         client_obd_list_lock(&cli->cl_loi_list_lock);
918         CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
919         if (body->oa.o_valid & OBD_MD_FLGRANT)
920                 cli->cl_avail_grant += body->oa.o_grant;
921         /* waiters are woken in brw_interpret */
922         client_obd_list_unlock(&cli->cl_loi_list_lock);
923 }
924
925 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
926                               void *key, obd_count vallen, void *val,
927                               struct ptlrpc_request_set *set);
928
929 static int osc_shrink_grant_interpret(const struct lu_env *env,
930                                       struct ptlrpc_request *req,
931                                       void *aa, int rc)
932 {
933         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
934         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
935         struct ost_body *body;
936         
937         if (rc != 0) {
938                 client_obd_list_lock(&cli->cl_loi_list_lock);
939                 cli->cl_avail_grant += oa->o_grant;
940                 client_obd_list_unlock(&cli->cl_loi_list_lock);
941                 GOTO(out, rc);
942         }
943
944         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
945         LASSERT(body);
946         osc_update_grant(cli, body);
947 out:
948         OBD_FREE_PTR(oa);
949         return rc;        
950 }
951
952 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
953 {
954         client_obd_list_lock(&cli->cl_loi_list_lock);
955         oa->o_grant = cli->cl_avail_grant / 4;
956         cli->cl_avail_grant -= oa->o_grant; 
957         client_obd_list_unlock(&cli->cl_loi_list_lock);
958         oa->o_flags |= OBD_FL_SHRINK_GRANT;
959         osc_update_next_shrink(cli);
960 }
961
962 static int osc_shrink_grant(struct client_obd *cli)
963 {
964         int    rc = 0;
965         struct ost_body     *body;
966         ENTRY;
967
968         OBD_ALLOC_PTR(body);
969         if (!body)
970                 RETURN(-ENOMEM);
971
972         osc_announce_cached(cli, &body->oa, 0);
973         osc_shrink_grant_local(cli, &body->oa);
974         rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
975                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
976                                 sizeof(*body), body, NULL);
977         if (rc) {
978                 client_obd_list_lock(&cli->cl_loi_list_lock);
979                 cli->cl_avail_grant += body->oa.o_grant;
980                 client_obd_list_unlock(&cli->cl_loi_list_lock);
981         }
982         if (body)
983                OBD_FREE_PTR(body);
984         RETURN(rc);
985 }
986
987 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
988 static int osc_should_shrink_grant(struct client_obd *client)
989 {
990         cfs_time_t time = cfs_time_current();
991         cfs_time_t next_shrink = client->cl_next_shrink_grant;
992         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
993                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
994                     client->cl_avail_grant > GRANT_SHRINK_LIMIT)
995                         return 1;
996                 else
997                         osc_update_next_shrink(client);
998         }
999         return 0;
1000 }
1001
1002 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1003 {
1004         struct client_obd *client;
1005
1006         list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
1007                 if (osc_should_shrink_grant(client))
1008                         osc_shrink_grant(client);
1009         }
1010         return 0;
1011 }
1012
1013 static int osc_add_shrink_grant(struct client_obd *client)
1014 {
1015         int rc;
1016
1017         rc = ptlrpc_add_timeout_client(GRANT_SHRINK_INTERVAL, 
1018                                          TIMEOUT_GRANT,
1019                                          osc_grant_shrink_grant_cb, NULL,
1020                                          &client->cl_grant_shrink_list);
1021         if (rc) {
1022                 CERROR("add grant client %s error %d\n", 
1023                         client->cl_import->imp_obd->obd_name, rc);
1024                 return rc;
1025         }
1026         CDEBUG(D_CACHE, "add grant client %s \n", 
1027                client->cl_import->imp_obd->obd_name);
1028         osc_update_next_shrink(client);
1029         return 0; 
1030 }
1031
1032 static int osc_del_shrink_grant(struct client_obd *client)
1033 {
1034         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list, 
1035                                          TIMEOUT_GRANT);
1036 }
1037
1038 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1039 {
1040         client_obd_list_lock(&cli->cl_loi_list_lock);
1041         cli->cl_avail_grant = ocd->ocd_grant;
1042         client_obd_list_unlock(&cli->cl_loi_list_lock);
1043
1044         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1045             list_empty(&cli->cl_grant_shrink_list))
1046                 osc_add_shrink_grant(cli);
1047
1048         CDEBUG(D_CACHE, "setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1049                cli->cl_avail_grant, cli->cl_lost_grant);
1050         LASSERT(cli->cl_avail_grant >= 0);
1051 }
1052
1053 /* We assume that the reason this OSC got a short read is because it read
1054  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1055  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1056  * this stripe never got written at or beyond this stripe offset yet. */
1057 static void handle_short_read(int nob_read, obd_count page_count,
1058                               struct brw_page **pga)
1059 {
1060         char *ptr;
1061         int i = 0;
1062
1063         /* skip bytes read OK */
1064         while (nob_read > 0) {
1065                 LASSERT (page_count > 0);
1066
1067                 if (pga[i]->count > nob_read) {
1068                         /* EOF inside this page */
1069                         ptr = cfs_kmap(pga[i]->pg) +
1070                                 (pga[i]->off & ~CFS_PAGE_MASK);
1071                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1072                         cfs_kunmap(pga[i]->pg);
1073                         page_count--;
1074                         i++;
1075                         break;
1076                 }
1077
1078                 nob_read -= pga[i]->count;
1079                 page_count--;
1080                 i++;
1081         }
1082
1083         /* zero remaining pages */
1084         while (page_count-- > 0) {
1085                 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1086                 memset(ptr, 0, pga[i]->count);
1087                 cfs_kunmap(pga[i]->pg);
1088                 i++;
1089         }
1090 }
1091
1092 static int check_write_rcs(struct ptlrpc_request *req,
1093                            int requested_nob, int niocount,
1094                            obd_count page_count, struct brw_page **pga)
1095 {
1096         int    *remote_rcs, i;
1097
1098         /* return error if any niobuf was in error */
1099         remote_rcs = lustre_swab_repbuf(req, REQ_REC_OFF + 1,
1100                                         sizeof(*remote_rcs) * niocount, NULL);
1101         if (remote_rcs == NULL) {
1102                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1103                 return(-EPROTO);
1104         }
1105         if (lustre_msg_swabbed(req->rq_repmsg))
1106                 for (i = 0; i < niocount; i++)
1107                         __swab32s(&remote_rcs[i]);
1108
1109         for (i = 0; i < niocount; i++) {
1110                 if (remote_rcs[i] < 0)
1111                         return(remote_rcs[i]);
1112
1113                 if (remote_rcs[i] != 0) {
1114                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1115                                 i, remote_rcs[i], req);
1116                         return(-EPROTO);
1117                 }
1118         }
1119
1120         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1121                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1122                        req->rq_bulk->bd_nob_transferred, requested_nob);
1123                 return(-EPROTO);
1124         }
1125
1126         return (0);
1127 }
1128
1129 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1130 {
1131         if (p1->flag != p2->flag) {
1132                 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1133                                   OBD_BRW_NOCACHE|OBD_BRW_SYNC);
1134
1135                 /* warn if we try to combine flags that we don't know to be
1136                  * safe to combine */
1137                 if ((p1->flag & mask) != (p2->flag & mask))
1138                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
1139                                "same brw?\n", p1->flag, p2->flag);
1140                 return 0;
1141         }
1142
1143         return (p1->off + p1->count == p2->off);
1144 }
1145
1146 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1147                                    struct brw_page **pga, int opc,
1148                                    cksum_type_t cksum_type)
1149 {
1150         __u32 cksum;
1151         int i = 0;
1152
1153         LASSERT (pg_count > 0);
1154         cksum = init_checksum(cksum_type);
1155         while (nob > 0 && pg_count > 0) {
1156                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1157                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1158                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1159
1160                 /* corrupt the data before we compute the checksum, to
1161                  * simulate an OST->client data error */
1162                 if (i == 0 && opc == OST_READ &&
1163                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1164                         memcpy(ptr + off, "bad1", min(4, nob));
1165                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1166                 cfs_kunmap(pga[i]->pg);
1167                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1168                                off, cksum);
1169
1170                 nob -= pga[i]->count;
1171                 pg_count--;
1172                 i++;
1173         }
1174         /* For sending we only compute the wrong checksum instead
1175          * of corrupting the data so it is still correct on a redo */
1176         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1177                 cksum++;
1178
1179         return cksum;
1180 }
1181
1182 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1183                                 struct lov_stripe_md *lsm, obd_count page_count,
1184                                 struct brw_page **pga,
1185                                 struct ptlrpc_request **reqp,
1186                                 struct obd_capa *ocapa, int reserve)
1187 {
1188         struct ptlrpc_request   *req;
1189         struct ptlrpc_bulk_desc *desc;
1190         struct ost_body         *body;
1191         struct obd_ioobj        *ioobj;
1192         struct niobuf_remote    *niobuf;
1193         int niocount, i, requested_nob, opc, rc;
1194         struct osc_brw_async_args *aa;
1195         struct req_capsule      *pill;
1196         struct brw_page *pg_prev;
1197
1198         ENTRY;
1199         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1200                 RETURN(-ENOMEM); /* Recoverable */
1201         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1202                 RETURN(-EINVAL); /* Fatal */
1203
1204         if ((cmd & OBD_BRW_WRITE) != 0) {
1205                 opc = OST_WRITE;
1206                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1207                                                 cli->cl_import->imp_rq_pool,
1208                                                 &RQF_OST_BRW);
1209         } else {
1210                 opc = OST_READ;
1211                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW);
1212         }
1213         if (req == NULL)
1214                 RETURN(-ENOMEM);
1215
1216         for (niocount = i = 1; i < page_count; i++) {
1217                 if (!can_merge_pages(pga[i - 1], pga[i]))
1218                         niocount++;
1219         }
1220
1221         pill = &req->rq_pill;
1222         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1223                              niocount * sizeof(*niobuf));
1224         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1225
1226         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1227         if (rc) {
1228                 ptlrpc_request_free(req);
1229                 RETURN(rc);
1230         }
1231         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1232         ptlrpc_at_set_req_timeout(req);
1233
1234         if (opc == OST_WRITE)
1235                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1236                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1237         else
1238                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1239                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1240
1241         if (desc == NULL)
1242                 GOTO(out, rc = -ENOMEM);
1243         /* NB request now owns desc and will free it when it gets freed */
1244
1245         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1246         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1247         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1248         LASSERT(body && ioobj && niobuf);
1249
1250         body->oa = *oa;
1251
1252         obdo_to_ioobj(oa, ioobj);
1253         ioobj->ioo_bufcnt = niocount;
1254         osc_pack_capa(req, body, ocapa);
1255         LASSERT (page_count > 0);
1256         pg_prev = pga[0];
1257         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1258                 struct brw_page *pg = pga[i];
1259
1260                 LASSERT(pg->count > 0);
1261                 LASSERTF((pg->off & ~CFS_PAGE_MASK) + pg->count <= CFS_PAGE_SIZE,
1262                          "i: %d pg: %p off: "LPU64", count: %u\n", i, pg,
1263                          pg->off, pg->count);
1264 #ifdef __linux__
1265                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1266                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1267                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1268                          i, page_count,
1269                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1270                          pg_prev->pg, page_private(pg_prev->pg),
1271                          pg_prev->pg->index, pg_prev->off);
1272 #else
1273                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1274                          "i %d p_c %u\n", i, page_count);
1275 #endif
1276                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1277                         (pg->flag & OBD_BRW_SRVLOCK));
1278
1279                 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~CFS_PAGE_MASK,
1280                                       pg->count);
1281                 requested_nob += pg->count;
1282
1283                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1284                         niobuf--;
1285                         niobuf->len += pg->count;
1286                 } else {
1287                         niobuf->offset = pg->off;
1288                         niobuf->len    = pg->count;
1289                         niobuf->flags  = pg->flag;
1290                 }
1291                 pg_prev = pg;
1292         }
1293
1294         LASSERTF((void *)(niobuf - niocount) ==
1295                 lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2,
1296                                niocount * sizeof(*niobuf)),
1297                 "want %p - real %p\n", lustre_msg_buf(req->rq_reqmsg,
1298                 REQ_REC_OFF + 2, niocount * sizeof(*niobuf)),
1299                 (void *)(niobuf - niocount));
1300
1301         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1302         if (osc_should_shrink_grant(cli))
1303                 osc_shrink_grant_local(cli, &body->oa); 
1304
1305         /* size[REQ_REC_OFF] still sizeof (*body) */
1306         if (opc == OST_WRITE) {
1307                 if (unlikely(cli->cl_checksum) &&
1308                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1309                         /* store cl_cksum_type in a local variable since
1310                          * it can be changed via lprocfs */
1311                         cksum_type_t cksum_type = cli->cl_cksum_type;
1312
1313                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1314                                 oa->o_flags = body->oa.o_flags = 0;
1315                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1316                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1317                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1318                                                              page_count, pga,
1319                                                              OST_WRITE,
1320                                                              cksum_type);
1321                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1322                                body->oa.o_cksum);
1323                         /* save this in 'oa', too, for later checking */
1324                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1325                         oa->o_flags |= cksum_type_pack(cksum_type);
1326                 } else {
1327                         /* clear out the checksum flag, in case this is a
1328                          * resend but cl_checksum is no longer set. b=11238 */
1329                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1330                 }
1331                 oa->o_cksum = body->oa.o_cksum;
1332                 /* 1 RC per niobuf */
1333                 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER,
1334                                      sizeof(__u32) * niocount);
1335         } else {
1336                 if (unlikely(cli->cl_checksum) &&
1337                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1338                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1339                                 body->oa.o_flags = 0;
1340                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1341                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1342                 }
1343                 req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_SERVER, 0);
1344                 /* 1 RC for the whole I/O */
1345         }
1346         ptlrpc_request_set_replen(req);
1347
1348         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1349         aa = ptlrpc_req_async_args(req);
1350         aa->aa_oa = oa;
1351         aa->aa_requested_nob = requested_nob;
1352         aa->aa_nio_count = niocount;
1353         aa->aa_page_count = page_count;
1354         aa->aa_resends = 0;
1355         aa->aa_ppga = pga;
1356         aa->aa_cli = cli;
1357         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1358         if (ocapa && reserve)
1359                 aa->aa_ocapa = capa_get(ocapa);
1360
1361         *reqp = req;
1362         RETURN(0);
1363
1364  out:
1365         ptlrpc_req_finished(req);
1366         RETURN(rc);
1367 }
1368
1369 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1370                                 __u32 client_cksum, __u32 server_cksum, int nob,
1371                                 obd_count page_count, struct brw_page **pga,
1372                                 cksum_type_t client_cksum_type)
1373 {
1374         __u32 new_cksum;
1375         char *msg;
1376         cksum_type_t cksum_type;
1377
1378         if (server_cksum == client_cksum) {
1379                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1380                 return 0;
1381         }
1382
1383         if (oa->o_valid & OBD_MD_FLFLAGS)
1384                 cksum_type = cksum_type_unpack(oa->o_flags);
1385         else
1386                 cksum_type = OBD_CKSUM_CRC32;
1387
1388         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1389                                       cksum_type);
1390
1391         if (cksum_type != client_cksum_type)
1392                 msg = "the server did not use the checksum type specified in "
1393                       "the original request - likely a protocol problem";
1394         else if (new_cksum == server_cksum)
1395                 msg = "changed on the client after we checksummed it - "
1396                       "likely false positive due to mmap IO (bug 11742)";
1397         else if (new_cksum == client_cksum)
1398                 msg = "changed in transit before arrival at OST";
1399         else
1400                 msg = "changed in transit AND doesn't match the original - "
1401                       "likely false positive due to mmap IO (bug 11742)";
1402
1403         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inum "
1404                            LPU64"/"LPU64" object "LPU64"/"LPU64" extent "
1405                            "["LPU64"-"LPU64"]\n",
1406                            msg, libcfs_nid2str(peer->nid),
1407                            oa->o_valid & OBD_MD_FLFID ? oa->o_fid : (__u64)0,
1408                            oa->o_valid & OBD_MD_FLFID ? oa->o_generation :
1409                                                         (__u64)0,
1410                            oa->o_id,
1411                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_gr : (__u64)0,
1412                            pga[0]->off,
1413                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1414         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1415                "client csum now %x\n", client_cksum, client_cksum_type,
1416                server_cksum, cksum_type, new_cksum);
1417         return 1;
1418 }
1419
1420 /* Note rc enters this function as number of bytes transferred */
1421 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1422 {
1423         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1424         const lnet_process_id_t *peer =
1425                         &req->rq_import->imp_connection->c_peer;
1426         struct client_obd *cli = aa->aa_cli;
1427         struct ost_body *body;
1428         __u32 client_cksum = 0;
1429         ENTRY;
1430
1431         if (rc < 0 && rc != -EDQUOT)
1432                 RETURN(rc);
1433
1434         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1435         body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
1436                                   lustre_swab_ost_body);
1437         if (body == NULL) {
1438                 CDEBUG(D_INFO, "Can't unpack body\n");
1439                 RETURN(-EPROTO);
1440         }
1441
1442         /* set/clear over quota flag for a uid/gid */
1443         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1444             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA))
1445                 lquota_setdq(quota_interface, cli, body->oa.o_uid,
1446                              body->oa.o_gid, body->oa.o_valid,
1447                              body->oa.o_flags);
1448
1449         if (rc < 0)
1450                 RETURN(rc);
1451
1452         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1453                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1454
1455         osc_update_grant(cli, body);
1456
1457         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1458                 if (rc > 0) {
1459                         CERROR("Unexpected +ve rc %d\n", rc);
1460                         RETURN(-EPROTO);
1461                 }
1462                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1463
1464                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1465                         RETURN(-EAGAIN);
1466
1467                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1468                     check_write_checksum(&body->oa, peer, client_cksum,
1469                                          body->oa.o_cksum, aa->aa_requested_nob,
1470                                          aa->aa_page_count, aa->aa_ppga,
1471                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1472                         RETURN(-EAGAIN);
1473
1474                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1475                                      aa->aa_page_count, aa->aa_ppga);
1476                 GOTO(out, rc);
1477         }
1478
1479         /* The rest of this function executes only for OST_READs */
1480
1481         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1482         if (rc < 0)
1483                 GOTO(out, rc);
1484
1485         if (rc > aa->aa_requested_nob) {
1486                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1487                        aa->aa_requested_nob);
1488                 RETURN(-EPROTO);
1489         }
1490
1491         if (rc != req->rq_bulk->bd_nob_transferred) {
1492                 CERROR ("Unexpected rc %d (%d transferred)\n",
1493                         rc, req->rq_bulk->bd_nob_transferred);
1494                 return (-EPROTO);
1495         }
1496
1497         if (rc < aa->aa_requested_nob)
1498                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1499
1500         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1501                 static int cksum_counter;
1502                 __u32      server_cksum = body->oa.o_cksum;
1503                 char      *via;
1504                 char      *router;
1505                 cksum_type_t cksum_type;
1506
1507                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1508                         cksum_type = cksum_type_unpack(body->oa.o_flags);
1509                 else
1510                         cksum_type = OBD_CKSUM_CRC32;
1511                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1512                                                  aa->aa_ppga, OST_READ,
1513                                                  cksum_type);
1514
1515                 if (peer->nid == req->rq_bulk->bd_sender) {
1516                         via = router = "";
1517                 } else {
1518                         via = " via ";
1519                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1520                 }
1521
1522                 if (server_cksum == ~0 && rc > 0) {
1523                         CERROR("Protocol error: server %s set the 'checksum' "
1524                                "bit, but didn't send a checksum.  Not fatal, "
1525                                "but please notify on http://bugzilla.lustre.org/\n",
1526                                libcfs_nid2str(peer->nid));
1527                 } else if (server_cksum != client_cksum) {
1528                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1529                                            "%s%s%s inum "LPU64"/"LPU64" object "
1530                                            LPU64"/"LPU64" extent "
1531                                            "["LPU64"-"LPU64"]\n",
1532                                            req->rq_import->imp_obd->obd_name,
1533                                            libcfs_nid2str(peer->nid),
1534                                            via, router,
1535                                            body->oa.o_valid & OBD_MD_FLFID ?
1536                                                 body->oa.o_fid : (__u64)0,
1537                                            body->oa.o_valid & OBD_MD_FLFID ?
1538                                                 body->oa.o_generation :(__u64)0,
1539                                            body->oa.o_id,
1540                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1541                                                 body->oa.o_gr : (__u64)0,
1542                                            aa->aa_ppga[0]->off,
1543                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1544                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1545                                                                         1);
1546                         CERROR("client %x, server %x, cksum_type %x\n",
1547                                client_cksum, server_cksum, cksum_type);
1548                         cksum_counter = 0;
1549                         aa->aa_oa->o_cksum = client_cksum;
1550                         rc = -EAGAIN;
1551                 } else {
1552                         cksum_counter++;
1553                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1554                         rc = 0;
1555                 }
1556         } else if (unlikely(client_cksum)) {
1557                 static int cksum_missed;
1558
1559                 cksum_missed++;
1560                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1561                         CERROR("Checksum %u requested from %s but not sent\n",
1562                                cksum_missed, libcfs_nid2str(peer->nid));
1563         } else {
1564                 rc = 0;
1565         }
1566 out:
1567         if (rc >= 0)
1568                 *aa->aa_oa = body->oa;
1569
1570         RETURN(rc);
1571 }
1572
1573 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1574                             struct lov_stripe_md *lsm,
1575                             obd_count page_count, struct brw_page **pga,
1576                             struct obd_capa *ocapa)
1577 {
1578         struct ptlrpc_request *req;
1579         int                    rc;
1580         cfs_waitq_t            waitq;
1581         int                    resends = 0;
1582         struct l_wait_info     lwi;
1583
1584         ENTRY;
1585
1586         cfs_waitq_init(&waitq);
1587
1588 restart_bulk:
1589         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1590                                   page_count, pga, &req, ocapa, 0);
1591         if (rc != 0)
1592                 return (rc);
1593
1594         rc = ptlrpc_queue_wait(req);
1595
1596         if (rc == -ETIMEDOUT && req->rq_resend) {
1597                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1598                 ptlrpc_req_finished(req);
1599                 goto restart_bulk;
1600         }
1601
1602         rc = osc_brw_fini_request(req, rc);
1603
1604         ptlrpc_req_finished(req);
1605         if (osc_recoverable_error(rc)) {
1606                 resends++;
1607                 if (!osc_should_resend(resends, &exp->exp_obd->u.cli)) {
1608                         CERROR("too many resend retries, returning error\n");
1609                         RETURN(-EIO);
1610                 }
1611
1612                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1613                 l_wait_event(waitq, 0, &lwi);
1614
1615                 goto restart_bulk;
1616         }
1617
1618         RETURN (rc);
1619 }
1620
1621 int osc_brw_redo_request(struct ptlrpc_request *request,
1622                          struct osc_brw_async_args *aa)
1623 {
1624         struct ptlrpc_request *new_req;
1625         struct ptlrpc_request_set *set = request->rq_set;
1626         struct osc_brw_async_args *new_aa;
1627         struct osc_async_page *oap;
1628         int rc = 0;
1629         ENTRY;
1630
1631         if (!osc_should_resend(aa->aa_resends, aa->aa_cli)) {
1632                 CERROR("too many resend retries, returning error\n");
1633                 RETURN(-EIO);
1634         }
1635
1636         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1637
1638         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1639                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1640                                   aa->aa_cli, aa->aa_oa,
1641                                   NULL /* lsm unused by osc currently */,
1642                                   aa->aa_page_count, aa->aa_ppga,
1643                                   &new_req, aa->aa_ocapa, 0);
1644         if (rc)
1645                 RETURN(rc);
1646
1647         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1648
1649         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1650                 if (oap->oap_request != NULL) {
1651                         LASSERTF(request == oap->oap_request,
1652                                  "request %p != oap_request %p\n",
1653                                  request, oap->oap_request);
1654                         if (oap->oap_interrupted) {
1655                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1656                                 ptlrpc_req_finished(new_req);
1657                                 RETURN(-EINTR);
1658                         }
1659                 }
1660         }
1661         /* New request takes over pga and oaps from old request.
1662          * Note that copying a list_head doesn't work, need to move it... */
1663         aa->aa_resends++;
1664         new_req->rq_interpret_reply = request->rq_interpret_reply;
1665         new_req->rq_async_args = request->rq_async_args;
1666         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1667
1668         new_aa = ptlrpc_req_async_args(new_req);
1669
1670         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1671         list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1672         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1673
1674         list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1675                 if (oap->oap_request) {
1676                         ptlrpc_req_finished(oap->oap_request);
1677                         oap->oap_request = ptlrpc_request_addref(new_req);
1678                 }
1679         }
1680
1681         new_aa->aa_ocapa = aa->aa_ocapa;
1682         aa->aa_ocapa = NULL;
1683
1684         /* use ptlrpc_set_add_req is safe because interpret functions work
1685          * in check_set context. only one way exist with access to request
1686          * from different thread got -EINTR - this way protected with
1687          * cl_loi_list_lock */
1688         ptlrpc_set_add_req(set, new_req);
1689
1690         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1691
1692         DEBUG_REQ(D_INFO, new_req, "new request");
1693         RETURN(0);
1694 }
1695
1696 /*
1697  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1698  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1699  * fine for our small page arrays and doesn't require allocation.  its an
1700  * insertion sort that swaps elements that are strides apart, shrinking the
1701  * stride down until its '1' and the array is sorted.
1702  */
1703 static void sort_brw_pages(struct brw_page **array, int num)
1704 {
1705         int stride, i, j;
1706         struct brw_page *tmp;
1707
1708         if (num == 1)
1709                 return;
1710         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1711                 ;
1712
1713         do {
1714                 stride /= 3;
1715                 for (i = stride ; i < num ; i++) {
1716                         tmp = array[i];
1717                         j = i;
1718                         while (j >= stride && array[j - stride]->off > tmp->off) {
1719                                 array[j] = array[j - stride];
1720                                 j -= stride;
1721                         }
1722                         array[j] = tmp;
1723                 }
1724         } while (stride > 1);
1725 }
1726
1727 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1728 {
1729         int count = 1;
1730         int offset;
1731         int i = 0;
1732
1733         LASSERT (pages > 0);
1734         offset = pg[i]->off & ~CFS_PAGE_MASK;
1735
1736         for (;;) {
1737                 pages--;
1738                 if (pages == 0)         /* that's all */
1739                         return count;
1740
1741                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1742                         return count;   /* doesn't end on page boundary */
1743
1744                 i++;
1745                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1746                 if (offset != 0)        /* doesn't start on page boundary */
1747                         return count;
1748
1749                 count++;
1750         }
1751 }
1752
1753 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1754 {
1755         struct brw_page **ppga;
1756         int i;
1757
1758         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1759         if (ppga == NULL)
1760                 return NULL;
1761
1762         for (i = 0; i < count; i++)
1763                 ppga[i] = pga + i;
1764         return ppga;
1765 }
1766
1767 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1768 {
1769         LASSERT(ppga != NULL);
1770         OBD_FREE(ppga, sizeof(*ppga) * count);
1771 }
1772
1773 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1774                    obd_count page_count, struct brw_page *pga,
1775                    struct obd_trans_info *oti)
1776 {
1777         struct obdo *saved_oa = NULL;
1778         struct brw_page **ppga, **orig;
1779         struct obd_import *imp = class_exp2cliimp(exp);
1780         struct client_obd *cli;
1781         int rc, page_count_orig;
1782         ENTRY;
1783
1784         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1785         cli = &imp->imp_obd->u.cli;
1786
1787         if (cmd & OBD_BRW_CHECK) {
1788                 /* The caller just wants to know if there's a chance that this
1789                  * I/O can succeed */
1790
1791                 if (imp->imp_invalid)
1792                         RETURN(-EIO);
1793                 RETURN(0);
1794         }
1795
1796         /* test_brw with a failed create can trip this, maybe others. */
1797         LASSERT(cli->cl_max_pages_per_rpc);
1798
1799         rc = 0;
1800
1801         orig = ppga = osc_build_ppga(pga, page_count);
1802         if (ppga == NULL)
1803                 RETURN(-ENOMEM);
1804         page_count_orig = page_count;
1805
1806         sort_brw_pages(ppga, page_count);
1807         while (page_count) {
1808                 obd_count pages_per_brw;
1809
1810                 if (page_count > cli->cl_max_pages_per_rpc)
1811                         pages_per_brw = cli->cl_max_pages_per_rpc;
1812                 else
1813                         pages_per_brw = page_count;
1814
1815                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1816
1817                 if (saved_oa != NULL) {
1818                         /* restore previously saved oa */
1819                         *oinfo->oi_oa = *saved_oa;
1820                 } else if (page_count > pages_per_brw) {
1821                         /* save a copy of oa (brw will clobber it) */
1822                         OBDO_ALLOC(saved_oa);
1823                         if (saved_oa == NULL)
1824                                 GOTO(out, rc = -ENOMEM);
1825                         *saved_oa = *oinfo->oi_oa;
1826                 }
1827
1828                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1829                                       pages_per_brw, ppga, oinfo->oi_capa);
1830
1831                 if (rc != 0)
1832                         break;
1833
1834                 page_count -= pages_per_brw;
1835                 ppga += pages_per_brw;
1836         }
1837
1838 out:
1839         osc_release_ppga(orig, page_count_orig);
1840
1841         if (saved_oa != NULL)
1842                 OBDO_FREE(saved_oa);
1843
1844         RETURN(rc);
1845 }
1846
1847 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1848  * the dirty accounting.  Writeback completes or truncate happens before
1849  * writing starts.  Must be called with the loi lock held. */
1850 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1851                            int sent)
1852 {
1853         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1854 }
1855
1856
1857 /* This maintains the lists of pending pages to read/write for a given object
1858  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1859  * to quickly find objects that are ready to send an RPC. */
1860 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1861                          int cmd)
1862 {
1863         int optimal;
1864         ENTRY;
1865
1866         if (lop->lop_num_pending == 0)
1867                 RETURN(0);
1868
1869         /* if we have an invalid import we want to drain the queued pages
1870          * by forcing them through rpcs that immediately fail and complete
1871          * the pages.  recovery relies on this to empty the queued pages
1872          * before canceling the locks and evicting down the llite pages */
1873         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1874                 RETURN(1);
1875
1876         /* stream rpcs in queue order as long as as there is an urgent page
1877          * queued.  this is our cheap solution for good batching in the case
1878          * where writepage marks some random page in the middle of the file
1879          * as urgent because of, say, memory pressure */
1880         if (!list_empty(&lop->lop_urgent)) {
1881                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1882                 RETURN(1);
1883         }
1884         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1885         optimal = cli->cl_max_pages_per_rpc;
1886         if (cmd & OBD_BRW_WRITE) {
1887                 /* trigger a write rpc stream as long as there are dirtiers
1888                  * waiting for space.  as they're waiting, they're not going to
1889                  * create more pages to coallesce with what's waiting.. */
1890                 if (!list_empty(&cli->cl_cache_waiters)) {
1891                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
1892                         RETURN(1);
1893                 }
1894                 /* +16 to avoid triggering rpcs that would want to include pages
1895                  * that are being queued but which can't be made ready until
1896                  * the queuer finishes with the page. this is a wart for
1897                  * llite::commit_write() */
1898                 optimal += 16;
1899         }
1900         if (lop->lop_num_pending >= optimal)
1901                 RETURN(1);
1902
1903         RETURN(0);
1904 }
1905
1906 static int lop_makes_hprpc(struct loi_oap_pages *lop)
1907 {
1908         struct osc_async_page *oap;
1909         ENTRY;
1910
1911         if (list_empty(&lop->lop_urgent))
1912                 RETURN(0);
1913
1914         oap = list_entry(lop->lop_urgent.next,
1915                          struct osc_async_page, oap_urgent_item);
1916
1917         if (oap->oap_async_flags & ASYNC_HP) {
1918                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
1919                 RETURN(1);
1920         }
1921
1922         RETURN(0);
1923 }
1924
1925 static void on_list(struct list_head *item, struct list_head *list,
1926                     int should_be_on)
1927 {
1928         if (list_empty(item) && should_be_on)
1929                 list_add_tail(item, list);
1930         else if (!list_empty(item) && !should_be_on)
1931                 list_del_init(item);
1932 }
1933
1934 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1935  * can find pages to build into rpcs quickly */
1936 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1937 {
1938         if (lop_makes_hprpc(&loi->loi_write_lop) ||
1939             lop_makes_hprpc(&loi->loi_read_lop)) {
1940                 /* HP rpc */
1941                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
1942                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
1943         } else {
1944                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
1945                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
1946                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
1947                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1948         }
1949
1950         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1951                 loi->loi_write_lop.lop_num_pending);
1952
1953         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1954                 loi->loi_read_lop.lop_num_pending);
1955 }
1956
1957 static void lop_update_pending(struct client_obd *cli,
1958                                struct loi_oap_pages *lop, int cmd, int delta)
1959 {
1960         lop->lop_num_pending += delta;
1961         if (cmd & OBD_BRW_WRITE)
1962                 cli->cl_pending_w_pages += delta;
1963         else
1964                 cli->cl_pending_r_pages += delta;
1965 }
1966
1967 /**
1968  * this is called when a sync waiter receives an interruption.  Its job is to
1969  * get the caller woken as soon as possible.  If its page hasn't been put in an
1970  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
1971  * desiring interruption which will forcefully complete the rpc once the rpc
1972  * has timed out.
1973  */
1974 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
1975 {
1976         struct loi_oap_pages *lop;
1977         struct lov_oinfo *loi;
1978         int rc = -EBUSY;
1979         ENTRY;
1980
1981         LASSERT(!oap->oap_interrupted);
1982         oap->oap_interrupted = 1;
1983
1984         /* ok, it's been put in an rpc. only one oap gets a request reference */
1985         if (oap->oap_request != NULL) {
1986                 ptlrpc_mark_interrupted(oap->oap_request);
1987                 ptlrpcd_wake(oap->oap_request);
1988                 ptlrpc_req_finished(oap->oap_request);
1989                 oap->oap_request = NULL;
1990         }
1991
1992         /*
1993          * page completion may be called only if ->cpo_prep() method was
1994          * executed by osc_io_submit(), that also adds page the to pending list
1995          */
1996         if (!list_empty(&oap->oap_pending_item)) {
1997                 list_del_init(&oap->oap_pending_item);
1998                 list_del_init(&oap->oap_urgent_item);
1999
2000                 loi = oap->oap_loi;
2001                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2002                         &loi->loi_write_lop : &loi->loi_read_lop;
2003                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2004                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2005                 rc = oap->oap_caller_ops->ap_completion(env,
2006                                           oap->oap_caller_data,
2007                                           oap->oap_cmd, NULL, -EINTR);
2008         }
2009
2010         RETURN(rc);
2011 }
2012
2013 /* this is trying to propogate async writeback errors back up to the
2014  * application.  As an async write fails we record the error code for later if
2015  * the app does an fsync.  As long as errors persist we force future rpcs to be
2016  * sync so that the app can get a sync error and break the cycle of queueing
2017  * pages for which writeback will fail. */
2018 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2019                            int rc)
2020 {
2021         if (rc) {
2022                 if (!ar->ar_rc)
2023                         ar->ar_rc = rc;
2024
2025                 ar->ar_force_sync = 1;
2026                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2027                 return;
2028
2029         }
2030
2031         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2032                 ar->ar_force_sync = 0;
2033 }
2034
2035 void osc_oap_to_pending(struct osc_async_page *oap)
2036 {
2037         struct loi_oap_pages *lop;
2038
2039         if (oap->oap_cmd & OBD_BRW_WRITE)
2040                 lop = &oap->oap_loi->loi_write_lop;
2041         else
2042                 lop = &oap->oap_loi->loi_read_lop;
2043
2044         if (oap->oap_async_flags & ASYNC_HP)
2045                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2046         else if (oap->oap_async_flags & ASYNC_URGENT)
2047                 list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2048         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2049         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2050 }
2051
2052 /* this must be called holding the loi list lock to give coverage to exit_cache,
2053  * async_flag maintenance, and oap_request */
2054 static void osc_ap_completion(const struct lu_env *env,
2055                               struct client_obd *cli, struct obdo *oa,
2056                               struct osc_async_page *oap, int sent, int rc)
2057 {
2058         __u64 xid = 0;
2059
2060         ENTRY;
2061         if (oap->oap_request != NULL) {
2062                 xid = ptlrpc_req_xid(oap->oap_request);
2063                 ptlrpc_req_finished(oap->oap_request);
2064                 oap->oap_request = NULL;
2065         }
2066
2067         oap->oap_async_flags = 0;
2068         oap->oap_interrupted = 0;
2069
2070         if (oap->oap_cmd & OBD_BRW_WRITE) {
2071                 osc_process_ar(&cli->cl_ar, xid, rc);
2072                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2073         }
2074
2075         if (rc == 0 && oa != NULL) {
2076                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2077                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2078                 if (oa->o_valid & OBD_MD_FLMTIME)
2079                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2080                 if (oa->o_valid & OBD_MD_FLATIME)
2081                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2082                 if (oa->o_valid & OBD_MD_FLCTIME)
2083                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2084         }
2085
2086         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2087                                                 oap->oap_cmd, oa, rc);
2088
2089         /* ll_ap_completion (from llite) drops PG_locked. so, a new
2090          * I/O on the page could start, but OSC calls it under lock
2091          * and thus we can add oap back to pending safely */
2092         if (rc)
2093                 /* upper layer wants to leave the page on pending queue */
2094                 osc_oap_to_pending(oap);
2095         else
2096                 osc_exit_cache(cli, oap, sent);
2097         EXIT;
2098 }
2099
2100 static int brw_interpret(const struct lu_env *env,
2101                          struct ptlrpc_request *req, void *data, int rc)
2102 {
2103         struct osc_brw_async_args *aa = data;
2104         struct client_obd *cli;
2105         int async;
2106         ENTRY;
2107
2108         rc = osc_brw_fini_request(req, rc);
2109         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2110         if (osc_recoverable_error(rc)) {
2111                 rc = osc_brw_redo_request(req, aa);
2112                 if (rc == 0)
2113                         RETURN(0);
2114         }
2115
2116         if (aa->aa_ocapa) {
2117                 capa_put(aa->aa_ocapa);
2118                 aa->aa_ocapa = NULL;
2119         }
2120
2121         cli = aa->aa_cli;
2122
2123         client_obd_list_lock(&cli->cl_loi_list_lock);
2124
2125         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2126          * is called so we know whether to go to sync BRWs or wait for more
2127          * RPCs to complete */
2128         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2129                 cli->cl_w_in_flight--;
2130         else
2131                 cli->cl_r_in_flight--;
2132
2133         async = list_empty(&aa->aa_oaps);
2134         if (!async) { /* from osc_send_oap_rpc() */
2135                 struct osc_async_page *oap, *tmp;
2136                 /* the caller may re-use the oap after the completion call so
2137                  * we need to clean it up a little */
2138                 list_for_each_entry_safe(oap, tmp, &aa->aa_oaps, oap_rpc_item) {
2139                         list_del_init(&oap->oap_rpc_item);
2140                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2141                 }
2142                 OBDO_FREE(aa->aa_oa);
2143         } else { /* from async_internal() */
2144                 int i;
2145                 for (i = 0; i < aa->aa_page_count; i++)
2146                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2147         }
2148         osc_wake_cache_waiters(cli);
2149         osc_check_rpcs(env, cli);
2150         client_obd_list_unlock(&cli->cl_loi_list_lock);
2151         if (!async)
2152                 cl_req_completion(env, aa->aa_clerq, rc);
2153         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2154         RETURN(rc);
2155 }
2156
2157 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2158                                             struct client_obd *cli,
2159                                             struct list_head *rpc_list,
2160                                             int page_count, int cmd)
2161 {
2162         struct ptlrpc_request *req;
2163         struct brw_page **pga = NULL;
2164         struct osc_brw_async_args *aa;
2165         struct obdo *oa = NULL;
2166         const struct obd_async_page_ops *ops = NULL;
2167         void *caller_data = NULL;
2168         struct osc_async_page *oap;
2169         struct osc_async_page *tmp;
2170         struct ost_body *body;
2171         struct cl_req *clerq = NULL;
2172         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2173         struct ldlm_lock *lock = NULL;
2174         struct cl_req_attr crattr;
2175         int i, rc;
2176
2177         ENTRY;
2178         LASSERT(!list_empty(rpc_list));
2179
2180         memset(&crattr, 0, sizeof crattr);
2181         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2182         if (pga == NULL)
2183                 GOTO(out, req = ERR_PTR(-ENOMEM));
2184
2185         OBDO_ALLOC(oa);
2186         if (oa == NULL)
2187                 GOTO(out, req = ERR_PTR(-ENOMEM));
2188
2189         i = 0;
2190         list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2191                 struct cl_page *page = osc_oap2cl_page(oap);
2192                 if (ops == NULL) {
2193                         ops = oap->oap_caller_ops;
2194                         caller_data = oap->oap_caller_data;
2195
2196                         clerq = cl_req_alloc(env, page, crt,
2197                                              1 /* only 1-object rpcs for
2198                                                 * now */);
2199                         if (IS_ERR(clerq))
2200                                 GOTO(out, req = (void *)clerq);
2201                         lock = oap->oap_ldlm_lock;
2202                 }
2203                 pga[i] = &oap->oap_brw_page;
2204                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2205                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2206                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2207                 i++;
2208                 cl_req_page_add(env, clerq, page);
2209         }
2210
2211         /* always get the data for the obdo for the rpc */
2212         LASSERT(ops != NULL);
2213         crattr.cra_oa = oa;
2214         crattr.cra_capa = NULL;
2215         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2216         if (lock) {
2217                 oa->o_handle = lock->l_remote_handle;
2218                 oa->o_valid |= OBD_MD_FLHANDLE;
2219         }
2220
2221         rc = cl_req_prep(env, clerq);
2222         if (rc != 0) {
2223                 CERROR("cl_req_prep failed: %d\n", rc);
2224                 GOTO(out, req = ERR_PTR(rc));
2225         }
2226
2227         sort_brw_pages(pga, page_count);
2228         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2229                                   pga, &req, crattr.cra_capa, 1);
2230         if (rc != 0) {
2231                 CERROR("prep_req failed: %d\n", rc);
2232                 GOTO(out, req = ERR_PTR(rc));
2233         }
2234
2235         /* Need to update the timestamps after the request is built in case
2236          * we race with setattr (locally or in queue at OST).  If OST gets
2237          * later setattr before earlier BRW (as determined by the request xid),
2238          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2239          * way to do this in a single call.  bug 10150 */
2240         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2241         cl_req_attr_set(env, clerq, &crattr,
2242                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2243
2244         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2245         aa = ptlrpc_req_async_args(req);
2246         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2247         list_splice(rpc_list, &aa->aa_oaps);
2248         CFS_INIT_LIST_HEAD(rpc_list);
2249         aa->aa_clerq = clerq;
2250 out:
2251         capa_put(crattr.cra_capa);
2252         if (IS_ERR(req)) {
2253                 if (oa)
2254                         OBDO_FREE(oa);
2255                 if (pga)
2256                         OBD_FREE(pga, sizeof(*pga) * page_count);
2257                 /* this should happen rarely and is pretty bad, it makes the
2258                  * pending list not follow the dirty order */
2259                 client_obd_list_lock(&cli->cl_loi_list_lock);
2260                 list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2261                         list_del_init(&oap->oap_rpc_item);
2262
2263                         /* queued sync pages can be torn down while the pages
2264                          * were between the pending list and the rpc */
2265                         if (oap->oap_interrupted) {
2266                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2267                                 osc_ap_completion(env, cli, NULL, oap, 0,
2268                                                   oap->oap_count);
2269                                 continue;
2270                         }
2271                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2272                 }
2273                 if (clerq && !IS_ERR(clerq))
2274                         cl_req_completion(env, clerq, PTR_ERR(req));
2275         }
2276         RETURN(req);
2277 }
2278
2279 /**
2280  * prepare pages for ASYNC io and put pages in send queue.
2281  *
2282  * \param cli -
2283  * \param loi -
2284  * \param cmd - OBD_BRW_* macroses
2285  * \param lop - pending pages
2286  *
2287  * \return zero if pages successfully add to send queue.
2288  * \return not zere if error occurring.
2289  */
2290 static int
2291 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2292                  struct lov_oinfo *loi,
2293                  int cmd, struct loi_oap_pages *lop)
2294 {
2295         struct ptlrpc_request *req;
2296         obd_count page_count = 0;
2297         struct osc_async_page *oap = NULL, *tmp;
2298         struct osc_brw_async_args *aa;
2299         const struct obd_async_page_ops *ops;
2300         CFS_LIST_HEAD(rpc_list);
2301         unsigned int ending_offset;
2302         unsigned  starting_offset = 0;
2303         int srvlock = 0;
2304         struct cl_object *clob = NULL;
2305         ENTRY;
2306
2307         /* If there are HP OAPs we need to handle at least 1 of them,
2308          * move it the beginning of the pending list for that. */
2309         if (!list_empty(&lop->lop_urgent)) {
2310                 oap = list_entry(lop->lop_urgent.next,
2311                                  struct osc_async_page, oap_urgent_item);
2312                 if (oap->oap_async_flags & ASYNC_HP)
2313                         list_move(&oap->oap_pending_item, &lop->lop_pending);
2314         }
2315
2316         /* first we find the pages we're allowed to work with */
2317         list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2318                                  oap_pending_item) {
2319                 ops = oap->oap_caller_ops;
2320
2321                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2322                          "magic 0x%x\n", oap, oap->oap_magic);
2323
2324                 if (clob == NULL) {
2325                         /* pin object in memory, so that completion call-backs
2326                          * can be safely called under client_obd_list lock. */
2327                         clob = osc_oap2cl_page(oap)->cp_obj;
2328                         cl_object_get(clob);
2329                 }
2330
2331                 if (page_count != 0 &&
2332                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2333                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2334                                " oap %p, page %p, srvlock %u\n",
2335                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2336                         break;
2337                 }
2338                 /* in llite being 'ready' equates to the page being locked
2339                  * until completion unlocks it.  commit_write submits a page
2340                  * as not ready because its unlock will happen unconditionally
2341                  * as the call returns.  if we race with commit_write giving
2342                  * us that page we dont' want to create a hole in the page
2343                  * stream, so we stop and leave the rpc to be fired by
2344                  * another dirtier or kupdated interval (the not ready page
2345                  * will still be on the dirty list).  we could call in
2346                  * at the end of ll_file_write to process the queue again. */
2347                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2348                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2349                                                     cmd);
2350                         if (rc < 0)
2351                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2352                                                 "instead of ready\n", oap,
2353                                                 oap->oap_page, rc);
2354                         switch (rc) {
2355                         case -EAGAIN:
2356                                 /* llite is telling us that the page is still
2357                                  * in commit_write and that we should try
2358                                  * and put it in an rpc again later.  we
2359                                  * break out of the loop so we don't create
2360                                  * a hole in the sequence of pages in the rpc
2361                                  * stream.*/
2362                                 oap = NULL;
2363                                 break;
2364                         case -EINTR:
2365                                 /* the io isn't needed.. tell the checks
2366                                  * below to complete the rpc with EINTR */
2367                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2368                                 oap->oap_count = -EINTR;
2369                                 break;
2370                         case 0:
2371                                 oap->oap_async_flags |= ASYNC_READY;
2372                                 break;
2373                         default:
2374                                 LASSERTF(0, "oap %p page %p returned %d "
2375                                             "from make_ready\n", oap,
2376                                             oap->oap_page, rc);
2377                                 break;
2378                         }
2379                 }
2380                 if (oap == NULL)
2381                         break;
2382                 /*
2383                  * Page submitted for IO has to be locked. Either by
2384                  * ->ap_make_ready() or by higher layers.
2385                  */
2386 #if defined(__KERNEL__) && defined(__linux__)
2387                 {
2388                         struct cl_page *page;
2389
2390                         page = osc_oap2cl_page(oap);
2391
2392                         if (page->cp_type == CPT_CACHEABLE &&
2393                             !(PageLocked(oap->oap_page) &&
2394                               (CheckWriteback(oap->oap_page, cmd)))) {
2395                                 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2396                                        oap->oap_page,
2397                                        (long)oap->oap_page->flags,
2398                                        oap->oap_async_flags);
2399                                 LBUG();
2400                         }
2401                 }
2402 #endif
2403                 /* If there is a gap at the start of this page, it can't merge
2404                  * with any previous page, so we'll hand the network a
2405                  * "fragmented" page array that it can't transfer in 1 RDMA */
2406                 if (page_count != 0 && oap->oap_page_off != 0)
2407                         break;
2408
2409                 /* take the page out of our book-keeping */
2410                 list_del_init(&oap->oap_pending_item);
2411                 lop_update_pending(cli, lop, cmd, -1);
2412                 list_del_init(&oap->oap_urgent_item);
2413
2414                 if (page_count == 0)
2415                         starting_offset = (oap->oap_obj_off+oap->oap_page_off) &
2416                                           (PTLRPC_MAX_BRW_SIZE - 1);
2417
2418                 /* ask the caller for the size of the io as the rpc leaves. */
2419                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2420                         oap->oap_count =
2421                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2422                                                       cmd);
2423                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2424                 }
2425                 if (oap->oap_count <= 0) {
2426                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2427                                oap->oap_count);
2428                         osc_ap_completion(env, cli, NULL,
2429                                           oap, 0, oap->oap_count);
2430                         continue;
2431                 }
2432
2433                 /* now put the page back in our accounting */
2434                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
2435                 if (page_count == 0)
2436                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2437                 if (++page_count >= cli->cl_max_pages_per_rpc)
2438                         break;
2439
2440                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2441                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2442                  * have the same alignment as the initial writes that allocated
2443                  * extents on the server. */
2444                 ending_offset = (oap->oap_obj_off + oap->oap_page_off +
2445                                  oap->oap_count) & (PTLRPC_MAX_BRW_SIZE - 1);
2446                 if (ending_offset == 0)
2447                         break;
2448
2449                 /* If there is a gap at the end of this page, it can't merge
2450                  * with any subsequent pages, so we'll hand the network a
2451                  * "fragmented" page array that it can't transfer in 1 RDMA */
2452                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2453                         break;
2454         }
2455
2456         osc_wake_cache_waiters(cli);
2457
2458         loi_list_maint(cli, loi);
2459
2460         client_obd_list_unlock(&cli->cl_loi_list_lock);
2461
2462         if (clob != NULL)
2463                 cl_object_put(env, clob);
2464
2465         if (page_count == 0) {
2466                 client_obd_list_lock(&cli->cl_loi_list_lock);
2467                 RETURN(0);
2468         }
2469
2470         req = osc_build_req(env, cli, &rpc_list, page_count, cmd);
2471         if (IS_ERR(req)) {
2472                 LASSERT(list_empty(&rpc_list));
2473                 loi_list_maint(cli, loi);
2474                 RETURN(PTR_ERR(req));
2475         }
2476
2477         aa = ptlrpc_req_async_args(req);
2478
2479         if (cmd == OBD_BRW_READ) {
2480                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2481                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2482                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2483                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2484         } else {
2485                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2486                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2487                                  cli->cl_w_in_flight);
2488                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2489                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2490         }
2491         ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2492
2493         client_obd_list_lock(&cli->cl_loi_list_lock);
2494
2495         if (cmd == OBD_BRW_READ)
2496                 cli->cl_r_in_flight++;
2497         else
2498                 cli->cl_w_in_flight++;
2499
2500         /* queued sync pages can be torn down while the pages
2501          * were between the pending list and the rpc */
2502         tmp = NULL;
2503         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2504                 /* only one oap gets a request reference */
2505                 if (tmp == NULL)
2506                         tmp = oap;
2507                 if (oap->oap_interrupted && !req->rq_intr) {
2508                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2509                                oap, req);
2510                         ptlrpc_mark_interrupted(req);
2511                 }
2512         }
2513         if (tmp != NULL)
2514                 tmp->oap_request = ptlrpc_request_addref(req);
2515
2516         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2517                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2518
2519         req->rq_interpret_reply = brw_interpret;
2520         ptlrpcd_add_req(req, PSCOPE_BRW);
2521         RETURN(1);
2522 }
2523
2524 #define LOI_DEBUG(LOI, STR, args...)                                     \
2525         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2526                !list_empty(&(LOI)->loi_ready_item) ||                    \
2527                !list_empty(&(LOI)->loi_hp_ready_item),                   \
2528                (LOI)->loi_write_lop.lop_num_pending,                     \
2529                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
2530                (LOI)->loi_read_lop.lop_num_pending,                      \
2531                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
2532                args)                                                     \
2533
2534 /* This is called by osc_check_rpcs() to find which objects have pages that
2535  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2536 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2537 {
2538         ENTRY;
2539
2540         /* First return objects that have blocked locks so that they
2541          * will be flushed quickly and other clients can get the lock,
2542          * then objects which have pages ready to be stuffed into RPCs */
2543         if (!list_empty(&cli->cl_loi_hp_ready_list))
2544                 RETURN(list_entry(cli->cl_loi_hp_ready_list.next,
2545                                   struct lov_oinfo, loi_hp_ready_item));
2546         if (!list_empty(&cli->cl_loi_ready_list))
2547                 RETURN(list_entry(cli->cl_loi_ready_list.next,
2548                                   struct lov_oinfo, loi_ready_item));
2549
2550         /* then if we have cache waiters, return all objects with queued
2551          * writes.  This is especially important when many small files
2552          * have filled up the cache and not been fired into rpcs because
2553          * they don't pass the nr_pending/object threshhold */
2554         if (!list_empty(&cli->cl_cache_waiters) &&
2555             !list_empty(&cli->cl_loi_write_list))
2556                 RETURN(list_entry(cli->cl_loi_write_list.next,
2557                                   struct lov_oinfo, loi_write_item));
2558
2559         /* then return all queued objects when we have an invalid import
2560          * so that they get flushed */
2561         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2562                 if (!list_empty(&cli->cl_loi_write_list))
2563                         RETURN(list_entry(cli->cl_loi_write_list.next,
2564                                           struct lov_oinfo, loi_write_item));
2565                 if (!list_empty(&cli->cl_loi_read_list))
2566                         RETURN(list_entry(cli->cl_loi_read_list.next,
2567                                           struct lov_oinfo, loi_read_item));
2568         }
2569         RETURN(NULL);
2570 }
2571
2572 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2573 {
2574         struct osc_async_page *oap;
2575         int hprpc = 0;
2576
2577         if (!list_empty(&loi->loi_write_lop.lop_urgent)) {
2578                 oap = list_entry(loi->loi_write_lop.lop_urgent.next,
2579                                  struct osc_async_page, oap_urgent_item);
2580                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2581         }
2582
2583         if (!hprpc && !list_empty(&loi->loi_read_lop.lop_urgent)) {
2584                 oap = list_entry(loi->loi_write_lop.lop_urgent.next,
2585                                  struct osc_async_page, oap_urgent_item);
2586                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2587         }
2588
2589         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2590 }
2591
2592 /* called with the loi list lock held */
2593 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2594 {
2595         struct lov_oinfo *loi;
2596         int rc = 0, race_counter = 0;
2597         ENTRY;
2598
2599         while ((loi = osc_next_loi(cli)) != NULL) {
2600                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2601
2602                 if (osc_max_rpc_in_flight(cli, loi))
2603                         break;
2604
2605                 /* attempt some read/write balancing by alternating between
2606                  * reads and writes in an object.  The makes_rpc checks here
2607                  * would be redundant if we were getting read/write work items
2608                  * instead of objects.  we don't want send_oap_rpc to drain a
2609                  * partial read pending queue when we're given this object to
2610                  * do io on writes while there are cache waiters */
2611                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2612                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2613                                               &loi->loi_write_lop);
2614                         if (rc < 0)
2615                                 break;
2616                         if (rc > 0)
2617                                 race_counter = 0;
2618                         else
2619                                 race_counter++;
2620                 }
2621                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2622                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2623                                               &loi->loi_read_lop);
2624                         if (rc < 0)
2625                                 break;
2626                         if (rc > 0)
2627                                 race_counter = 0;
2628                         else
2629                                 race_counter++;
2630                 }
2631
2632                 /* attempt some inter-object balancing by issueing rpcs
2633                  * for each object in turn */
2634                 if (!list_empty(&loi->loi_hp_ready_item))
2635                         list_del_init(&loi->loi_hp_ready_item);
2636                 if (!list_empty(&loi->loi_ready_item))
2637                         list_del_init(&loi->loi_ready_item);
2638                 if (!list_empty(&loi->loi_write_item))
2639                         list_del_init(&loi->loi_write_item);
2640                 if (!list_empty(&loi->loi_read_item))
2641                         list_del_init(&loi->loi_read_item);
2642
2643                 loi_list_maint(cli, loi);
2644
2645                 /* send_oap_rpc fails with 0 when make_ready tells it to
2646                  * back off.  llite's make_ready does this when it tries
2647                  * to lock a page queued for write that is already locked.
2648                  * we want to try sending rpcs from many objects, but we
2649                  * don't want to spin failing with 0.  */
2650                 if (race_counter == 10)
2651                         break;
2652         }
2653         EXIT;
2654 }
2655
2656 /* we're trying to queue a page in the osc so we're subject to the
2657  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2658  * If the osc's queued pages are already at that limit, then we want to sleep
2659  * until there is space in the osc's queue for us.  We also may be waiting for
2660  * write credits from the OST if there are RPCs in flight that may return some
2661  * before we fall back to sync writes.
2662  *
2663  * We need this know our allocation was granted in the presence of signals */
2664 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2665 {
2666         int rc;
2667         ENTRY;
2668         client_obd_list_lock(&cli->cl_loi_list_lock);
2669         rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2670         client_obd_list_unlock(&cli->cl_loi_list_lock);
2671         RETURN(rc);
2672 };
2673
2674 /**
2675  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2676  * is available.
2677  */
2678 int osc_enter_cache_try(const struct lu_env *env,
2679                         struct client_obd *cli, struct lov_oinfo *loi,
2680                         struct osc_async_page *oap, int transient)
2681 {
2682         int has_grant;
2683
2684         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2685         if (has_grant) {
2686                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2687                 if (transient) {
2688                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2689                         atomic_inc(&obd_dirty_transit_pages);
2690                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2691                 }
2692         }
2693         return has_grant;
2694 }
2695
2696 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2697  * grant or cache space. */
2698 static int osc_enter_cache(const struct lu_env *env,
2699                            struct client_obd *cli, struct lov_oinfo *loi,
2700                            struct osc_async_page *oap)
2701 {
2702         struct osc_cache_waiter ocw;
2703         struct l_wait_info lwi = { 0 };
2704
2705         ENTRY;
2706
2707         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2708                "grant: %lu\n", cli->cl_dirty, atomic_read(&obd_dirty_pages),
2709                cli->cl_dirty_max, obd_max_dirty_pages,
2710                cli->cl_lost_grant, cli->cl_avail_grant);
2711
2712         /* force the caller to try sync io.  this can jump the list
2713          * of queued writes and create a discontiguous rpc stream */
2714         if (cli->cl_dirty_max < CFS_PAGE_SIZE || cli->cl_ar.ar_force_sync ||
2715             loi->loi_ar.ar_force_sync)
2716                 RETURN(-EDQUOT);
2717
2718         /* Hopefully normal case - cache space and write credits available */
2719         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2720             atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2721             osc_enter_cache_try(env, cli, loi, oap, 0))
2722                 RETURN(0);
2723
2724         /* Make sure that there are write rpcs in flight to wait for.  This
2725          * is a little silly as this object may not have any pending but
2726          * other objects sure might. */
2727         if (cli->cl_w_in_flight) {
2728                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2729                 cfs_waitq_init(&ocw.ocw_waitq);
2730                 ocw.ocw_oap = oap;
2731                 ocw.ocw_rc = 0;
2732
2733                 loi_list_maint(cli, loi);
2734                 osc_check_rpcs(env, cli);
2735                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2736
2737                 CDEBUG(D_CACHE, "sleeping for cache space\n");
2738                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2739
2740                 client_obd_list_lock(&cli->cl_loi_list_lock);
2741                 if (!list_empty(&ocw.ocw_entry)) {
2742                         list_del(&ocw.ocw_entry);
2743                         RETURN(-EINTR);
2744                 }
2745                 RETURN(ocw.ocw_rc);
2746         }
2747
2748         RETURN(-EDQUOT);
2749 }
2750
2751
2752 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2753                         struct lov_oinfo *loi, cfs_page_t *page,
2754                         obd_off offset, const struct obd_async_page_ops *ops,
2755                         void *data, void **res, int nocache,
2756                         struct lustre_handle *lockh)
2757 {
2758         struct osc_async_page *oap;
2759
2760         ENTRY;
2761
2762         if (!page)
2763                 return size_round(sizeof(*oap));
2764
2765         oap = *res;
2766         oap->oap_magic = OAP_MAGIC;
2767         oap->oap_cli = &exp->exp_obd->u.cli;
2768         oap->oap_loi = loi;
2769
2770         oap->oap_caller_ops = ops;
2771         oap->oap_caller_data = data;
2772
2773         oap->oap_page = page;
2774         oap->oap_obj_off = offset;
2775         if (!client_is_remote(exp) &&
2776             cfs_capable(CFS_CAP_SYS_RESOURCE))
2777                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2778
2779         LASSERT(!(offset & ~CFS_PAGE_MASK));
2780
2781         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2782         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2783         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2784         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2785
2786         spin_lock_init(&oap->oap_lock);
2787         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2788         RETURN(0);
2789 }
2790
2791 struct osc_async_page *oap_from_cookie(void *cookie)
2792 {
2793         struct osc_async_page *oap = cookie;
2794         if (oap->oap_magic != OAP_MAGIC)
2795                 return ERR_PTR(-EINVAL);
2796         return oap;
2797 };
2798
2799 int osc_queue_async_io(const struct lu_env *env,
2800                        struct obd_export *exp, struct lov_stripe_md *lsm,
2801                        struct lov_oinfo *loi, void *cookie,
2802                        int cmd, obd_off off, int count,
2803                        obd_flag brw_flags, enum async_flags async_flags)
2804 {
2805         struct client_obd *cli = &exp->exp_obd->u.cli;
2806         struct osc_async_page *oap;
2807         int rc = 0;
2808         ENTRY;
2809
2810         oap = oap_from_cookie(cookie);
2811         if (IS_ERR(oap))
2812                 RETURN(PTR_ERR(oap));
2813
2814         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2815                 RETURN(-EIO);
2816
2817         if (!list_empty(&oap->oap_pending_item) ||
2818             !list_empty(&oap->oap_urgent_item) ||
2819             !list_empty(&oap->oap_rpc_item))
2820                 RETURN(-EBUSY);
2821
2822         /* check if the file's owner/group is over quota */
2823         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2824                 struct cl_object *obj;
2825                 struct cl_attr    attr; /* XXX put attr into thread info */
2826
2827                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2828
2829                 cl_object_attr_lock(obj);
2830                 rc = cl_object_attr_get(env, obj, &attr);
2831                 cl_object_attr_unlock(obj);
2832
2833                 if (rc == 0 && lquota_chkdq(quota_interface, cli, attr.cat_uid,
2834                                             attr.cat_gid) == NO_QUOTA)
2835                         rc = -EDQUOT;
2836                 if (rc)
2837                         RETURN(rc);
2838         }
2839
2840         if (loi == NULL)
2841                 loi = lsm->lsm_oinfo[0];
2842
2843         client_obd_list_lock(&cli->cl_loi_list_lock);
2844
2845         LASSERT(off + count <= CFS_PAGE_SIZE);
2846         oap->oap_cmd = cmd;
2847         oap->oap_page_off = off;
2848         oap->oap_count = count;
2849         oap->oap_brw_flags = brw_flags;
2850         oap->oap_async_flags = async_flags;
2851
2852         if (cmd & OBD_BRW_WRITE) {
2853                 rc = osc_enter_cache(env, cli, loi, oap);
2854                 if (rc) {
2855                         client_obd_list_unlock(&cli->cl_loi_list_lock);
2856                         RETURN(rc);
2857                 }
2858         }
2859
2860         osc_oap_to_pending(oap);
2861         loi_list_maint(cli, loi);
2862
2863         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2864                   cmd);
2865
2866         osc_check_rpcs(env, cli);
2867         client_obd_list_unlock(&cli->cl_loi_list_lock);
2868
2869         RETURN(0);
2870 }
2871
2872 /* aka (~was & now & flag), but this is more clear :) */
2873 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2874
2875 int osc_set_async_flags_base(struct client_obd *cli,
2876                              struct lov_oinfo *loi, struct osc_async_page *oap,
2877                              obd_flag async_flags)
2878 {
2879         struct loi_oap_pages *lop;
2880         ENTRY;
2881
2882         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2883                 RETURN(-EIO);
2884
2885         if (oap->oap_cmd & OBD_BRW_WRITE) {
2886                 lop = &loi->loi_write_lop;
2887         } else {
2888                 lop = &loi->loi_read_lop;
2889         }
2890
2891         if (list_empty(&oap->oap_pending_item))
2892                 RETURN(-EINVAL);
2893
2894         if ((oap->oap_async_flags & async_flags) == async_flags)
2895                 RETURN(0);
2896
2897         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2898                 oap->oap_async_flags |= ASYNC_READY;
2899
2900         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
2901             list_empty(&oap->oap_rpc_item)) {
2902                 if (oap->oap_async_flags & ASYNC_HP)
2903                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2904                 else
2905                         list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2906                 oap->oap_async_flags |= ASYNC_URGENT;
2907                 loi_list_maint(cli, loi);
2908         }
2909
2910         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2911                         oap->oap_async_flags);
2912         RETURN(0);
2913 }
2914
2915 int osc_teardown_async_page(struct obd_export *exp,
2916                             struct lov_stripe_md *lsm,
2917                             struct lov_oinfo *loi, void *cookie)
2918 {
2919         struct client_obd *cli = &exp->exp_obd->u.cli;
2920         struct loi_oap_pages *lop;
2921         struct osc_async_page *oap;
2922         int rc = 0;
2923         ENTRY;
2924
2925         oap = oap_from_cookie(cookie);
2926         if (IS_ERR(oap))
2927                 RETURN(PTR_ERR(oap));
2928
2929         if (loi == NULL)
2930                 loi = lsm->lsm_oinfo[0];
2931
2932         if (oap->oap_cmd & OBD_BRW_WRITE) {
2933                 lop = &loi->loi_write_lop;
2934         } else {
2935                 lop = &loi->loi_read_lop;
2936         }
2937
2938         client_obd_list_lock(&cli->cl_loi_list_lock);
2939
2940         if (!list_empty(&oap->oap_rpc_item))
2941                 GOTO(out, rc = -EBUSY);
2942
2943         osc_exit_cache(cli, oap, 0);
2944         osc_wake_cache_waiters(cli);
2945
2946         if (!list_empty(&oap->oap_urgent_item)) {
2947                 list_del_init(&oap->oap_urgent_item);
2948                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
2949         }
2950         if (!list_empty(&oap->oap_pending_item)) {
2951                 list_del_init(&oap->oap_pending_item);
2952                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2953         }
2954         loi_list_maint(cli, loi);
2955         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2956 out:
2957         client_obd_list_unlock(&cli->cl_loi_list_lock);
2958         RETURN(rc);
2959 }
2960
2961 static void osc_set_lock_data_with_check(struct ldlm_lock *lock,
2962                                          struct ldlm_enqueue_info *einfo,
2963                                          int flags)
2964 {
2965         void *data = einfo->ei_cbdata;
2966
2967         LASSERT(lock != NULL);
2968         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2969         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2970         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2971         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2972
2973         lock_res_and_lock(lock);
2974         spin_lock(&osc_ast_guard);
2975         LASSERT(lock->l_ast_data == NULL || lock->l_ast_data == data);
2976         lock->l_ast_data = data;
2977         spin_unlock(&osc_ast_guard);
2978         unlock_res_and_lock(lock);
2979 }
2980
2981 static void osc_set_data_with_check(struct lustre_handle *lockh,
2982                                     struct ldlm_enqueue_info *einfo,
2983                                     int flags)
2984 {
2985         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2986
2987         if (lock != NULL) {
2988                 osc_set_lock_data_with_check(lock, einfo, flags);
2989                 LDLM_LOCK_PUT(lock);
2990         } else
2991                 CERROR("lockh %p, data %p - client evicted?\n",
2992                        lockh, einfo->ei_cbdata);
2993 }
2994
2995 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2996                              ldlm_iterator_t replace, void *data)
2997 {
2998         struct ldlm_res_id res_id;
2999         struct obd_device *obd = class_exp2obd(exp);
3000
3001         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_gr, &res_id);
3002         ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3003         return 0;
3004 }
3005
3006 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3007                             obd_enqueue_update_f upcall, void *cookie,
3008                             int *flags, int rc)
3009 {
3010         int intent = *flags & LDLM_FL_HAS_INTENT;
3011         ENTRY;
3012
3013         if (intent) {
3014                 /* The request was created before ldlm_cli_enqueue call. */
3015                 if (rc == ELDLM_LOCK_ABORTED) {
3016                         struct ldlm_reply *rep;
3017                         rep = req_capsule_server_get(&req->rq_pill,
3018                                                      &RMF_DLM_REP);
3019
3020                         LASSERT(rep != NULL);
3021                         if (rep->lock_policy_res1)
3022                                 rc = rep->lock_policy_res1;
3023                 }
3024         }
3025
3026         if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3027                 *flags |= LDLM_FL_LVB_READY;
3028                 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3029                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3030         }
3031
3032         /* Call the update callback. */
3033         rc = (*upcall)(cookie, rc);
3034         RETURN(rc);
3035 }
3036
3037 static int osc_enqueue_interpret(const struct lu_env *env,
3038                                  struct ptlrpc_request *req,
3039                                  struct osc_enqueue_args *aa, int rc)
3040 {
3041         struct ldlm_lock *lock;
3042         struct lustre_handle handle;
3043         __u32 mode;
3044
3045         /* Make a local copy of a lock handle and a mode, because aa->oa_*
3046          * might be freed anytime after lock upcall has been called. */
3047         lustre_handle_copy(&handle, aa->oa_lockh);
3048         mode = aa->oa_ei->ei_mode;
3049
3050         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3051          * be valid. */
3052         lock = ldlm_handle2lock(&handle);
3053
3054         /* Take an additional reference so that a blocking AST that
3055          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3056          * to arrive after an upcall has been executed by
3057          * osc_enqueue_fini(). */
3058         ldlm_lock_addref(&handle, mode);
3059
3060         /* Complete obtaining the lock procedure. */
3061         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3062                                    mode, aa->oa_flags, aa->oa_lvb,
3063                                    sizeof(*aa->oa_lvb), lustre_swab_ost_lvb,
3064                                    &handle, rc);
3065         /* Complete osc stuff. */
3066         rc = osc_enqueue_fini(req, aa->oa_lvb,
3067                               aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3068
3069         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3070
3071         /* Release the lock for async request. */
3072         if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3073                 /*
3074                  * Releases a reference taken by ldlm_cli_enqueue(), if it is
3075                  * not already released by
3076                  * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3077                  */
3078                 ldlm_lock_decref(&handle, mode);
3079
3080         LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3081                  aa->oa_lockh, req, aa);
3082         ldlm_lock_decref(&handle, mode);
3083         LDLM_LOCK_PUT(lock);
3084         return rc;
3085 }
3086
3087 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3088                         struct lov_oinfo *loi, int flags,
3089                         struct ost_lvb *lvb, __u32 mode, int rc)
3090 {
3091         if (rc == ELDLM_OK) {
3092                 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3093                 __u64 tmp;
3094
3095                 LASSERT(lock != NULL);
3096                 loi->loi_lvb = *lvb;
3097                 tmp = loi->loi_lvb.lvb_size;
3098                 /* Extend KMS up to the end of this lock and no further
3099                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3100                 if (tmp > lock->l_policy_data.l_extent.end)
3101                         tmp = lock->l_policy_data.l_extent.end + 1;
3102                 if (tmp >= loi->loi_kms) {
3103                         LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3104                                    ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3105                         loi_kms_set(loi, tmp);
3106                 } else {
3107                         LDLM_DEBUG(lock, "lock acquired, setting rss="
3108                                    LPU64"; leaving kms="LPU64", end="LPU64,
3109                                    loi->loi_lvb.lvb_size, loi->loi_kms,
3110                                    lock->l_policy_data.l_extent.end);
3111                 }
3112                 ldlm_lock_allow_match(lock);
3113                 LDLM_LOCK_PUT(lock);
3114         } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3115                 loi->loi_lvb = *lvb;
3116                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3117                        " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3118                 rc = ELDLM_OK;
3119         }
3120 }
3121 EXPORT_SYMBOL(osc_update_enqueue);
3122
3123 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3124
3125 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3126  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3127  * other synchronous requests, however keeping some locks and trying to obtain
3128  * others may take a considerable amount of time in a case of ost failure; and
3129  * when other sync requests do not get released lock from a client, the client
3130  * is excluded from the cluster -- such scenarious make the life difficult, so
3131  * release locks just after they are obtained. */
3132 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3133                      int *flags, ldlm_policy_data_t *policy,
3134                      struct ost_lvb *lvb, int kms_valid,
3135                      obd_enqueue_update_f upcall, void *cookie,
3136                      struct ldlm_enqueue_info *einfo,
3137                      struct lustre_handle *lockh,
3138                      struct ptlrpc_request_set *rqset, int async)
3139 {
3140         struct obd_device *obd = exp->exp_obd;
3141         struct ptlrpc_request *req = NULL;
3142         int intent = *flags & LDLM_FL_HAS_INTENT;
3143         ldlm_mode_t mode;
3144         int rc;
3145         ENTRY;
3146
3147         /* Filesystem lock extents are extended to page boundaries so that
3148          * dealing with the page cache is a little smoother.  */
3149         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3150         policy->l_extent.end |= ~CFS_PAGE_MASK;
3151
3152         /*
3153          * kms is not valid when either object is completely fresh (so that no
3154          * locks are cached), or object was evicted. In the latter case cached
3155          * lock cannot be used, because it would prime inode state with
3156          * potentially stale LVB.
3157          */
3158         if (!kms_valid)
3159                 goto no_match;
3160
3161         /* Next, search for already existing extent locks that will cover us */
3162         /* If we're trying to read, we also search for an existing PW lock.  The
3163          * VFS and page cache already protect us locally, so lots of readers/
3164          * writers can share a single PW lock.
3165          *
3166          * There are problems with conversion deadlocks, so instead of
3167          * converting a read lock to a write lock, we'll just enqueue a new
3168          * one.
3169          *
3170          * At some point we should cancel the read lock instead of making them
3171          * send us a blocking callback, but there are problems with canceling
3172          * locks out from other users right now, too. */
3173         mode = einfo->ei_mode;
3174         if (einfo->ei_mode == LCK_PR)
3175                 mode |= LCK_PW;
3176         mode = ldlm_lock_match(obd->obd_namespace,
3177                                *flags | LDLM_FL_LVB_READY, res_i