lustre/osc/osc_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  */
  39
  40 #ifndef EXPORT_SYMTAB
  41 # define EXPORT_SYMTAB
  42 #endif
  43 #define DEBUG_SUBSYSTEM S_OSC
  44
  45 #include <libcfs/libcfs.h>
  46
  47 #ifndef __KERNEL__
  48 # include <liblustre.h>
  49 #endif
  50
  51 #include <lustre_dlm.h>
  52 #include <lustre_net.h>
  53 #include <lustre/lustre_user.h>
  54 #include <obd_cksum.h>
  55 #include <obd_ost.h>
  56 #include <obd_lov.h>
  57
  58 #ifdef  __CYGWIN__
  59 # include <ctype.h>
  60 #endif
  61
  62 #include <lustre_ha.h>
  63 #include <lprocfs_status.h>
  64 #include <lustre_log.h>
  65 #include <lustre_debug.h>
  66 #include <lustre_param.h>
  67 #include "osc_internal.h"
  68
  69 static quota_interface_t *quota_interface = NULL;
  70 extern quota_interface_t osc_quota_interface;
  71
  72 static void osc_release_ppga(struct brw_page **ppga, obd_count count);
  73 static int brw_interpret(const struct lu_env *env,
  74                          struct ptlrpc_request *req, void *data, int rc);
  75 int osc_cleanup(struct obd_device *obd);
  76
  77 /* Pack OSC object metadata for disk storage (LE byte order). */
  78 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
  79                       struct lov_stripe_md *lsm)
  80 {
  81         int lmm_size;
  82         ENTRY;
  83
  84         lmm_size = sizeof(**lmmp);
  85         if (!lmmp)
  86                 RETURN(lmm_size);
  87
  88         if (*lmmp && !lsm) {
  89                 OBD_FREE(*lmmp, lmm_size);
  90                 *lmmp = NULL;
  91                 RETURN(0);
  92         }
  93
  94         if (!*lmmp) {
  95                 OBD_ALLOC(*lmmp, lmm_size);
  96                 if (!*lmmp)
  97                         RETURN(-ENOMEM);
  98         }
  99
 100         if (lsm) {
 101                 LASSERT(lsm->lsm_object_id);
 102                 LASSERT_SEQ_IS_MDT(lsm->lsm_object_seq);
 103                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
 104                 (*lmmp)->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
 105         }
 106
 107         RETURN(lmm_size);
 108 }
 109
 110 /* Unpack OSC object metadata from disk storage (LE byte order). */
 111 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
 112                         struct lov_mds_md *lmm, int lmm_bytes)
 113 {
 114         int lsm_size;
 115         struct obd_import *imp = class_exp2cliimp(exp);
 116         ENTRY;
 117
 118         if (lmm != NULL) {
 119                 if (lmm_bytes < sizeof (*lmm)) {
 120                         CERROR("lov_mds_md too small: %d, need %d\n",
 121                                lmm_bytes, (int)sizeof(*lmm));
 122                         RETURN(-EINVAL);
 123                 }
 124                 /* XXX LOV_MAGIC etc check? */
 125
 126                 if (lmm->lmm_object_id == 0) {
 127                         CERROR("lov_mds_md: zero lmm_object_id\n");
 128                         RETURN(-EINVAL);
 129                 }
 130         }
 131
 132         lsm_size = lov_stripe_md_size(1);
 133         if (lsmp == NULL)
 134                 RETURN(lsm_size);
 135
 136         if (*lsmp != NULL && lmm == NULL) {
 137                 OBD_FREE((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
 138                 OBD_FREE(*lsmp, lsm_size);
 139                 *lsmp = NULL;
 140                 RETURN(0);
 141         }
 142
 143         if (*lsmp == NULL) {
 144                 OBD_ALLOC(*lsmp, lsm_size);
 145                 if (*lsmp == NULL)
 146                         RETURN(-ENOMEM);
 147                 OBD_ALLOC((*lsmp)->lsm_oinfo[0], sizeof(struct lov_oinfo));
 148                 if ((*lsmp)->lsm_oinfo[0] == NULL) {
 149                         OBD_FREE(*lsmp, lsm_size);
 150                         RETURN(-ENOMEM);
 151                 }
 152                 loi_init((*lsmp)->lsm_oinfo[0]);
 153         }
 154
 155         if (lmm != NULL) {
 156                 /* XXX zero *lsmp? */
 157                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
 158                 (*lsmp)->lsm_object_seq = le64_to_cpu (lmm->lmm_object_seq);
 159                 LASSERT((*lsmp)->lsm_object_id);
 160                 LASSERT_SEQ_IS_MDT((*lsmp)->lsm_object_seq);
 161         }
 162
 163         if (imp != NULL &&
 164             (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
 165                 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
 166         else
 167                 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
 168
 169         RETURN(lsm_size);
 170 }
 171
 172 static inline void osc_pack_capa(struct ptlrpc_request *req,
 173                                  struct ost_body *body, void *capa)
 174 {
 175         struct obd_capa *oc = (struct obd_capa *)capa;
 176         struct lustre_capa *c;
 177
 178         if (!capa)
 179                 return;
 180
 181         c = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
 182         LASSERT(c);
 183         capa_cpy(c, oc);
 184         body->oa.o_valid |= OBD_MD_FLOSSCAPA;
 185         DEBUG_CAPA(D_SEC, c, "pack");
 186 }
 187
 188 static inline void osc_pack_req_body(struct ptlrpc_request *req,
 189                                      struct obd_info *oinfo)
 190 {
 191         struct ost_body *body;
 192
 193         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 194         LASSERT(body);
 195
 196         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
 197         osc_pack_capa(req, body, oinfo->oi_capa);
 198 }
 199
 200 static inline void osc_set_capa_size(struct ptlrpc_request *req,
 201                                      const struct req_msg_field *field,
 202                                      struct obd_capa *oc)
 203 {
 204         if (oc == NULL)
 205                 req_capsule_set_size(&req->rq_pill, field, RCL_CLIENT, 0);
 206         else
 207                 /* it is already calculated as sizeof struct obd_capa */
 208                 ;
 209 }
 210
 211 static int osc_getattr_interpret(const struct lu_env *env,
 212                                  struct ptlrpc_request *req,
 213                                  struct osc_async_args *aa, int rc)
 214 {
 215         struct ost_body *body;
 216         ENTRY;
 217
 218         if (rc != 0)
 219                 GOTO(out, rc);
 220
 221         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 222         if (body) {
 223                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 224                 lustre_get_wire_obdo(aa->aa_oi->oi_oa, &body->oa);
 225
 226                 /* This should really be sent by the OST */
 227                 aa->aa_oi->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 228                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
 229         } else {
 230                 CDEBUG(D_INFO, "can't unpack ost_body\n");
 231                 rc = -EPROTO;
 232                 aa->aa_oi->oi_oa->o_valid = 0;
 233         }
 234 out:
 235         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
 236         RETURN(rc);
 237 }
 238
 239 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
 240                              struct ptlrpc_request_set *set)
 241 {
 242         struct ptlrpc_request *req;
 243         struct osc_async_args *aa;
 244         int                    rc;
 245         ENTRY;
 246
 247         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
 248         if (req == NULL)
 249                 RETURN(-ENOMEM);
 250
 251         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
 252         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
 253         if (rc) {
 254                 ptlrpc_request_free(req);
 255                 RETURN(rc);
 256         }
 257
 258         osc_pack_req_body(req, oinfo);
 259
 260         ptlrpc_request_set_replen(req);
 261         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
 262
 263         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
 264         aa = ptlrpc_req_async_args(req);
 265         aa->aa_oi = oinfo;
 266
 267         ptlrpc_set_add_req(set, req);
 268         RETURN(0);
 269 }
 270
 271 static int osc_getattr(struct obd_export *exp, struct obd_info *oinfo)
 272 {
 273         struct ptlrpc_request *req;
 274         struct ost_body       *body;
 275         int                    rc;
 276         ENTRY;
 277
 278         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
 279         if (req == NULL)
 280                 RETURN(-ENOMEM);
 281
 282         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
 283         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
 284         if (rc) {
 285                 ptlrpc_request_free(req);
 286                 RETURN(rc);
 287         }
 288
 289         osc_pack_req_body(req, oinfo);
 290
 291         ptlrpc_request_set_replen(req);
 292
 293         rc = ptlrpc_queue_wait(req);
 294         if (rc)
 295                 GOTO(out, rc);
 296
 297         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 298         if (body == NULL)
 299                 GOTO(out, rc = -EPROTO);
 300
 301         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 302         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
 303
 304         /* This should really be sent by the OST */
 305         oinfo->oi_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 306         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
 307
 308         EXIT;
 309  out:
 310         ptlrpc_req_finished(req);
 311         return rc;
 312 }
 313
 314 static int osc_setattr(struct obd_export *exp, struct obd_info *oinfo,
 315                        struct obd_trans_info *oti)
 316 {
 317         struct ptlrpc_request *req;
 318         struct ost_body       *body;
 319         int                    rc;
 320         ENTRY;
 321
 322         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
 323
 324         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 325         if (req == NULL)
 326                 RETURN(-ENOMEM);
 327
 328         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
 329         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 330         if (rc) {
 331                 ptlrpc_request_free(req);
 332                 RETURN(rc);
 333         }
 334
 335         osc_pack_req_body(req, oinfo);
 336
 337         ptlrpc_request_set_replen(req);
 338
 339         rc = ptlrpc_queue_wait(req);
 340         if (rc)
 341                 GOTO(out, rc);
 342
 343         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 344         if (body == NULL)
 345                 GOTO(out, rc = -EPROTO);
 346
 347         lustre_get_wire_obdo(oinfo->oi_oa, &body->oa);
 348
 349         EXIT;
 350 out:
 351         ptlrpc_req_finished(req);
 352         RETURN(rc);
 353 }
 354
 355 static int osc_setattr_interpret(const struct lu_env *env,
 356                                  struct ptlrpc_request *req,
 357                                  struct osc_setattr_args *sa, int rc)
 358 {
 359         struct ost_body *body;
 360         ENTRY;
 361
 362         if (rc != 0)
 363                 GOTO(out, rc);
 364
 365         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 366         if (body == NULL)
 367                 GOTO(out, rc = -EPROTO);
 368
 369         lustre_get_wire_obdo(sa->sa_oa, &body->oa);
 370 out:
 371         rc = sa->sa_upcall(sa->sa_cookie, rc);
 372         RETURN(rc);
 373 }
 374
 375 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
 376                            struct obd_trans_info *oti,
 377                            obd_enqueue_update_f upcall, void *cookie,
 378                            struct ptlrpc_request_set *rqset)
 379 {
 380         struct ptlrpc_request   *req;
 381         struct osc_setattr_args *sa;
 382         int                      rc;
 383         ENTRY;
 384
 385         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 386         if (req == NULL)
 387                 RETURN(-ENOMEM);
 388
 389         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
 390         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 391         if (rc) {
 392                 ptlrpc_request_free(req);
 393                 RETURN(rc);
 394         }
 395
 396         if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
 397                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
 398
 399         osc_pack_req_body(req, oinfo);
 400
 401         ptlrpc_request_set_replen(req);
 402
 403         /* do mds to ost setattr asynchronously */
 404         if (!rqset) {
 405                 /* Do not wait for response. */
 406                 ptlrpcd_add_req(req, PSCOPE_OTHER);
 407         } else {
 408                 req->rq_interpret_reply =
 409                         (ptlrpc_interpterer_t)osc_setattr_interpret;
 410
 411                 CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
 412                 sa = ptlrpc_req_async_args(req);
 413                 sa->sa_oa = oinfo->oi_oa;
 414                 sa->sa_upcall = upcall;
 415                 sa->sa_cookie = cookie;
 416
 417                 if (rqset == PTLRPCD_SET)
 418                         ptlrpcd_add_req(req, PSCOPE_OTHER);
 419                 else
 420                         ptlrpc_set_add_req(rqset, req);
 421         }
 422
 423         RETURN(0);
 424 }
 425
 426 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
 427                              struct obd_trans_info *oti,
 428                              struct ptlrpc_request_set *rqset)
 429 {
 430         return osc_setattr_async_base(exp, oinfo, oti,
 431                                       oinfo->oi_cb_up, oinfo, rqset);
 432 }
 433
 434 int osc_real_create(struct obd_export *exp, struct obdo *oa,
 435                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
 436 {
 437         struct ptlrpc_request *req;
 438         struct ost_body       *body;
 439         struct lov_stripe_md  *lsm;
 440         int                    rc;
 441         ENTRY;
 442
 443         LASSERT(oa);
 444         LASSERT(ea);
 445
 446         lsm = *ea;
 447         if (!lsm) {
 448                 rc = obd_alloc_memmd(exp, &lsm);
 449                 if (rc < 0)
 450                         RETURN(rc);
 451         }
 452
 453         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
 454         if (req == NULL)
 455                 GOTO(out, rc = -ENOMEM);
 456
 457         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 458         if (rc) {
 459                 ptlrpc_request_free(req);
 460                 GOTO(out, rc);
 461         }
 462
 463         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 464         LASSERT(body);
 465         lustre_set_wire_obdo(&body->oa, oa);
 466
 467         ptlrpc_request_set_replen(req);
 468
 469         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 470             oa->o_flags == OBD_FL_DELORPHAN) {
 471                 DEBUG_REQ(D_HA, req,
 472                           "delorphan from OST integration");
 473                 /* Don't resend the delorphan req */
 474                 req->rq_no_resend = req->rq_no_delay = 1;
 475         }
 476
 477         rc = ptlrpc_queue_wait(req);
 478         if (rc)
 479                 GOTO(out_req, rc);
 480
 481         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 482         if (body == NULL)
 483                 GOTO(out_req, rc = -EPROTO);
 484
 485         lustre_get_wire_obdo(oa, &body->oa);
 486
 487         /* This should really be sent by the OST */
 488         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 489         oa->o_valid |= OBD_MD_FLBLKSZ;
 490
 491         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
 492          * have valid lsm_oinfo data structs, so don't go touching that.
 493          * This needs to be fixed in a big way.
 494          */
 495         lsm->lsm_object_id = oa->o_id;
 496         lsm->lsm_object_seq = oa->o_seq;
 497         *ea = lsm;
 498
 499         if (oti != NULL) {
 500                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
 501
 502                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
 503                         if (!oti->oti_logcookies)
 504                                 oti_alloc_cookies(oti, 1);
 505                         *oti->oti_logcookies = oa->o_lcookie;
 506                 }
 507         }
 508
 509         CDEBUG(D_HA, "transno: "LPD64"\n",
 510                lustre_msg_get_transno(req->rq_repmsg));
 511 out_req:
 512         ptlrpc_req_finished(req);
 513 out:
 514         if (rc && !*ea)
 515                 obd_free_memmd(exp, &lsm);
 516         RETURN(rc);
 517 }
 518
 519 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
 520                    obd_enqueue_update_f upcall, void *cookie,
 521                    struct ptlrpc_request_set *rqset)
 522 {
 523         struct ptlrpc_request   *req;
 524         struct osc_setattr_args *sa;
 525         struct ost_body         *body;
 526         int                      rc;
 527         ENTRY;
 528
 529         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
 530         if (req == NULL)
 531                 RETURN(-ENOMEM);
 532
 533         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
 534         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
 535         if (rc) {
 536                 ptlrpc_request_free(req);
 537                 RETURN(rc);
 538         }
 539         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 540         ptlrpc_at_set_req_timeout(req);
 541
 542         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 543         LASSERT(body);
 544         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
 545         osc_pack_capa(req, body, oinfo->oi_capa);
 546
 547         ptlrpc_request_set_replen(req);
 548
 549
 550         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
 551         CLASSERT (sizeof(*sa) <= sizeof(req->rq_async_args));
 552         sa = ptlrpc_req_async_args(req);
 553         sa->sa_oa     = oinfo->oi_oa;
 554         sa->sa_upcall = upcall;
 555         sa->sa_cookie = cookie;
 556         if (rqset == PTLRPCD_SET)
 557                 ptlrpcd_add_req(req, PSCOPE_OTHER);
 558         else
 559                 ptlrpc_set_add_req(rqset, req);
 560
 561         RETURN(0);
 562 }
 563
 564 static int osc_punch(struct obd_export *exp, struct obd_info *oinfo,
 565                      struct obd_trans_info *oti,
 566                      struct ptlrpc_request_set *rqset)
 567 {
 568         oinfo->oi_oa->o_size   = oinfo->oi_policy.l_extent.start;
 569         oinfo->oi_oa->o_blocks = oinfo->oi_policy.l_extent.end;
 570         oinfo->oi_oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 571         return osc_punch_base(exp, oinfo,
 572                               oinfo->oi_cb_up, oinfo, rqset);
 573 }
 574
 575 static int osc_sync_interpret(const struct lu_env *env,
 576                               struct ptlrpc_request *req,
 577                               void *arg, int rc)
 578 {
 579         struct osc_async_args *aa = arg;
 580         struct ost_body *body;
 581         ENTRY;
 582
 583         if (rc)
 584                 GOTO(out, rc);
 585
 586         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 587         if (body == NULL) {
 588                 CERROR ("can't unpack ost_body\n");
 589                 GOTO(out, rc = -EPROTO);
 590         }
 591
 592         *aa->aa_oi->oi_oa = body->oa;
 593 out:
 594         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
 595         RETURN(rc);
 596 }
 597
 598 static int osc_sync(struct obd_export *exp, struct obd_info *oinfo,
 599                     obd_size start, obd_size end,
 600                     struct ptlrpc_request_set *set)
 601 {
 602         struct ptlrpc_request *req;
 603         struct ost_body       *body;
 604         struct osc_async_args *aa;
 605         int                    rc;
 606         ENTRY;
 607
 608         if (!oinfo->oi_oa) {
 609                 CDEBUG(D_INFO, "oa NULL\n");
 610                 RETURN(-EINVAL);
 611         }
 612
 613         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
 614         if (req == NULL)
 615                 RETURN(-ENOMEM);
 616
 617         osc_set_capa_size(req, &RMF_CAPA1, oinfo->oi_capa);
 618         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
 619         if (rc) {
 620                 ptlrpc_request_free(req);
 621                 RETURN(rc);
 622         }
 623
 624         /* overload the size and blocks fields in the oa with start/end */
 625         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 626         LASSERT(body);
 627         lustre_set_wire_obdo(&body->oa, oinfo->oi_oa);
 628         body->oa.o_size = start;
 629         body->oa.o_blocks = end;
 630         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 631         osc_pack_capa(req, body, oinfo->oi_capa);
 632
 633         ptlrpc_request_set_replen(req);
 634         req->rq_interpret_reply = osc_sync_interpret;
 635
 636         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
 637         aa = ptlrpc_req_async_args(req);
 638         aa->aa_oi = oinfo;
 639
 640         ptlrpc_set_add_req(set, req);
 641         RETURN (0);
 642 }
 643
 644 /* Find and cancel locally locks matched by @mode in the resource found by
 645  * @objid. Found locks are added into @cancel list. Returns the amount of
 646  * locks added to @cancels list. */
 647 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
 648                                    cfs_list_t *cancels,
 649                                    ldlm_mode_t mode, int lock_flags)
 650 {
 651         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 652         struct ldlm_res_id res_id;
 653         struct ldlm_resource *res;
 654         int count;
 655         ENTRY;
 656
 657         osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
 658         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
 659         if (res == NULL)
 660                 RETURN(0);
 661
 662         LDLM_RESOURCE_ADDREF(res);
 663         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
 664                                            lock_flags, 0, NULL);
 665         LDLM_RESOURCE_DELREF(res);
 666         ldlm_resource_putref(res);
 667         RETURN(count);
 668 }
 669
 670 static int osc_destroy_interpret(const struct lu_env *env,
 671                                  struct ptlrpc_request *req, void *data,
 672                                  int rc)
 673 {
 674         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 675
 676         cfs_atomic_dec(&cli->cl_destroy_in_flight);
 677         cfs_waitq_signal(&cli->cl_destroy_waitq);
 678         return 0;
 679 }
 680
 681 static int osc_can_send_destroy(struct client_obd *cli)
 682 {
 683         if (cfs_atomic_inc_return(&cli->cl_destroy_in_flight) <=
 684             cli->cl_max_rpcs_in_flight) {
 685                 /* The destroy request can be sent */
 686                 return 1;
 687         }
 688         if (cfs_atomic_dec_return(&cli->cl_destroy_in_flight) <
 689             cli->cl_max_rpcs_in_flight) {
 690                 /*
 691                  * The counter has been modified between the two atomic
 692                  * operations.
 693                  */
 694                 cfs_waitq_signal(&cli->cl_destroy_waitq);
 695         }
 696         return 0;
 697 }
 698
 699 /* Destroy requests can be async always on the client, and we don't even really
 700  * care about the return code since the client cannot do anything at all about
 701  * a destroy failure.
 702  * When the MDS is unlinking a filename, it saves the file objects into a
 703  * recovery llog, and these object records are cancelled when the OST reports
 704  * they were destroyed and sync'd to disk (i.e. transaction committed).
 705  * If the client dies, or the OST is down when the object should be destroyed,
 706  * the records are not cancelled, and when the OST reconnects to the MDS next,
 707  * it will retrieve the llog unlink logs and then sends the log cancellation
 708  * cookies to the MDS after committing destroy transactions. */
 709 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
 710                        struct lov_stripe_md *ea, struct obd_trans_info *oti,
 711                        struct obd_export *md_export, void *capa)
 712 {
 713         struct client_obd     *cli = &exp->exp_obd->u.cli;
 714         struct ptlrpc_request *req;
 715         struct ost_body       *body;
 716         CFS_LIST_HEAD(cancels);
 717         int rc, count;
 718         ENTRY;
 719
 720         if (!oa) {
 721                 CDEBUG(D_INFO, "oa NULL\n");
 722                 RETURN(-EINVAL);
 723         }
 724
 725         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
 726                                         LDLM_FL_DISCARD_DATA);
 727
 728         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
 729         if (req == NULL) {
 730                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 731                 RETURN(-ENOMEM);
 732         }
 733
 734         osc_set_capa_size(req, &RMF_CAPA1, (struct obd_capa *)capa);
 735         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
 736                                0, &cancels, count);
 737         if (rc) {
 738                 ptlrpc_request_free(req);
 739                 RETURN(rc);
 740         }
 741
 742         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 743         ptlrpc_at_set_req_timeout(req);
 744
 745         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
 746                 oa->o_lcookie = *oti->oti_logcookies;
 747         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 748         LASSERT(body);
 749         lustre_set_wire_obdo(&body->oa, oa);
 750
 751         osc_pack_capa(req, body, (struct obd_capa *)capa);
 752         ptlrpc_request_set_replen(req);
 753
 754         /* don't throttle destroy RPCs for the MDT */
 755         if (!(cli->cl_import->imp_connect_flags_orig & OBD_CONNECT_MDS)) {
 756                 req->rq_interpret_reply = osc_destroy_interpret;
 757                 if (!osc_can_send_destroy(cli)) {
 758                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
 759                                                           NULL);
 760
 761                         /*
 762                          * Wait until the number of on-going destroy RPCs drops
 763                          * under max_rpc_in_flight
 764                          */
 765                         l_wait_event_exclusive(cli->cl_destroy_waitq,
 766                                                osc_can_send_destroy(cli), &lwi);
 767                 }
 768         }
 769
 770         /* Do not wait for response */
 771         ptlrpcd_add_req(req, PSCOPE_OTHER);
 772         RETURN(0);
 773 }
 774
 775 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 776                                 long writing_bytes)
 777 {
 778         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
 779
 780         LASSERT(!(oa->o_valid & bits));
 781
 782         oa->o_valid |= bits;
 783         client_obd_list_lock(&cli->cl_loi_list_lock);
 784         oa->o_dirty = cli->cl_dirty;
 785         if (cli->cl_dirty - cli->cl_dirty_transit > cli->cl_dirty_max) {
 786                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
 787                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
 788                 oa->o_undirty = 0;
 789         } else if (cfs_atomic_read(&obd_dirty_pages) -
 790                    cfs_atomic_read(&obd_dirty_transit_pages) >
 791                    obd_max_dirty_pages + 1){
 792                 /* The cfs_atomic_read() allowing the cfs_atomic_inc() are
 793                  * not covered by a lock thus they may safely race and trip
 794                  * this CERROR() unless we add in a small fudge factor (+1). */
 795                 CERROR("dirty %d - %d > system dirty_max %d\n",
 796                        cfs_atomic_read(&obd_dirty_pages),
 797                        cfs_atomic_read(&obd_dirty_transit_pages),
 798                        obd_max_dirty_pages);
 799                 oa->o_undirty = 0;
 800         } else if (cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff) {
 801                 CERROR("dirty %lu - dirty_max %lu too big???\n",
 802                        cli->cl_dirty, cli->cl_dirty_max);
 803                 oa->o_undirty = 0;
 804         } else {
 805                 long max_in_flight = (cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT)*
 806                                 (cli->cl_max_rpcs_in_flight + 1);
 807                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
 808         }
 809         oa->o_grant = cli->cl_avail_grant;
 810         oa->o_dropped = cli->cl_lost_grant;
 811         cli->cl_lost_grant = 0;
 812         client_obd_list_unlock(&cli->cl_loi_list_lock);
 813         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
 814                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 815
 816 }
 817
 818 static void osc_update_next_shrink(struct client_obd *cli)
 819 {
 820         cli->cl_next_shrink_grant =
 821                 cfs_time_shift(cli->cl_grant_shrink_interval);
 822         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
 823                cli->cl_next_shrink_grant);
 824 }
 825
 826 /* caller must hold loi_list_lock */
 827 static void osc_consume_write_grant(struct client_obd *cli,
 828                                     struct brw_page *pga)
 829 {
 830         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
 831         LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
 832         cfs_atomic_inc(&obd_dirty_pages);
 833         cli->cl_dirty += CFS_PAGE_SIZE;
 834         cli->cl_avail_grant -= CFS_PAGE_SIZE;
 835         pga->flag |= OBD_BRW_FROM_GRANT;
 836         CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
 837                CFS_PAGE_SIZE, pga, pga->pg);
 838         LASSERT(cli->cl_avail_grant >= 0);
 839         osc_update_next_shrink(cli);
 840 }
 841
 842 /* the companion to osc_consume_write_grant, called when a brw has completed.
 843  * must be called with the loi lock held. */
 844 static void osc_release_write_grant(struct client_obd *cli,
 845                                     struct brw_page *pga, int sent)
 846 {
 847         int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
 848         ENTRY;
 849
 850         LASSERT_SPIN_LOCKED(&cli->cl_loi_list_lock.lock);
 851         if (!(pga->flag & OBD_BRW_FROM_GRANT)) {
 852                 EXIT;
 853                 return;
 854         }
 855
 856         pga->flag &= ~OBD_BRW_FROM_GRANT;
 857         cfs_atomic_dec(&obd_dirty_pages);
 858         cli->cl_dirty -= CFS_PAGE_SIZE;
 859         if (pga->flag & OBD_BRW_NOCACHE) {
 860                 pga->flag &= ~OBD_BRW_NOCACHE;
 861                 cfs_atomic_dec(&obd_dirty_transit_pages);
 862                 cli->cl_dirty_transit -= CFS_PAGE_SIZE;
 863         }
 864         if (!sent) {
 865                 cli->cl_lost_grant += CFS_PAGE_SIZE;
 866                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
 867                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
 868         } else if (CFS_PAGE_SIZE != blocksize && pga->count != CFS_PAGE_SIZE) {
 869                 /* For short writes we shouldn't count parts of pages that
 870                  * span a whole block on the OST side, or our accounting goes
 871                  * wrong.  Should match the code in filter_grant_check. */
 872                 int offset = pga->off & ~CFS_PAGE_MASK;
 873                 int count = pga->count + (offset & (blocksize - 1));
 874                 int end = (offset + pga->count) & (blocksize - 1);
 875                 if (end)
 876                         count += blocksize - end;
 877
 878                 cli->cl_lost_grant += CFS_PAGE_SIZE - count;
 879                 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
 880                        CFS_PAGE_SIZE - count, cli->cl_lost_grant,
 881                        cli->cl_avail_grant, cli->cl_dirty);
 882         }
 883
 884         EXIT;
 885 }
 886
 887 static unsigned long rpcs_in_flight(struct client_obd *cli)
 888 {
 889         return cli->cl_r_in_flight + cli->cl_w_in_flight;
 890 }
 891
 892 /* caller must hold loi_list_lock */
 893 void osc_wake_cache_waiters(struct client_obd *cli)
 894 {
 895         cfs_list_t *l, *tmp;
 896         struct osc_cache_waiter *ocw;
 897
 898         ENTRY;
 899         cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 900                 /* if we can't dirty more, we must wait until some is written */
 901                 if ((cli->cl_dirty + CFS_PAGE_SIZE > cli->cl_dirty_max) ||
 902                    (cfs_atomic_read(&obd_dirty_pages) + 1 >
 903                     obd_max_dirty_pages)) {
 904                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
 905                                "osc max %ld, sys max %d\n", cli->cl_dirty,
 906                                cli->cl_dirty_max, obd_max_dirty_pages);
 907                         return;
 908                 }
 909
 910                 /* if still dirty cache but no grant wait for pending RPCs that
 911                  * may yet return us some grant before doing sync writes */
 912                 if (cli->cl_w_in_flight && cli->cl_avail_grant < CFS_PAGE_SIZE) {
 913                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
 914                                cli->cl_w_in_flight);
 915                         return;
 916                 }
 917
 918                 ocw = cfs_list_entry(l, struct osc_cache_waiter, ocw_entry);
 919                 cfs_list_del_init(&ocw->ocw_entry);
 920                 if (cli->cl_avail_grant < CFS_PAGE_SIZE) {
 921                         /* no more RPCs in flight to return grant, do sync IO */
 922                         ocw->ocw_rc = -EDQUOT;
 923                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
 924                 } else {
 925                         osc_consume_write_grant(cli,
 926                                                 &ocw->ocw_oap->oap_brw_page);
 927                 }
 928
 929                 cfs_waitq_signal(&ocw->ocw_waitq);
 930         }
 931
 932         EXIT;
 933 }
 934
 935 static void __osc_update_grant(struct client_obd *cli, obd_size grant)
 936 {
 937         client_obd_list_lock(&cli->cl_loi_list_lock);
 938         cli->cl_avail_grant += grant;
 939         client_obd_list_unlock(&cli->cl_loi_list_lock);
 940 }
 941
 942 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 943 {
 944         if (body->oa.o_valid & OBD_MD_FLGRANT) {
 945                 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
 946                 __osc_update_grant(cli, body->oa.o_grant);
 947         }
 948 }
 949
 950 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
 951                               void *key, obd_count vallen, void *val,
 952                               struct ptlrpc_request_set *set);
 953
 954 static int osc_shrink_grant_interpret(const struct lu_env *env,
 955                                       struct ptlrpc_request *req,
 956                                       void *aa, int rc)
 957 {
 958         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 959         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
 960         struct ost_body *body;
 961
 962         if (rc != 0) {
 963                 __osc_update_grant(cli, oa->o_grant);
 964                 GOTO(out, rc);
 965         }
 966
 967         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 968         LASSERT(body);
 969         osc_update_grant(cli, body);
 970 out:
 971         OBDO_FREE(oa);
 972         return rc;
 973 }
 974
 975 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
 976 {
 977         client_obd_list_lock(&cli->cl_loi_list_lock);
 978         oa->o_grant = cli->cl_avail_grant / 4;
 979         cli->cl_avail_grant -= oa->o_grant;
 980         client_obd_list_unlock(&cli->cl_loi_list_lock);
 981         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
 982                 oa->o_valid |= OBD_MD_FLFLAGS;
 983                 oa->o_flags = 0;
 984         }
 985         oa->o_flags |= OBD_FL_SHRINK_GRANT;
 986         osc_update_next_shrink(cli);
 987 }
 988
 989 /* Shrink the current grant, either from some large amount to enough for a
 990  * full set of in-flight RPCs, or if we have already shrunk to that limit
 991  * then to enough for a single RPC.  This avoids keeping more grant than
 992  * needed, and avoids shrinking the grant piecemeal. */
 993 static int osc_shrink_grant(struct client_obd *cli)
 994 {
 995         long target = (cli->cl_max_rpcs_in_flight + 1) *
 996                       cli->cl_max_pages_per_rpc;
 997
 998         client_obd_list_lock(&cli->cl_loi_list_lock);
 999         if (cli->cl_avail_grant <= target)
1000                 target = cli->cl_max_pages_per_rpc;
1001         client_obd_list_unlock(&cli->cl_loi_list_lock);
1002
1003         return osc_shrink_grant_to_target(cli, target);
1004 }
1005
1006 int osc_shrink_grant_to_target(struct client_obd *cli, long target)
1007 {
1008         int    rc = 0;
1009         struct ost_body     *body;
1010         ENTRY;
1011
1012         client_obd_list_lock(&cli->cl_loi_list_lock);
1013         /* Don't shrink if we are already above or below the desired limit
1014          * We don't want to shrink below a single RPC, as that will negatively
1015          * impact block allocation and long-term performance. */
1016         if (target < cli->cl_max_pages_per_rpc)
1017                 target = cli->cl_max_pages_per_rpc;
1018
1019         if (target >= cli->cl_avail_grant) {
1020                 client_obd_list_unlock(&cli->cl_loi_list_lock);
1021                 RETURN(0);
1022         }
1023         client_obd_list_unlock(&cli->cl_loi_list_lock);
1024
1025         OBD_ALLOC_PTR(body);
1026         if (!body)
1027                 RETURN(-ENOMEM);
1028
1029         osc_announce_cached(cli, &body->oa, 0);
1030
1031         client_obd_list_lock(&cli->cl_loi_list_lock);
1032         body->oa.o_grant = cli->cl_avail_grant - target;
1033         cli->cl_avail_grant = target;
1034         client_obd_list_unlock(&cli->cl_loi_list_lock);
1035         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
1036                 body->oa.o_valid |= OBD_MD_FLFLAGS;
1037                 body->oa.o_flags = 0;
1038         }
1039         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
1040         osc_update_next_shrink(cli);
1041
1042         rc = osc_set_info_async(cli->cl_import->imp_obd->obd_self_export,
1043                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
1044                                 sizeof(*body), body, NULL);
1045         if (rc != 0)
1046                 __osc_update_grant(cli, body->oa.o_grant);
1047         OBD_FREE_PTR(body);
1048         RETURN(rc);
1049 }
1050
1051 #define GRANT_SHRINK_LIMIT PTLRPC_MAX_BRW_SIZE
1052 static int osc_should_shrink_grant(struct client_obd *client)
1053 {
1054         cfs_time_t time = cfs_time_current();
1055         cfs_time_t next_shrink = client->cl_next_shrink_grant;
1056
1057         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
1058              OBD_CONNECT_GRANT_SHRINK) == 0)
1059                 return 0;
1060
1061         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
1062                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
1063                     client->cl_avail_grant > GRANT_SHRINK_LIMIT)
1064                         return 1;
1065                 else
1066                         osc_update_next_shrink(client);
1067         }
1068         return 0;
1069 }
1070
1071 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
1072 {
1073         struct client_obd *client;
1074
1075         cfs_list_for_each_entry(client, &item->ti_obd_list,
1076                                 cl_grant_shrink_list) {
1077                 if (osc_should_shrink_grant(client))
1078                         osc_shrink_grant(client);
1079         }
1080         return 0;
1081 }
1082
1083 static int osc_add_shrink_grant(struct client_obd *client)
1084 {
1085         int rc;
1086
1087         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1088                                        TIMEOUT_GRANT,
1089                                        osc_grant_shrink_grant_cb, NULL,
1090                                        &client->cl_grant_shrink_list);
1091         if (rc) {
1092                 CERROR("add grant client %s error %d\n",
1093                         client->cl_import->imp_obd->obd_name, rc);
1094                 return rc;
1095         }
1096         CDEBUG(D_CACHE, "add grant client %s \n",
1097                client->cl_import->imp_obd->obd_name);
1098         osc_update_next_shrink(client);
1099         return 0;
1100 }
1101
1102 static int osc_del_shrink_grant(struct client_obd *client)
1103 {
1104         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1105                                          TIMEOUT_GRANT);
1106 }
1107
1108 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1109 {
1110         /*
1111          * ocd_grant is the total grant amount we're expect to hold: if we've
1112          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1113          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1114          *
1115          * race is tolerable here: if we're evicted, but imp_state already
1116          * left EVICTED state, then cl_dirty must be 0 already.
1117          */
1118         client_obd_list_lock(&cli->cl_loi_list_lock);
1119         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1120                 cli->cl_avail_grant = ocd->ocd_grant;
1121         else
1122                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1123
1124         if (cli->cl_avail_grant < 0) {
1125                 CWARN("%s: available grant < 0, the OSS is probably not running"
1126                       " with patch from bug20278 (%ld) \n",
1127                       cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant);
1128                 /* workaround for 1.6 servers which do not have
1129                  * the patch from bug20278 */
1130                 cli->cl_avail_grant = ocd->ocd_grant;
1131         }
1132
1133         client_obd_list_unlock(&cli->cl_loi_list_lock);
1134
1135         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld \n",
1136                cli->cl_import->imp_obd->obd_name,
1137                cli->cl_avail_grant, cli->cl_lost_grant);
1138
1139         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1140             cfs_list_empty(&cli->cl_grant_shrink_list))
1141                 osc_add_shrink_grant(cli);
1142 }
1143
1144 /* We assume that the reason this OSC got a short read is because it read
1145  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1146  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1147  * this stripe never got written at or beyond this stripe offset yet. */
1148 static void handle_short_read(int nob_read, obd_count page_count,
1149                               struct brw_page **pga)
1150 {
1151         char *ptr;
1152         int i = 0;
1153
1154         /* skip bytes read OK */
1155         while (nob_read > 0) {
1156                 LASSERT (page_count > 0);
1157
1158                 if (pga[i]->count > nob_read) {
1159                         /* EOF inside this page */
1160                         ptr = cfs_kmap(pga[i]->pg) +
1161                                 (pga[i]->off & ~CFS_PAGE_MASK);
1162                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1163                         cfs_kunmap(pga[i]->pg);
1164                         page_count--;
1165                         i++;
1166                         break;
1167                 }
1168
1169                 nob_read -= pga[i]->count;
1170                 page_count--;
1171                 i++;
1172         }
1173
1174         /* zero remaining pages */
1175         while (page_count-- > 0) {
1176                 ptr = cfs_kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1177                 memset(ptr, 0, pga[i]->count);
1178                 cfs_kunmap(pga[i]->pg);
1179                 i++;
1180         }
1181 }
1182
1183 static int check_write_rcs(struct ptlrpc_request *req,
1184                            int requested_nob, int niocount,
1185                            obd_count page_count, struct brw_page **pga)
1186 {
1187         int     i;
1188         __u32   *remote_rcs;
1189
1190         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1191                                                   sizeof(*remote_rcs) *
1192                                                   niocount);
1193         if (remote_rcs == NULL) {
1194                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1195                 return(-EPROTO);
1196         }
1197
1198         /* return error if any niobuf was in error */
1199         for (i = 0; i < niocount; i++) {
1200                 if ((int)remote_rcs[i] < 0)
1201                         return(remote_rcs[i]);
1202
1203                 if (remote_rcs[i] != 0) {
1204                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1205                                 i, remote_rcs[i], req);
1206                         return(-EPROTO);
1207                 }
1208         }
1209
1210         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1211                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1212                        req->rq_bulk->bd_nob_transferred, requested_nob);
1213                 return(-EPROTO);
1214         }
1215
1216         return (0);
1217 }
1218
1219 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1220 {
1221         if (p1->flag != p2->flag) {
1222                 unsigned mask = ~(OBD_BRW_FROM_GRANT|
1223                                   OBD_BRW_NOCACHE|OBD_BRW_SYNC|OBD_BRW_ASYNC);
1224
1225                 /* warn if we try to combine flags that we don't know to be
1226                  * safe to combine */
1227                 if ((p1->flag & mask) != (p2->flag & mask))
1228                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
1229                                "same brw?\n", p1->flag, p2->flag);
1230                 return 0;
1231         }
1232
1233         return (p1->off + p1->count == p2->off);
1234 }
1235
1236 static obd_count osc_checksum_bulk(int nob, obd_count pg_count,
1237                                    struct brw_page **pga, int opc,
1238                                    cksum_type_t cksum_type)
1239 {
1240         __u32 cksum;
1241         int i = 0;
1242
1243         LASSERT (pg_count > 0);
1244         cksum = init_checksum(cksum_type);
1245         while (nob > 0 && pg_count > 0) {
1246                 unsigned char *ptr = cfs_kmap(pga[i]->pg);
1247                 int off = pga[i]->off & ~CFS_PAGE_MASK;
1248                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1249
1250                 /* corrupt the data before we compute the checksum, to
1251                  * simulate an OST->client data error */
1252                 if (i == 0 && opc == OST_READ &&
1253                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))
1254                         memcpy(ptr + off, "bad1", min(4, nob));
1255                 cksum = compute_checksum(cksum, ptr + off, count, cksum_type);
1256                 cfs_kunmap(pga[i]->pg);
1257                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d checksum %x\n",
1258                                off, cksum);
1259
1260                 nob -= pga[i]->count;
1261                 pg_count--;
1262                 i++;
1263         }
1264         /* For sending we only compute the wrong checksum instead
1265          * of corrupting the data so it is still correct on a redo */
1266         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1267                 cksum++;
1268
1269         return cksum;
1270 }
1271
1272 static int osc_brw_prep_request(int cmd, struct client_obd *cli,struct obdo *oa,
1273                                 struct lov_stripe_md *lsm, obd_count page_count,
1274                                 struct brw_page **pga,
1275                                 struct ptlrpc_request **reqp,
1276                                 struct obd_capa *ocapa, int reserve,
1277                                 int resend)
1278 {
1279         struct ptlrpc_request   *req;
1280         struct ptlrpc_bulk_desc *desc;
1281         struct ost_body         *body;
1282         struct obd_ioobj        *ioobj;
1283         struct niobuf_remote    *niobuf;
1284         int niocount, i, requested_nob, opc, rc;
1285         struct osc_brw_async_args *aa;
1286         struct req_capsule      *pill;
1287         struct brw_page *pg_prev;
1288
1289         ENTRY;
1290         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1291                 RETURN(-ENOMEM); /* Recoverable */
1292         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1293                 RETURN(-EINVAL); /* Fatal */
1294
1295         if ((cmd & OBD_BRW_WRITE) != 0) {
1296                 opc = OST_WRITE;
1297                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1298                                                 cli->cl_import->imp_rq_pool,
1299                                                 &RQF_OST_BRW_WRITE);
1300         } else {
1301                 opc = OST_READ;
1302                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1303         }
1304         if (req == NULL)
1305                 RETURN(-ENOMEM);
1306
1307         for (niocount = i = 1; i < page_count; i++) {
1308                 if (!can_merge_pages(pga[i - 1], pga[i]))
1309                         niocount++;
1310         }
1311
1312         pill = &req->rq_pill;
1313         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1314                              sizeof(*ioobj));
1315         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1316                              niocount * sizeof(*niobuf));
1317         osc_set_capa_size(req, &RMF_CAPA1, ocapa);
1318
1319         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1320         if (rc) {
1321                 ptlrpc_request_free(req);
1322                 RETURN(rc);
1323         }
1324         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1325         ptlrpc_at_set_req_timeout(req);
1326
1327         if (opc == OST_WRITE)
1328                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1329                                             BULK_GET_SOURCE, OST_BULK_PORTAL);
1330         else
1331                 desc = ptlrpc_prep_bulk_imp(req, page_count,
1332                                             BULK_PUT_SINK, OST_BULK_PORTAL);
1333
1334         if (desc == NULL)
1335                 GOTO(out, rc = -ENOMEM);
1336         /* NB request now owns desc and will free it when it gets freed */
1337
1338         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1339         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1340         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1341         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1342
1343         lustre_set_wire_obdo(&body->oa, oa);
1344
1345         obdo_to_ioobj(oa, ioobj);
1346         ioobj->ioo_bufcnt = niocount;
1347         osc_pack_capa(req, body, ocapa);
1348         LASSERT (page_count > 0);
1349         pg_prev = pga[0];
1350         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1351                 struct brw_page *pg = pga[i];
1352                 int poff = pg->off & ~CFS_PAGE_MASK;
1353
1354                 LASSERT(pg->count > 0);
1355                 /* make sure there is no gap in the middle of page array */
1356                 LASSERTF(page_count == 1 ||
1357                          (ergo(i == 0, poff + pg->count == CFS_PAGE_SIZE) &&
1358                           ergo(i > 0 && i < page_count - 1,
1359                                poff == 0 && pg->count == CFS_PAGE_SIZE)   &&
1360                           ergo(i == page_count - 1, poff == 0)),
1361                          "i: %d/%d pg: %p off: "LPU64", count: %u\n",
1362                          i, page_count, pg, pg->off, pg->count);
1363 #ifdef __linux__
1364                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1365                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
1366                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
1367                          i, page_count,
1368                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1369                          pg_prev->pg, page_private(pg_prev->pg),
1370                          pg_prev->pg->index, pg_prev->off);
1371 #else
1372                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1373                          "i %d p_c %u\n", i, page_count);
1374 #endif
1375                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1376                         (pg->flag & OBD_BRW_SRVLOCK));
1377
1378                 ptlrpc_prep_bulk_page(desc, pg->pg, poff, pg->count);
1379                 requested_nob += pg->count;
1380
1381                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1382                         niobuf--;
1383                         niobuf->len += pg->count;
1384                 } else {
1385                         niobuf->offset = pg->off;
1386                         niobuf->len    = pg->count;
1387                         niobuf->flags  = pg->flag;
1388                 }
1389                 pg_prev = pg;
1390         }
1391
1392         LASSERTF((void *)(niobuf - niocount) ==
1393                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1394                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1395                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1396
1397         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1398         if (resend) {
1399                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1400                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1401                         body->oa.o_flags = 0;
1402                 }
1403                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1404         }
1405
1406         if (osc_should_shrink_grant(cli))
1407                 osc_shrink_grant_local(cli, &body->oa);
1408
1409         /* size[REQ_REC_OFF] still sizeof (*body) */
1410         if (opc == OST_WRITE) {
1411                 if (unlikely(cli->cl_checksum) &&
1412                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1413                         /* store cl_cksum_type in a local variable since
1414                          * it can be changed via lprocfs */
1415                         cksum_type_t cksum_type = cli->cl_cksum_type;
1416
1417                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1418                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1419                                 body->oa.o_flags = 0;
1420                         }
1421                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1422                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1423                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1424                                                              page_count, pga,
1425                                                              OST_WRITE,
1426                                                              cksum_type);
1427                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1428                                body->oa.o_cksum);
1429                         /* save this in 'oa', too, for later checking */
1430                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1431                         oa->o_flags |= cksum_type_pack(cksum_type);
1432                 } else {
1433                         /* clear out the checksum flag, in case this is a
1434                          * resend but cl_checksum is no longer set. b=11238 */
1435                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1436                 }
1437                 oa->o_cksum = body->oa.o_cksum;
1438                 /* 1 RC per niobuf */
1439                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1440                                      sizeof(__u32) * niocount);
1441         } else {
1442                 if (unlikely(cli->cl_checksum) &&
1443                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1444                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1445                                 body->oa.o_flags = 0;
1446                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1447                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1448                 }
1449         }
1450         ptlrpc_request_set_replen(req);
1451
1452         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1453         aa = ptlrpc_req_async_args(req);
1454         aa->aa_oa = oa;
1455         aa->aa_requested_nob = requested_nob;
1456         aa->aa_nio_count = niocount;
1457         aa->aa_page_count = page_count;
1458         aa->aa_resends = 0;
1459         aa->aa_ppga = pga;
1460         aa->aa_cli = cli;
1461         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1462         if (ocapa && reserve)
1463                 aa->aa_ocapa = capa_get(ocapa);
1464
1465         *reqp = req;
1466         RETURN(0);
1467
1468  out:
1469         ptlrpc_req_finished(req);
1470         RETURN(rc);
1471 }
1472
1473 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1474                                 __u32 client_cksum, __u32 server_cksum, int nob,
1475                                 obd_count page_count, struct brw_page **pga,
1476                                 cksum_type_t client_cksum_type)
1477 {
1478         __u32 new_cksum;
1479         char *msg;
1480         cksum_type_t cksum_type;
1481
1482         if (server_cksum == client_cksum) {
1483                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1484                 return 0;
1485         }
1486
1487         /* If this is mmaped file - it can be changed at any time */
1488         if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_MMAP)
1489                 return 1;
1490
1491         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1492                                        oa->o_flags : 0);
1493         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1494                                       cksum_type);
1495
1496         if (cksum_type != client_cksum_type)
1497                 msg = "the server did not use the checksum type specified in "
1498                       "the original request - likely a protocol problem";
1499         else if (new_cksum == server_cksum)
1500                 msg = "changed on the client after we checksummed it - "
1501                       "likely false positive due to mmap IO (bug 11742)";
1502         else if (new_cksum == client_cksum)
1503                 msg = "changed in transit before arrival at OST";
1504         else
1505                 msg = "changed in transit AND doesn't match the original - "
1506                       "likely false positive due to mmap IO (bug 11742)";
1507
1508         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1509                            " object "LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1510                            msg, libcfs_nid2str(peer->nid),
1511                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1512                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1513                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1514                            oa->o_id,
1515                            oa->o_valid & OBD_MD_FLGROUP ? oa->o_seq : (__u64)0,
1516                            pga[0]->off,
1517                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1518         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1519                "client csum now %x\n", client_cksum, client_cksum_type,
1520                server_cksum, cksum_type, new_cksum);
1521         return 1;
1522 }
1523
1524 /* Note rc enters this function as number of bytes transferred */
1525 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1526 {
1527         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1528         const lnet_process_id_t *peer =
1529                         &req->rq_import->imp_connection->c_peer;
1530         struct client_obd *cli = aa->aa_cli;
1531         struct ost_body *body;
1532         __u32 client_cksum = 0;
1533         ENTRY;
1534
1535         if (rc < 0 && rc != -EDQUOT) {
1536                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1537                 RETURN(rc);
1538         }
1539
1540         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1541         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1542         if (body == NULL) {
1543                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1544                 RETURN(-EPROTO);
1545         }
1546
1547 #ifdef HAVE_QUOTA_SUPPORT
1548         /* set/clear over quota flag for a uid/gid */
1549         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1550             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1551                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1552
1553                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid "LPX64", flags %x\n",
1554                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1555                        body->oa.o_flags);
1556                 lquota_setdq(quota_interface, cli, qid, body->oa.o_valid,
1557                              body->oa.o_flags);
1558         }
1559 #endif
1560
1561         osc_update_grant(cli, body);
1562
1563         if (rc < 0)
1564                 RETURN(rc);
1565
1566         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1567                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1568
1569         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1570                 if (rc > 0) {
1571                         CERROR("Unexpected +ve rc %d\n", rc);
1572                         RETURN(-EPROTO);
1573                 }
1574                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1575
1576                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1577                         RETURN(-EAGAIN);
1578
1579                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1580                     check_write_checksum(&body->oa, peer, client_cksum,
1581                                          body->oa.o_cksum, aa->aa_requested_nob,
1582                                          aa->aa_page_count, aa->aa_ppga,
1583                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1584                         RETURN(-EAGAIN);
1585
1586                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1587                                      aa->aa_page_count, aa->aa_ppga);
1588                 GOTO(out, rc);
1589         }
1590
1591         /* The rest of this function executes only for OST_READs */
1592
1593         /* if unwrap_bulk failed, return -EAGAIN to retry */
1594         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1595         if (rc < 0)
1596                 GOTO(out, rc = -EAGAIN);
1597
1598         if (rc > aa->aa_requested_nob) {
1599                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1600                        aa->aa_requested_nob);
1601                 RETURN(-EPROTO);
1602         }
1603
1604         if (rc != req->rq_bulk->bd_nob_transferred) {
1605                 CERROR ("Unexpected rc %d (%d transferred)\n",
1606                         rc, req->rq_bulk->bd_nob_transferred);
1607                 return (-EPROTO);
1608         }
1609
1610         if (rc < aa->aa_requested_nob)
1611                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1612
1613         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1614                 static int cksum_counter;
1615                 __u32      server_cksum = body->oa.o_cksum;
1616                 char      *via;
1617                 char      *router;
1618                 cksum_type_t cksum_type;
1619
1620                 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1621                                                body->oa.o_flags : 0);
1622                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1623                                                  aa->aa_ppga, OST_READ,
1624                                                  cksum_type);
1625
1626                 if (peer->nid == req->rq_bulk->bd_sender) {
1627                         via = router = "";
1628                 } else {
1629                         via = " via ";
1630                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1631                 }
1632
1633                 if (server_cksum == ~0 && rc > 0) {
1634                         CERROR("Protocol error: server %s set the 'checksum' "
1635                                "bit, but didn't send a checksum.  Not fatal, "
1636                                "but please notify on http://bugs.whamcloud.com/\n",
1637                                libcfs_nid2str(peer->nid));
1638                 } else if (server_cksum != client_cksum) {
1639                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1640                                            "%s%s%s inode "DFID" object "
1641                                            LPU64"/"LPU64" extent "
1642                                            "["LPU64"-"LPU64"]\n",
1643                                            req->rq_import->imp_obd->obd_name,
1644                                            libcfs_nid2str(peer->nid),
1645                                            via, router,
1646                                            body->oa.o_valid & OBD_MD_FLFID ?
1647                                                 body->oa.o_parent_seq : (__u64)0,
1648                                            body->oa.o_valid & OBD_MD_FLFID ?
1649                                                 body->oa.o_parent_oid : 0,
1650                                            body->oa.o_valid & OBD_MD_FLFID ?
1651                                                 body->oa.o_parent_ver : 0,
1652                                            body->oa.o_id,
1653                                            body->oa.o_valid & OBD_MD_FLGROUP ?
1654                                                 body->oa.o_seq : (__u64)0,
1655                                            aa->aa_ppga[0]->off,
1656                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1657                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1658                                                                         1);
1659                         CERROR("client %x, server %x, cksum_type %x\n",
1660                                client_cksum, server_cksum, cksum_type);
1661                         cksum_counter = 0;
1662                         aa->aa_oa->o_cksum = client_cksum;
1663                         rc = -EAGAIN;
1664                 } else {
1665                         cksum_counter++;
1666                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1667                         rc = 0;
1668                 }
1669         } else if (unlikely(client_cksum)) {
1670                 static int cksum_missed;
1671
1672                 cksum_missed++;
1673                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1674                         CERROR("Checksum %u requested from %s but not sent\n",
1675                                cksum_missed, libcfs_nid2str(peer->nid));
1676         } else {
1677                 rc = 0;
1678         }
1679 out:
1680         if (rc >= 0)
1681                 lustre_get_wire_obdo(aa->aa_oa, &body->oa);
1682
1683         RETURN(rc);
1684 }
1685
1686 static int osc_brw_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1687                             struct lov_stripe_md *lsm,
1688                             obd_count page_count, struct brw_page **pga,
1689                             struct obd_capa *ocapa)
1690 {
1691         struct ptlrpc_request *req;
1692         int                    rc;
1693         cfs_waitq_t            waitq;
1694         int                    resends = 0;
1695         struct l_wait_info     lwi;
1696
1697         ENTRY;
1698
1699         cfs_waitq_init(&waitq);
1700
1701 restart_bulk:
1702         rc = osc_brw_prep_request(cmd, &exp->exp_obd->u.cli, oa, lsm,
1703                                   page_count, pga, &req, ocapa, 0, resends);
1704         if (rc != 0)
1705                 return (rc);
1706
1707         rc = ptlrpc_queue_wait(req);
1708
1709         if (rc == -ETIMEDOUT && req->rq_resend) {
1710                 DEBUG_REQ(D_HA, req,  "BULK TIMEOUT");
1711                 ptlrpc_req_finished(req);
1712                 goto restart_bulk;
1713         }
1714
1715         rc = osc_brw_fini_request(req, rc);
1716
1717         ptlrpc_req_finished(req);
1718         if (osc_recoverable_error(rc)) {
1719                 resends++;
1720                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1721                         CERROR("too many resend retries, returning error\n");
1722                         RETURN(-EIO);
1723                 }
1724
1725                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL);
1726                 l_wait_event(waitq, 0, &lwi);
1727
1728                 goto restart_bulk;
1729         }
1730
1731         RETURN (rc);
1732 }
1733
1734 int osc_brw_redo_request(struct ptlrpc_request *request,
1735                          struct osc_brw_async_args *aa)
1736 {
1737         struct ptlrpc_request *new_req;
1738         struct ptlrpc_request_set *set = request->rq_set;
1739         struct osc_brw_async_args *new_aa;
1740         struct osc_async_page *oap;
1741         int rc = 0;
1742         ENTRY;
1743
1744         if (!client_should_resend(aa->aa_resends, aa->aa_cli)) {
1745                 CERROR("too many resent retries, returning error\n");
1746                 RETURN(-EIO);
1747         }
1748
1749         DEBUG_REQ(D_ERROR, request, "redo for recoverable error");
1750
1751         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1752                                         OST_WRITE ? OBD_BRW_WRITE :OBD_BRW_READ,
1753                                   aa->aa_cli, aa->aa_oa,
1754                                   NULL /* lsm unused by osc currently */,
1755                                   aa->aa_page_count, aa->aa_ppga,
1756                                   &new_req, aa->aa_ocapa, 0, 1);
1757         if (rc)
1758                 RETURN(rc);
1759
1760         client_obd_list_lock(&aa->aa_cli->cl_loi_list_lock);
1761
1762         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1763                 if (oap->oap_request != NULL) {
1764                         LASSERTF(request == oap->oap_request,
1765                                  "request %p != oap_request %p\n",
1766                                  request, oap->oap_request);
1767                         if (oap->oap_interrupted) {
1768                                 client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1769                                 ptlrpc_req_finished(new_req);
1770                                 RETURN(-EINTR);
1771                         }
1772                 }
1773         }
1774         /* New request takes over pga and oaps from old request.
1775          * Note that copying a list_head doesn't work, need to move it... */
1776         aa->aa_resends++;
1777         new_req->rq_interpret_reply = request->rq_interpret_reply;
1778         new_req->rq_async_args = request->rq_async_args;
1779         new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1780
1781         new_aa = ptlrpc_req_async_args(new_req);
1782
1783         CFS_INIT_LIST_HEAD(&new_aa->aa_oaps);
1784         cfs_list_splice(&aa->aa_oaps, &new_aa->aa_oaps);
1785         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
1786
1787         cfs_list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1788                 if (oap->oap_request) {
1789                         ptlrpc_req_finished(oap->oap_request);
1790                         oap->oap_request = ptlrpc_request_addref(new_req);
1791                 }
1792         }
1793
1794         new_aa->aa_ocapa = aa->aa_ocapa;
1795         aa->aa_ocapa = NULL;
1796
1797         /* use ptlrpc_set_add_req is safe because interpret functions work
1798          * in check_set context. only one way exist with access to request
1799          * from different thread got -EINTR - this way protected with
1800          * cl_loi_list_lock */
1801         ptlrpc_set_add_req(set, new_req);
1802
1803         client_obd_list_unlock(&aa->aa_cli->cl_loi_list_lock);
1804
1805         DEBUG_REQ(D_INFO, new_req, "new request");
1806         RETURN(0);
1807 }
1808
1809 /*
1810  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1811  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1812  * fine for our small page arrays and doesn't require allocation.  its an
1813  * insertion sort that swaps elements that are strides apart, shrinking the
1814  * stride down until its '1' and the array is sorted.
1815  */
1816 static void sort_brw_pages(struct brw_page **array, int num)
1817 {
1818         int stride, i, j;
1819         struct brw_page *tmp;
1820
1821         if (num == 1)
1822                 return;
1823         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1824                 ;
1825
1826         do {
1827                 stride /= 3;
1828                 for (i = stride ; i < num ; i++) {
1829                         tmp = array[i];
1830                         j = i;
1831                         while (j >= stride && array[j - stride]->off > tmp->off) {
1832                                 array[j] = array[j - stride];
1833                                 j -= stride;
1834                         }
1835                         array[j] = tmp;
1836                 }
1837         } while (stride > 1);
1838 }
1839
1840 static obd_count max_unfragmented_pages(struct brw_page **pg, obd_count pages)
1841 {
1842         int count = 1;
1843         int offset;
1844         int i = 0;
1845
1846         LASSERT (pages > 0);
1847         offset = pg[i]->off & ~CFS_PAGE_MASK;
1848
1849         for (;;) {
1850                 pages--;
1851                 if (pages == 0)         /* that's all */
1852                         return count;
1853
1854                 if (offset + pg[i]->count < CFS_PAGE_SIZE)
1855                         return count;   /* doesn't end on page boundary */
1856
1857                 i++;
1858                 offset = pg[i]->off & ~CFS_PAGE_MASK;
1859                 if (offset != 0)        /* doesn't start on page boundary */
1860                         return count;
1861
1862                 count++;
1863         }
1864 }
1865
1866 static struct brw_page **osc_build_ppga(struct brw_page *pga, obd_count count)
1867 {
1868         struct brw_page **ppga;
1869         int i;
1870
1871         OBD_ALLOC(ppga, sizeof(*ppga) * count);
1872         if (ppga == NULL)
1873                 return NULL;
1874
1875         for (i = 0; i < count; i++)
1876                 ppga[i] = pga + i;
1877         return ppga;
1878 }
1879
1880 static void osc_release_ppga(struct brw_page **ppga, obd_count count)
1881 {
1882         LASSERT(ppga != NULL);
1883         OBD_FREE(ppga, sizeof(*ppga) * count);
1884 }
1885
1886 static int osc_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1887                    obd_count page_count, struct brw_page *pga,
1888                    struct obd_trans_info *oti)
1889 {
1890         struct obdo *saved_oa = NULL;
1891         struct brw_page **ppga, **orig;
1892         struct obd_import *imp = class_exp2cliimp(exp);
1893         struct client_obd *cli;
1894         int rc, page_count_orig;
1895         ENTRY;
1896
1897         LASSERT((imp != NULL) && (imp->imp_obd != NULL));
1898         cli = &imp->imp_obd->u.cli;
1899
1900         if (cmd & OBD_BRW_CHECK) {
1901                 /* The caller just wants to know if there's a chance that this
1902                  * I/O can succeed */
1903
1904                 if (imp->imp_invalid)
1905                         RETURN(-EIO);
1906                 RETURN(0);
1907         }
1908
1909         /* test_brw with a failed create can trip this, maybe others. */
1910         LASSERT(cli->cl_max_pages_per_rpc);
1911
1912         rc = 0;
1913
1914         orig = ppga = osc_build_ppga(pga, page_count);
1915         if (ppga == NULL)
1916                 RETURN(-ENOMEM);
1917         page_count_orig = page_count;
1918
1919         sort_brw_pages(ppga, page_count);
1920         while (page_count) {
1921                 obd_count pages_per_brw;
1922
1923                 if (page_count > cli->cl_max_pages_per_rpc)
1924                         pages_per_brw = cli->cl_max_pages_per_rpc;
1925                 else
1926                         pages_per_brw = page_count;
1927
1928                 pages_per_brw = max_unfragmented_pages(ppga, pages_per_brw);
1929
1930                 if (saved_oa != NULL) {
1931                         /* restore previously saved oa */
1932                         *oinfo->oi_oa = *saved_oa;
1933                 } else if (page_count > pages_per_brw) {
1934                         /* save a copy of oa (brw will clobber it) */
1935                         OBDO_ALLOC(saved_oa);
1936                         if (saved_oa == NULL)
1937                                 GOTO(out, rc = -ENOMEM);
1938                         *saved_oa = *oinfo->oi_oa;
1939                 }
1940
1941                 rc = osc_brw_internal(cmd, exp, oinfo->oi_oa, oinfo->oi_md,
1942                                       pages_per_brw, ppga, oinfo->oi_capa);
1943
1944                 if (rc != 0)
1945                         break;
1946
1947                 page_count -= pages_per_brw;
1948                 ppga += pages_per_brw;
1949         }
1950
1951 out:
1952         osc_release_ppga(orig, page_count_orig);
1953
1954         if (saved_oa != NULL)
1955                 OBDO_FREE(saved_oa);
1956
1957         RETURN(rc);
1958 }
1959
1960 /* The companion to osc_enter_cache(), called when @oap is no longer part of
1961  * the dirty accounting.  Writeback completes or truncate happens before
1962  * writing starts.  Must be called with the loi lock held. */
1963 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1964                            int sent)
1965 {
1966         osc_release_write_grant(cli, &oap->oap_brw_page, sent);
1967 }
1968
1969
1970 /* This maintains the lists of pending pages to read/write for a given object
1971  * (lop).  This is used by osc_check_rpcs->osc_next_loi() and loi_list_maint()
1972  * to quickly find objects that are ready to send an RPC. */
1973 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1974                          int cmd)
1975 {
1976         int optimal;
1977         ENTRY;
1978
1979         if (lop->lop_num_pending == 0)
1980                 RETURN(0);
1981
1982         /* if we have an invalid import we want to drain the queued pages
1983          * by forcing them through rpcs that immediately fail and complete
1984          * the pages.  recovery relies on this to empty the queued pages
1985          * before canceling the locks and evicting down the llite pages */
1986         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1987                 RETURN(1);
1988
1989         /* stream rpcs in queue order as long as as there is an urgent page
1990          * queued.  this is our cheap solution for good batching in the case
1991          * where writepage marks some random page in the middle of the file
1992          * as urgent because of, say, memory pressure */
1993         if (!cfs_list_empty(&lop->lop_urgent)) {
1994                 CDEBUG(D_CACHE, "urgent request forcing RPC\n");
1995                 RETURN(1);
1996         }
1997         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1998         optimal = cli->cl_max_pages_per_rpc;
1999         if (cmd & OBD_BRW_WRITE) {
2000                 /* trigger a write rpc stream as long as there are dirtiers
2001                  * waiting for space.  as they're waiting, they're not going to
2002                  * create more pages to coalesce with what's waiting.. */
2003                 if (!cfs_list_empty(&cli->cl_cache_waiters)) {
2004                         CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
2005                         RETURN(1);
2006                 }
2007                 /* +16 to avoid triggering rpcs that would want to include pages
2008                  * that are being queued but which can't be made ready until
2009                  * the queuer finishes with the page. this is a wart for
2010                  * llite::commit_write() */
2011                 optimal += 16;
2012         }
2013         if (lop->lop_num_pending >= optimal)
2014                 RETURN(1);
2015
2016         RETURN(0);
2017 }
2018
2019 static int lop_makes_hprpc(struct loi_oap_pages *lop)
2020 {
2021         struct osc_async_page *oap;
2022         ENTRY;
2023
2024         if (cfs_list_empty(&lop->lop_urgent))
2025                 RETURN(0);
2026
2027         oap = cfs_list_entry(lop->lop_urgent.next,
2028                          struct osc_async_page, oap_urgent_item);
2029
2030         if (oap->oap_async_flags & ASYNC_HP) {
2031                 CDEBUG(D_CACHE, "hp request forcing RPC\n");
2032                 RETURN(1);
2033         }
2034
2035         RETURN(0);
2036 }
2037
2038 static void on_list(cfs_list_t *item, cfs_list_t *list,
2039                     int should_be_on)
2040 {
2041         if (cfs_list_empty(item) && should_be_on)
2042                 cfs_list_add_tail(item, list);
2043         else if (!cfs_list_empty(item) && !should_be_on)
2044                 cfs_list_del_init(item);
2045 }
2046
2047 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
2048  * can find pages to build into rpcs quickly */
2049 void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
2050 {
2051         if (lop_makes_hprpc(&loi->loi_write_lop) ||
2052             lop_makes_hprpc(&loi->loi_read_lop)) {
2053                 /* HP rpc */
2054                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list, 0);
2055                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 1);
2056         } else {
2057                 on_list(&loi->loi_hp_ready_item, &cli->cl_loi_hp_ready_list, 0);
2058                 on_list(&loi->loi_ready_item, &cli->cl_loi_ready_list,
2059                         lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)||
2060                         lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
2061         }
2062
2063         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
2064                 loi->loi_write_lop.lop_num_pending);
2065
2066         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
2067                 loi->loi_read_lop.lop_num_pending);
2068 }
2069
2070 static void lop_update_pending(struct client_obd *cli,
2071                                struct loi_oap_pages *lop, int cmd, int delta)
2072 {
2073         lop->lop_num_pending += delta;
2074         if (cmd & OBD_BRW_WRITE)
2075                 cli->cl_pending_w_pages += delta;
2076         else
2077                 cli->cl_pending_r_pages += delta;
2078 }
2079
2080 /**
2081  * this is called when a sync waiter receives an interruption.  Its job is to
2082  * get the caller woken as soon as possible.  If its page hasn't been put in an
2083  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
2084  * desiring interruption which will forcefully complete the rpc once the rpc
2085  * has timed out.
2086  */
2087 int osc_oap_interrupted(const struct lu_env *env, struct osc_async_page *oap)
2088 {
2089         struct loi_oap_pages *lop;
2090         struct lov_oinfo *loi;
2091         int rc = -EBUSY;
2092         ENTRY;
2093
2094         LASSERT(!oap->oap_interrupted);
2095         oap->oap_interrupted = 1;
2096
2097         /* ok, it's been put in an rpc. only one oap gets a request reference */
2098         if (oap->oap_request != NULL) {
2099                 ptlrpc_mark_interrupted(oap->oap_request);
2100                 ptlrpcd_wake(oap->oap_request);
2101                 ptlrpc_req_finished(oap->oap_request);
2102                 oap->oap_request = NULL;
2103         }
2104
2105         /*
2106          * page completion may be called only if ->cpo_prep() method was
2107          * executed by osc_io_submit(), that also adds page the to pending list
2108          */
2109         if (!cfs_list_empty(&oap->oap_pending_item)) {
2110                 cfs_list_del_init(&oap->oap_pending_item);
2111                 cfs_list_del_init(&oap->oap_urgent_item);
2112
2113                 loi = oap->oap_loi;
2114                 lop = (oap->oap_cmd & OBD_BRW_WRITE) ?
2115                         &loi->loi_write_lop : &loi->loi_read_lop;
2116                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
2117                 loi_list_maint(oap->oap_cli, oap->oap_loi);
2118                 rc = oap->oap_caller_ops->ap_completion(env,
2119                                           oap->oap_caller_data,
2120                                           oap->oap_cmd, NULL, -EINTR);
2121         }
2122
2123         RETURN(rc);
2124 }
2125
2126 /* this is trying to propogate async writeback errors back up to the
2127  * application.  As an async write fails we record the error code for later if
2128  * the app does an fsync.  As long as errors persist we force future rpcs to be
2129  * sync so that the app can get a sync error and break the cycle of queueing
2130  * pages for which writeback will fail. */
2131 static void osc_process_ar(struct osc_async_rc *ar, __u64 xid,
2132                            int rc)
2133 {
2134         if (rc) {
2135                 if (!ar->ar_rc)
2136                         ar->ar_rc = rc;
2137
2138                 ar->ar_force_sync = 1;
2139                 ar->ar_min_xid = ptlrpc_sample_next_xid();
2140                 return;
2141
2142         }
2143
2144         if (ar->ar_force_sync && (xid >= ar->ar_min_xid))
2145                 ar->ar_force_sync = 0;
2146 }
2147
2148 void osc_oap_to_pending(struct osc_async_page *oap)
2149 {
2150         struct loi_oap_pages *lop;
2151
2152         if (oap->oap_cmd & OBD_BRW_WRITE)
2153                 lop = &oap->oap_loi->loi_write_lop;
2154         else
2155                 lop = &oap->oap_loi->loi_read_lop;
2156
2157         if (oap->oap_async_flags & ASYNC_HP)
2158                 cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2159         else if (oap->oap_async_flags & ASYNC_URGENT)
2160                 cfs_list_add_tail(&oap->oap_urgent_item, &lop->lop_urgent);
2161         cfs_list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2162         lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, 1);
2163 }
2164
2165 /* this must be called holding the loi list lock to give coverage to exit_cache,
2166  * async_flag maintenance, and oap_request */
2167 static void osc_ap_completion(const struct lu_env *env,
2168                               struct client_obd *cli, struct obdo *oa,
2169                               struct osc_async_page *oap, int sent, int rc)
2170 {
2171         __u64 xid = 0;
2172
2173         ENTRY;
2174         if (oap->oap_request != NULL) {
2175                 xid = ptlrpc_req_xid(oap->oap_request);
2176                 ptlrpc_req_finished(oap->oap_request);
2177                 oap->oap_request = NULL;
2178         }
2179
2180         cfs_spin_lock(&oap->oap_lock);
2181         oap->oap_async_flags = 0;
2182         cfs_spin_unlock(&oap->oap_lock);
2183         oap->oap_interrupted = 0;
2184
2185         if (oap->oap_cmd & OBD_BRW_WRITE) {
2186                 osc_process_ar(&cli->cl_ar, xid, rc);
2187                 osc_process_ar(&oap->oap_loi->loi_ar, xid, rc);
2188         }
2189
2190         if (rc == 0 && oa != NULL) {
2191                 if (oa->o_valid & OBD_MD_FLBLOCKS)
2192                         oap->oap_loi->loi_lvb.lvb_blocks = oa->o_blocks;
2193                 if (oa->o_valid & OBD_MD_FLMTIME)
2194                         oap->oap_loi->loi_lvb.lvb_mtime = oa->o_mtime;
2195                 if (oa->o_valid & OBD_MD_FLATIME)
2196                         oap->oap_loi->loi_lvb.lvb_atime = oa->o_atime;
2197                 if (oa->o_valid & OBD_MD_FLCTIME)
2198                         oap->oap_loi->loi_lvb.lvb_ctime = oa->o_ctime;
2199         }
2200
2201         rc = oap->oap_caller_ops->ap_completion(env, oap->oap_caller_data,
2202                                                 oap->oap_cmd, oa, rc);
2203
2204         /* ll_ap_completion (from llite) drops PG_locked. so, a new
2205          * I/O on the page could start, but OSC calls it under lock
2206          * and thus we can add oap back to pending safely */
2207         if (rc)
2208                 /* upper layer wants to leave the page on pending queue */
2209                 osc_oap_to_pending(oap);
2210         else
2211                 osc_exit_cache(cli, oap, sent);
2212         EXIT;
2213 }
2214
2215 static int brw_interpret(const struct lu_env *env,
2216                          struct ptlrpc_request *req, void *data, int rc)
2217 {
2218         struct osc_brw_async_args *aa = data;
2219         struct client_obd *cli;
2220         int async;
2221         ENTRY;
2222
2223         rc = osc_brw_fini_request(req, rc);
2224         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2225         if (osc_recoverable_error(rc)) {
2226                 /* Only retry once for mmaped files since the mmaped page
2227                  * might be modified at anytime. We have to retry at least
2228                  * once in case there WAS really a corruption of the page
2229                  * on the network, that was not caused by mmap() modifying
2230                  * the page. Bug11742 */
2231                 if ((rc == -EAGAIN) && (aa->aa_resends > 0) &&
2232                     aa->aa_oa->o_valid & OBD_MD_FLFLAGS &&
2233                     aa->aa_oa->o_flags & OBD_FL_MMAP) {
2234                         rc = 0;
2235                 } else {
2236                         rc = osc_brw_redo_request(req, aa);
2237                         if (rc == 0)
2238                                 RETURN(0);
2239                 }
2240         }
2241
2242         if (aa->aa_ocapa) {
2243                 capa_put(aa->aa_ocapa);
2244                 aa->aa_ocapa = NULL;
2245         }
2246
2247         cli = aa->aa_cli;
2248
2249         client_obd_list_lock(&cli->cl_loi_list_lock);
2250
2251         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2252          * is called so we know whether to go to sync BRWs or wait for more
2253          * RPCs to complete */
2254         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2255                 cli->cl_w_in_flight--;
2256         else
2257                 cli->cl_r_in_flight--;
2258
2259         async = cfs_list_empty(&aa->aa_oaps);
2260         if (!async) { /* from osc_send_oap_rpc() */
2261                 struct osc_async_page *oap, *tmp;
2262                 /* the caller may re-use the oap after the completion call so
2263                  * we need to clean it up a little */
2264                 cfs_list_for_each_entry_safe(oap, tmp, &aa->aa_oaps,
2265                                              oap_rpc_item) {
2266                         cfs_list_del_init(&oap->oap_rpc_item);
2267                         osc_ap_completion(env, cli, aa->aa_oa, oap, 1, rc);
2268                 }
2269                 OBDO_FREE(aa->aa_oa);
2270         } else { /* from async_internal() */
2271                 obd_count i;
2272                 for (i = 0; i < aa->aa_page_count; i++)
2273                         osc_release_write_grant(aa->aa_cli, aa->aa_ppga[i], 1);
2274         }
2275         osc_wake_cache_waiters(cli);
2276         osc_check_rpcs(env, cli);
2277         client_obd_list_unlock(&cli->cl_loi_list_lock);
2278         if (!async)
2279                 cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
2280                                   req->rq_bulk->bd_nob_transferred);
2281         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2282
2283         RETURN(rc);
2284 }
2285
2286 static struct ptlrpc_request *osc_build_req(const struct lu_env *env,
2287                                             struct client_obd *cli,
2288                                             cfs_list_t *rpc_list,
2289                                             int page_count, int cmd)
2290 {
2291         struct ptlrpc_request *req;
2292         struct brw_page **pga = NULL;
2293         struct osc_brw_async_args *aa;
2294         struct obdo *oa = NULL;
2295         const struct obd_async_page_ops *ops = NULL;
2296         void *caller_data = NULL;
2297         struct osc_async_page *oap;
2298         struct osc_async_page *tmp;
2299         struct ost_body *body;
2300         struct cl_req *clerq = NULL;
2301         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2302         struct ldlm_lock *lock = NULL;
2303         struct cl_req_attr crattr;
2304         int i, rc, mpflag = 0;
2305
2306         ENTRY;
2307         LASSERT(!cfs_list_empty(rpc_list));
2308
2309         if (cmd & OBD_BRW_MEMALLOC)
2310                 mpflag = cfs_memory_pressure_get_and_set();
2311
2312         memset(&crattr, 0, sizeof crattr);
2313         OBD_ALLOC(pga, sizeof(*pga) * page_count);
2314         if (pga == NULL)
2315                 GOTO(out, req = ERR_PTR(-ENOMEM));
2316
2317         OBDO_ALLOC(oa);
2318         if (oa == NULL)
2319                 GOTO(out, req = ERR_PTR(-ENOMEM));
2320
2321         i = 0;
2322         cfs_list_for_each_entry(oap, rpc_list, oap_rpc_item) {
2323                 struct cl_page *page = osc_oap2cl_page(oap);
2324                 if (ops == NULL) {
2325                         ops = oap->oap_caller_ops;
2326                         caller_data = oap->oap_caller_data;
2327
2328                         clerq = cl_req_alloc(env, page, crt,
2329                                              1 /* only 1-object rpcs for
2330                                                 * now */);
2331                         if (IS_ERR(clerq))
2332                                 GOTO(out, req = (void *)clerq);
2333                         lock = oap->oap_ldlm_lock;
2334                 }
2335                 pga[i] = &oap->oap_brw_page;
2336                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2337                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
2338                        pga[i]->pg, cfs_page_index(oap->oap_page), oap, pga[i]->flag);
2339                 i++;
2340                 cl_req_page_add(env, clerq, page);
2341         }
2342
2343         /* always get the data for the obdo for the rpc */
2344         LASSERT(ops != NULL);
2345         crattr.cra_oa = oa;
2346         crattr.cra_capa = NULL;
2347         cl_req_attr_set(env, clerq, &crattr, ~0ULL);
2348         if (lock) {
2349                 oa->o_handle = lock->l_remote_handle;
2350                 oa->o_valid |= OBD_MD_FLHANDLE;
2351         }
2352
2353         rc = cl_req_prep(env, clerq);
2354         if (rc != 0) {
2355                 CERROR("cl_req_prep failed: %d\n", rc);
2356                 GOTO(out, req = ERR_PTR(rc));
2357         }
2358
2359         sort_brw_pages(pga, page_count);
2360         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
2361                                   pga, &req, crattr.cra_capa, 1, 0);
2362         if (rc != 0) {
2363                 CERROR("prep_req failed: %d\n", rc);
2364                 GOTO(out, req = ERR_PTR(rc));
2365         }
2366
2367         if (cmd & OBD_BRW_MEMALLOC)
2368                 req->rq_memalloc = 1;
2369
2370         /* Need to update the timestamps after the request is built in case
2371          * we race with setattr (locally or in queue at OST).  If OST gets
2372          * later setattr before earlier BRW (as determined by the request xid),
2373          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2374          * way to do this in a single call.  bug 10150 */
2375         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2376         cl_req_attr_set(env, clerq, &crattr,
2377                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
2378
2379         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2380         aa = ptlrpc_req_async_args(req);
2381         CFS_INIT_LIST_HEAD(&aa->aa_oaps);
2382         cfs_list_splice(rpc_list, &aa->aa_oaps);
2383         CFS_INIT_LIST_HEAD(rpc_list);
2384         aa->aa_clerq = clerq;
2385 out:
2386         if (cmd & OBD_BRW_MEMALLOC)
2387                 cfs_memory_pressure_restore(mpflag);
2388
2389         capa_put(crattr.cra_capa);
2390         if (IS_ERR(req)) {
2391                 if (oa)
2392                         OBDO_FREE(oa);
2393                 if (pga)
2394                         OBD_FREE(pga, sizeof(*pga) * page_count);
2395                 /* this should happen rarely and is pretty bad, it makes the
2396                  * pending list not follow the dirty order */
2397                 client_obd_list_lock(&cli->cl_loi_list_lock);
2398                 cfs_list_for_each_entry_safe(oap, tmp, rpc_list, oap_rpc_item) {
2399                         cfs_list_del_init(&oap->oap_rpc_item);
2400
2401                         /* queued sync pages can be torn down while the pages
2402                          * were between the pending list and the rpc */
2403                         if (oap->oap_interrupted) {
2404                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
2405                                 osc_ap_completion(env, cli, NULL, oap, 0,
2406                                                   oap->oap_count);
2407                                 continue;
2408                         }
2409                         osc_ap_completion(env, cli, NULL, oap, 0, PTR_ERR(req));
2410                 }
2411                 if (clerq && !IS_ERR(clerq))
2412                         cl_req_completion(env, clerq, PTR_ERR(req));
2413         }
2414         RETURN(req);
2415 }
2416
2417 /**
2418  * prepare pages for ASYNC io and put pages in send queue.
2419  *
2420  * \param cmd OBD_BRW_* macroses
2421  * \param lop pending pages
2422  *
2423  * \return zero if no page added to send queue.
2424  * \return 1 if pages successfully added to send queue.
2425  * \return negative on errors.
2426  */
2427 static int
2428 osc_send_oap_rpc(const struct lu_env *env, struct client_obd *cli,
2429                  struct lov_oinfo *loi,
2430                  int cmd, struct loi_oap_pages *lop)
2431 {
2432         struct ptlrpc_request *req;
2433         obd_count page_count = 0;
2434         struct osc_async_page *oap = NULL, *tmp;
2435         struct osc_brw_async_args *aa;
2436         const struct obd_async_page_ops *ops;
2437         CFS_LIST_HEAD(rpc_list);
2438         int srvlock = 0, mem_tight = 0;
2439         struct cl_object *clob = NULL;
2440         obd_off starting_offset = OBD_OBJECT_EOF;
2441         unsigned int ending_offset;
2442         int starting_page_off = 0;
2443         ENTRY;
2444
2445         /* ASYNC_HP pages first. At present, when the lock the pages is
2446          * to be canceled, the pages covered by the lock will be sent out
2447          * with ASYNC_HP. We have to send out them as soon as possible. */
2448         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_urgent, oap_urgent_item) {
2449                 if (oap->oap_async_flags & ASYNC_HP)
2450                         cfs_list_move(&oap->oap_pending_item, &lop->lop_pending);
2451                 if (++page_count >= cli->cl_max_pages_per_rpc)
2452                         break;
2453         }
2454         page_count = 0;
2455
2456         /* first we find the pages we're allowed to work with */
2457         cfs_list_for_each_entry_safe(oap, tmp, &lop->lop_pending,
2458                                      oap_pending_item) {
2459                 ops = oap->oap_caller_ops;
2460
2461                 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
2462                          "magic 0x%x\n", oap, oap->oap_magic);
2463
2464                 if (clob == NULL) {
2465                         /* pin object in memory, so that completion call-backs
2466                          * can be safely called under client_obd_list lock. */
2467                         clob = osc_oap2cl_page(oap)->cp_obj;
2468                         cl_object_get(clob);
2469                 }
2470
2471                 if (page_count != 0 &&
2472                     srvlock != !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK)) {
2473                         CDEBUG(D_PAGE, "SRVLOCK flag mismatch,"
2474                                " oap %p, page %p, srvlock %u\n",
2475                                oap, oap->oap_brw_page.pg, (unsigned)!srvlock);
2476                         break;
2477                 }
2478
2479                 /* If there is a gap at the start of this page, it can't merge
2480                  * with any previous page, so we'll hand the network a
2481                  * "fragmented" page array that it can't transfer in 1 RDMA */
2482                 if (oap->oap_obj_off < starting_offset) {
2483                         if (starting_page_off != 0)
2484                                 break;
2485
2486                         starting_page_off = oap->oap_page_off;
2487                         starting_offset = oap->oap_obj_off + starting_page_off;
2488                 } else if (oap->oap_page_off != 0)
2489                         break;
2490
2491                 /* in llite being 'ready' equates to the page being locked
2492                  * until completion unlocks it.  commit_write submits a page
2493                  * as not ready because its unlock will happen unconditionally
2494                  * as the call returns.  if we race with commit_write giving
2495                  * us that page we don't want to create a hole in the page
2496                  * stream, so we stop and leave the rpc to be fired by
2497                  * another dirtier or kupdated interval (the not ready page
2498                  * will still be on the dirty list).  we could call in
2499                  * at the end of ll_file_write to process the queue again. */
2500                 if (!(oap->oap_async_flags & ASYNC_READY)) {
2501                         int rc = ops->ap_make_ready(env, oap->oap_caller_data,
2502                                                     cmd);
2503                         if (rc < 0)
2504                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
2505                                                 "instead of ready\n", oap,
2506                                                 oap->oap_page, rc);
2507                         switch (rc) {
2508                         case -EAGAIN:
2509                                 /* llite is telling us that the page is still
2510                                  * in commit_write and that we should try
2511                                  * and put it in an rpc again later.  we
2512                                  * break out of the loop so we don't create
2513                                  * a hole in the sequence of pages in the rpc
2514                                  * stream.*/
2515                                 oap = NULL;
2516                                 break;
2517                         case -EINTR:
2518                                 /* the io isn't needed.. tell the checks
2519                                  * below to complete the rpc with EINTR */
2520                                 cfs_spin_lock(&oap->oap_lock);
2521                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
2522                                 cfs_spin_unlock(&oap->oap_lock);
2523                                 oap->oap_count = -EINTR;
2524                                 break;
2525                         case 0:
2526                                 cfs_spin_lock(&oap->oap_lock);
2527                                 oap->oap_async_flags |= ASYNC_READY;
2528                                 cfs_spin_unlock(&oap->oap_lock);
2529                                 break;
2530                         default:
2531                                 LASSERTF(0, "oap %p page %p returned %d "
2532                                             "from make_ready\n", oap,
2533                                             oap->oap_page, rc);
2534                                 break;
2535                         }
2536                 }
2537                 if (oap == NULL)
2538                         break;
2539                 /*
2540                  * Page submitted for IO has to be locked. Either by
2541                  * ->ap_make_ready() or by higher layers.
2542                  */
2543 #if defined(__KERNEL__) && defined(__linux__)
2544                 {
2545                         struct cl_page *page;
2546
2547                         page = osc_oap2cl_page(oap);
2548
2549                         if (page->cp_type == CPT_CACHEABLE &&
2550                             !(PageLocked(oap->oap_page) &&
2551                               (CheckWriteback(oap->oap_page, cmd)))) {
2552                                 CDEBUG(D_PAGE, "page %p lost wb %lx/%x\n",
2553                                        oap->oap_page,
2554                                        (long)oap->oap_page->flags,
2555                                        oap->oap_async_flags);
2556                                 LBUG();
2557                         }
2558                 }
2559 #endif
2560
2561                 /* take the page out of our book-keeping */
2562                 cfs_list_del_init(&oap->oap_pending_item);
2563                 lop_update_pending(cli, lop, cmd, -1);
2564                 cfs_list_del_init(&oap->oap_urgent_item);
2565
2566                 /* ask the caller for the size of the io as the rpc leaves. */
2567                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE)) {
2568                         oap->oap_count =
2569                                 ops->ap_refresh_count(env, oap->oap_caller_data,
2570                                                       cmd);
2571                         LASSERT(oap->oap_page_off + oap->oap_count <= CFS_PAGE_SIZE);
2572                 }
2573                 if (oap->oap_count <= 0) {
2574                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
2575                                oap->oap_count);
2576                         osc_ap_completion(env, cli, NULL,
2577                                           oap, 0, oap->oap_count);
2578                         continue;
2579                 }
2580
2581                 /* now put the page back in our accounting */
2582                 cfs_list_add_tail(&oap->oap_rpc_item, &rpc_list);
2583                 if (page_count++ == 0)
2584                         srvlock = !!(oap->oap_brw_flags & OBD_BRW_SRVLOCK);
2585
2586                 if (oap->oap_brw_flags & OBD_BRW_MEMALLOC)
2587                         mem_tight = 1;
2588
2589                 /* End on a PTLRPC_MAX_BRW_SIZE boundary.  We want full-sized
2590                  * RPCs aligned on PTLRPC_MAX_BRW_SIZE boundaries to help reads
2591                  * have the same alignment as the initial writes that allocated
2592                  * extents on the server. */
2593                 ending_offset = oap->oap_obj_off + oap->oap_page_off +
2594                                 oap->oap_count;
2595                 if (!(ending_offset & (PTLRPC_MAX_BRW_SIZE - 1)))
2596                         break;
2597
2598                 if (page_count >= cli->cl_max_pages_per_rpc)
2599                         break;
2600
2601                 /* If there is a gap at the end of this page, it can't merge
2602                  * with any subsequent pages, so we'll hand the network a
2603                  * "fragmented" page array that it can't transfer in 1 RDMA */
2604                 if (oap->oap_page_off + oap->oap_count < CFS_PAGE_SIZE)
2605                         break;
2606         }
2607
2608         osc_wake_cache_waiters(cli);
2609
2610         loi_list_maint(cli, loi);
2611
2612         client_obd_list_unlock(&cli->cl_loi_list_lock);
2613
2614         if (clob != NULL)
2615                 cl_object_put(env, clob);
2616
2617         if (page_count == 0) {
2618                 client_obd_list_lock(&cli->cl_loi_list_lock);
2619                 RETURN(0);
2620         }
2621
2622         req = osc_build_req(env, cli, &rpc_list, page_count,
2623                             mem_tight ? (cmd | OBD_BRW_MEMALLOC) : cmd);
2624         if (IS_ERR(req)) {
2625                 LASSERT(cfs_list_empty(&rpc_list));
2626                 loi_list_maint(cli, loi);
2627                 RETURN(PTR_ERR(req));
2628         }
2629
2630         aa = ptlrpc_req_async_args(req);
2631
2632         starting_offset &= PTLRPC_MAX_BRW_SIZE - 1;
2633         if (cmd == OBD_BRW_READ) {
2634                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2635                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2636                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2637                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2638         } else {
2639                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2640                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
2641                                  cli->cl_w_in_flight);
2642                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2643                                       (starting_offset >> CFS_PAGE_SHIFT) + 1);
2644         }
2645         ptlrpc_lprocfs_brw(req, aa->aa_requested_nob);
2646
2647         client_obd_list_lock(&cli->cl_loi_list_lock);
2648
2649         if (cmd == OBD_BRW_READ)
2650                 cli->cl_r_in_flight++;
2651         else
2652                 cli->cl_w_in_flight++;
2653
2654         /* queued sync pages can be torn down while the pages
2655          * were between the pending list and the rpc */
2656         tmp = NULL;
2657         cfs_list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2658                 /* only one oap gets a request reference */
2659                 if (tmp == NULL)
2660                         tmp = oap;
2661                 if (oap->oap_interrupted && !req->rq_intr) {
2662                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
2663                                oap, req);
2664                         ptlrpc_mark_interrupted(req);
2665                 }
2666         }
2667         if (tmp != NULL)
2668                 tmp->oap_request = ptlrpc_request_addref(req);
2669
2670         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2671                   page_count, aa, cli->cl_r_in_flight, cli->cl_w_in_flight);
2672
2673         req->rq_interpret_reply = brw_interpret;
2674         ptlrpcd_add_req(req, PSCOPE_BRW);
2675         RETURN(1);
2676 }
2677
2678 #define LOI_DEBUG(LOI, STR, args...)                                     \
2679         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
2680                !cfs_list_empty(&(LOI)->loi_ready_item) ||                \
2681                !cfs_list_empty(&(LOI)->loi_hp_ready_item),               \
2682                (LOI)->loi_write_lop.lop_num_pending,                     \
2683                !cfs_list_empty(&(LOI)->loi_write_lop.lop_urgent),        \
2684                (LOI)->loi_read_lop.lop_num_pending,                      \
2685                !cfs_list_empty(&(LOI)->loi_read_lop.lop_urgent),         \
2686                args)                                                     \
2687
2688 /* This is called by osc_check_rpcs() to find which objects have pages that
2689  * we could be sending.  These lists are maintained by lop_makes_rpc(). */
2690 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
2691 {
2692         ENTRY;
2693
2694         /* First return objects that have blocked locks so that they
2695          * will be flushed quickly and other clients can get the lock,
2696          * then objects which have pages ready to be stuffed into RPCs */
2697         if (!cfs_list_empty(&cli->cl_loi_hp_ready_list))
2698                 RETURN(cfs_list_entry(cli->cl_loi_hp_ready_list.next,
2699                                       struct lov_oinfo, loi_hp_ready_item));
2700         if (!cfs_list_empty(&cli->cl_loi_ready_list))
2701                 RETURN(cfs_list_entry(cli->cl_loi_ready_list.next,
2702                                       struct lov_oinfo, loi_ready_item));
2703
2704         /* then if we have cache waiters, return all objects with queued
2705          * writes.  This is especially important when many small files
2706          * have filled up the cache and not been fired into rpcs because
2707          * they don't pass the nr_pending/object threshhold */
2708         if (!cfs_list_empty(&cli->cl_cache_waiters) &&
2709             !cfs_list_empty(&cli->cl_loi_write_list))
2710                 RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2711                                       struct lov_oinfo, loi_write_item));
2712
2713         /* then return all queued objects when we have an invalid import
2714          * so that they get flushed */
2715         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
2716                 if (!cfs_list_empty(&cli->cl_loi_write_list))
2717                         RETURN(cfs_list_entry(cli->cl_loi_write_list.next,
2718                                               struct lov_oinfo,
2719                                               loi_write_item));
2720                 if (!cfs_list_empty(&cli->cl_loi_read_list))
2721                         RETURN(cfs_list_entry(cli->cl_loi_read_list.next,
2722                                               struct lov_oinfo, loi_read_item));
2723         }
2724         RETURN(NULL);
2725 }
2726
2727 static int osc_max_rpc_in_flight(struct client_obd *cli, struct lov_oinfo *loi)
2728 {
2729         struct osc_async_page *oap;
2730         int hprpc = 0;
2731
2732         if (!cfs_list_empty(&loi->loi_write_lop.lop_urgent)) {
2733                 oap = cfs_list_entry(loi->loi_write_lop.lop_urgent.next,
2734                                      struct osc_async_page, oap_urgent_item);
2735                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2736         }
2737
2738         if (!hprpc && !cfs_list_empty(&loi->loi_read_lop.lop_urgent)) {
2739                 oap = cfs_list_entry(loi->loi_read_lop.lop_urgent.next,
2740                                      struct osc_async_page, oap_urgent_item);
2741                 hprpc = !!(oap->oap_async_flags & ASYNC_HP);
2742         }
2743
2744         return rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight + hprpc;
2745 }
2746
2747 /* called with the loi list lock held */
2748 void osc_check_rpcs(const struct lu_env *env, struct client_obd *cli)
2749 {
2750         struct lov_oinfo *loi;
2751         int rc = 0, race_counter = 0;
2752         ENTRY;
2753
2754         while ((loi = osc_next_loi(cli)) != NULL) {
2755                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
2756
2757                 if (osc_max_rpc_in_flight(cli, loi))
2758                         break;
2759
2760                 /* attempt some read/write balancing by alternating between
2761                  * reads and writes in an object.  The makes_rpc checks here
2762                  * would be redundant if we were getting read/write work items
2763                  * instead of objects.  we don't want send_oap_rpc to drain a
2764                  * partial read pending queue when we're given this object to
2765                  * do io on writes while there are cache waiters */
2766                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
2767                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_WRITE,
2768                                               &loi->loi_write_lop);
2769                         if (rc < 0) {
2770                                 CERROR("Write request failed with %d\n", rc);
2771
2772                                 /* osc_send_oap_rpc failed, mostly because of
2773                                  * memory pressure.
2774                                  *
2775                                  * It can't break here, because if:
2776                                  *  - a page was submitted by osc_io_submit, so
2777                                  *    page locked;
2778                                  *  - no request in flight
2779                                  *  - no subsequent request
2780                                  * The system will be in live-lock state,
2781                                  * because there is no chance to call
2782                                  * osc_io_unplug() and osc_check_rpcs() any
2783                                  * more. pdflush can't help in this case,
2784                                  * because it might be blocked at grabbing
2785                                  * the page lock as we mentioned.
2786                                  *
2787                                  * Anyway, continue to drain pages. */
2788                                 /* break; */
2789                         }
2790
2791                         if (rc > 0)
2792                                 race_counter = 0;
2793                         else if (rc == 0)
2794                                 race_counter++;
2795                 }
2796                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
2797                         rc = osc_send_oap_rpc(env, cli, loi, OBD_BRW_READ,
2798                                               &loi->loi_read_lop);
2799                         if (rc < 0)
2800                                 CERROR("Read request failed with %d\n", rc);
2801
2802                         if (rc > 0)
2803                                 race_counter = 0;
2804                         else if (rc == 0)
2805                                 race_counter++;
2806                 }
2807
2808                 /* attempt some inter-object balancing by issuing rpcs
2809                  * for each object in turn */
2810                 if (!cfs_list_empty(&loi->loi_hp_ready_item))
2811                         cfs_list_del_init(&loi->loi_hp_ready_item);
2812                 if (!cfs_list_empty(&loi->loi_ready_item))
2813                         cfs_list_del_init(&loi->loi_ready_item);
2814                 if (!cfs_list_empty(&loi->loi_write_item))
2815                         cfs_list_del_init(&loi->loi_write_item);
2816                 if (!cfs_list_empty(&loi->loi_read_item))
2817                         cfs_list_del_init(&loi->loi_read_item);
2818
2819                 loi_list_maint(cli, loi);
2820
2821                 /* send_oap_rpc fails with 0 when make_ready tells it to
2822                  * back off.  llite's make_ready does this when it tries
2823                  * to lock a page queued for write that is already locked.
2824                  * we want to try sending rpcs from many objects, but we
2825                  * don't want to spin failing with 0.  */
2826                 if (race_counter == 10)
2827                         break;
2828         }
2829         EXIT;
2830 }
2831
2832 /* we're trying to queue a page in the osc so we're subject to the
2833  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
2834  * If the osc's queued pages are already at that limit, then we want to sleep
2835  * until there is space in the osc's queue for us.  We also may be waiting for
2836  * write credits from the OST if there are RPCs in flight that may return some
2837  * before we fall back to sync writes.
2838  *
2839  * We need this know our allocation was granted in the presence of signals */
2840 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
2841 {
2842         int rc;
2843         ENTRY;
2844         client_obd_list_lock(&cli->cl_loi_list_lock);
2845         rc = cfs_list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
2846         client_obd_list_unlock(&cli->cl_loi_list_lock);
2847         RETURN(rc);
2848 };
2849
2850 /**
2851  * Non-blocking version of osc_enter_cache() that consumes grant only when it
2852  * is available.
2853  */
2854 int osc_enter_cache_try(const struct lu_env *env,
2855                         struct client_obd *cli, struct lov_oinfo *loi,
2856                         struct osc_async_page *oap, int transient)
2857 {
2858         int has_grant;
2859
2860         has_grant = cli->cl_avail_grant >= CFS_PAGE_SIZE;
2861         if (has_grant) {
2862                 osc_consume_write_grant(cli, &oap->oap_brw_page);
2863                 if (transient) {
2864                         cli->cl_dirty_transit += CFS_PAGE_SIZE;
2865                         cfs_atomic_inc(&obd_dirty_transit_pages);
2866                         oap->oap_brw_flags |= OBD_BRW_NOCACHE;
2867                 }
2868         }
2869         return has_grant;
2870 }
2871
2872 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
2873  * grant or cache space. */
2874 static int osc_enter_cache(const struct lu_env *env,
2875                            struct client_obd *cli, struct lov_oinfo *loi,
2876                            struct osc_async_page *oap)
2877 {
2878         struct osc_cache_waiter ocw;
2879         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
2880
2881         ENTRY;
2882
2883         CDEBUG(D_CACHE, "dirty: %ld/%d dirty_max: %ld/%d dropped: %lu "
2884                "grant: %lu\n", cli->cl_dirty, cfs_atomic_read(&obd_dirty_pages),
2885                cli->cl_dirty_max, obd_max_dirty_pages,
2886                cli->cl_lost_grant, cli->cl_avail_grant);
2887
2888         /* force the caller to try sync io.  this can jump the list
2889          * of queued writes and create a discontiguous rpc stream */
2890         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_GRANT) ||
2891             cli->cl_dirty_max < CFS_PAGE_SIZE     ||
2892             cli->cl_ar.ar_force_sync || loi->loi_ar.ar_force_sync)
2893                 RETURN(-EDQUOT);
2894
2895         /* Hopefully normal case - cache space and write credits available */
2896         if (cli->cl_dirty + CFS_PAGE_SIZE <= cli->cl_dirty_max &&
2897             cfs_atomic_read(&obd_dirty_pages) + 1 <= obd_max_dirty_pages &&
2898             osc_enter_cache_try(env, cli, loi, oap, 0))
2899                 RETURN(0);
2900
2901         /* It is safe to block as a cache waiter as long as there is grant
2902          * space available or the hope of additional grant being returned
2903          * when an in flight write completes.  Using the write back cache
2904          * if possible is preferable to sending the data synchronously
2905          * because write pages can then be merged in to large requests.
2906          * The addition of this cache waiter will causing pending write
2907          * pages to be sent immediately. */
2908         if (cli->cl_w_in_flight || cli->cl_avail_grant >= CFS_PAGE_SIZE) {
2909                 cfs_list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
2910                 cfs_waitq_init(&ocw.ocw_waitq);
2911                 ocw.ocw_oap = oap;
2912                 ocw.ocw_rc = 0;
2913
2914                 loi_list_maint(cli, loi);
2915                 osc_check_rpcs(env, cli);
2916                 client_obd_list_unlock(&cli->cl_loi_list_lock);
2917
2918                 CDEBUG(D_CACHE, "sleeping for cache space\n");
2919                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
2920
2921                 client_obd_list_lock(&cli->cl_loi_list_lock);
2922                 if (!cfs_list_empty(&ocw.ocw_entry)) {
2923                         cfs_list_del(&ocw.ocw_entry);
2924                         RETURN(-EINTR);
2925                 }
2926                 RETURN(ocw.ocw_rc);
2927         }
2928
2929         RETURN(-EDQUOT);
2930 }
2931
2932
2933 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2934                         struct lov_oinfo *loi, cfs_page_t *page,
2935                         obd_off offset, const struct obd_async_page_ops *ops,
2936                         void *data, void **res, int nocache,
2937                         struct lustre_handle *lockh)
2938 {
2939         struct osc_async_page *oap;
2940
2941         ENTRY;
2942
2943         if (!page)
2944                 return cfs_size_round(sizeof(*oap));
2945
2946         oap = *res;
2947         oap->oap_magic = OAP_MAGIC;
2948         oap->oap_cli = &exp->exp_obd->u.cli;
2949         oap->oap_loi = loi;
2950
2951         oap->oap_caller_ops = ops;
2952         oap->oap_caller_data = data;
2953
2954         oap->oap_page = page;
2955         oap->oap_obj_off = offset;
2956         if (!client_is_remote(exp) &&
2957             cfs_capable(CFS_CAP_SYS_RESOURCE))
2958                 oap->oap_brw_flags = OBD_BRW_NOQUOTA;
2959
2960         LASSERT(!(offset & ~CFS_PAGE_MASK));
2961
2962         CFS_INIT_LIST_HEAD(&oap->oap_pending_item);
2963         CFS_INIT_LIST_HEAD(&oap->oap_urgent_item);
2964         CFS_INIT_LIST_HEAD(&oap->oap_rpc_item);
2965         CFS_INIT_LIST_HEAD(&oap->oap_page_list);
2966
2967         cfs_spin_lock_init(&oap->oap_lock);
2968         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2969         RETURN(0);
2970 }
2971
2972 int osc_queue_async_io(const struct lu_env *env, struct obd_export *exp,
2973                        struct lov_stripe_md *lsm, struct lov_oinfo *loi,
2974                        struct osc_async_page *oap, int cmd, int off,
2975                        int count, obd_flag brw_flags, enum async_flags async_flags)
2976 {
2977         struct client_obd *cli = &exp->exp_obd->u.cli;
2978         int rc = 0;
2979         ENTRY;
2980
2981         if (oap->oap_magic != OAP_MAGIC)
2982                 RETURN(-EINVAL);
2983
2984         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2985                 RETURN(-EIO);
2986
2987         if (!cfs_list_empty(&oap->oap_pending_item) ||
2988             !cfs_list_empty(&oap->oap_urgent_item) ||
2989             !cfs_list_empty(&oap->oap_rpc_item))
2990                 RETURN(-EBUSY);
2991
2992         /* check if the file's owner/group is over quota */
2993         if ((cmd & OBD_BRW_WRITE) && !(cmd & OBD_BRW_NOQUOTA)) {
2994                 struct cl_object *obj;
2995                 struct cl_attr    attr; /* XXX put attr into thread info */
2996                 unsigned int qid[MAXQUOTAS];
2997
2998                 obj = cl_object_top(osc_oap2cl_page(oap)->cp_obj);
2999
3000                 cl_object_attr_lock(obj);
3001                 rc = cl_object_attr_get(env, obj, &attr);
3002                 cl_object_attr_unlock(obj);
3003
3004                 qid[USRQUOTA] = attr.cat_uid;
3005                 qid[GRPQUOTA] = attr.cat_gid;
3006                 if (rc == 0 &&
3007                     lquota_chkdq(quota_interface, cli, qid) == NO_QUOTA)
3008                         rc = -EDQUOT;
3009                 if (rc)
3010                         RETURN(rc);
3011         }
3012
3013         if (loi == NULL)
3014                 loi = lsm->lsm_oinfo[0];
3015
3016         client_obd_list_lock(&cli->cl_loi_list_lock);
3017
3018         LASSERT(off + count <= CFS_PAGE_SIZE);
3019         oap->oap_cmd = cmd;
3020         oap->oap_page_off = off;
3021         oap->oap_count = count;
3022         oap->oap_brw_flags = brw_flags;
3023         /* Give a hint to OST that requests are coming from kswapd - bug19529 */
3024         if (cfs_memory_pressure_get())
3025                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
3026         cfs_spin_lock(&oap->oap_lock);
3027         oap->oap_async_flags = async_flags;
3028         cfs_spin_unlock(&oap->oap_lock);
3029
3030         if (cmd & OBD_BRW_WRITE) {
3031                 rc = osc_enter_cache(env, cli, loi, oap);
3032                 if (rc) {
3033                         client_obd_list_unlock(&cli->cl_loi_list_lock);
3034                         RETURN(rc);
3035                 }
3036         }
3037
3038         osc_oap_to_pending(oap);
3039         loi_list_maint(cli, loi);
3040
3041         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
3042                   cmd);
3043
3044         osc_check_rpcs(env, cli);
3045         client_obd_list_unlock(&cli->cl_loi_list_lock);
3046
3047         RETURN(0);
3048 }
3049
3050 /* aka (~was & now & flag), but this is more clear :) */
3051 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
3052
3053 int osc_set_async_flags_base(struct client_obd *cli,
3054                              struct lov_oinfo *loi, struct osc_async_page *oap,
3055                              obd_flag async_flags)
3056 {
3057         struct loi_oap_pages *lop;
3058         int flags = 0;
3059         ENTRY;
3060
3061         LASSERT(!cfs_list_empty(&oap->oap_pending_item));
3062
3063         if (oap->oap_cmd & OBD_BRW_WRITE) {
3064                 lop = &loi->loi_write_lop;
3065         } else {
3066                 lop = &loi->loi_read_lop;
3067         }
3068
3069         if ((oap->oap_async_flags & async_flags) == async_flags)
3070                 RETURN(0);
3071
3072         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
3073                 flags |= ASYNC_READY;
3074
3075         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT) &&
3076             cfs_list_empty(&oap->oap_rpc_item)) {
3077                 if (oap->oap_async_flags & ASYNC_HP)
3078                         cfs_list_add(&oap->oap_urgent_item, &lop->lop_urgent);
3079                 else
3080                         cfs_list_add_tail(&oap->oap_urgent_item,
3081                                           &lop->lop_urgent);
3082                 flags |= ASYNC_URGENT;
3083                 loi_list_maint(cli, loi);
3084         }
3085         cfs_spin_lock(&oap->oap_lock);
3086         oap->oap_async_flags |= flags;
3087         cfs_spin_unlock(&oap->oap_lock);
3088
3089         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
3090                         oap->oap_async_flags);
3091         RETURN(0);
3092 }
3093
3094 int osc_teardown_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
3095                             struct lov_oinfo *loi, struct osc_async_page *oap)
3096 {
3097         struct client_obd *cli = &exp->exp_obd->u.cli;
3098         struct loi_oap_pages *lop;
3099         int rc = 0;
3100         ENTRY;
3101
3102         if (oap->oap_magic != OAP_MAGIC)
3103                 RETURN(-EINVAL);
3104
3105         if (loi == NULL)
3106                 loi = lsm->lsm_oinfo[0];
3107
3108         if (oap->oap_cmd & OBD_BRW_WRITE) {
3109                 lop = &loi->loi_write_lop;
3110         } else {
3111                 lop = &loi->loi_read_lop;
3112         }
3113
3114         client_obd_list_lock(&cli->cl_loi_list_lock);
3115
3116         if (!cfs_list_empty(&oap->oap_rpc_item))
3117                 GOTO(out, rc = -EBUSY);
3118
3119         osc_exit_cache(cli, oap, 0);
3120         osc_wake_cache_waiters(cli);
3121
3122         if (!cfs_list_empty(&oap->oap_urgent_item)) {
3123                 cfs_list_del_init(&oap->oap_urgent_item);
3124                 cfs_spin_lock(&oap->oap_lock);
3125                 oap->oap_async_flags &= ~(ASYNC_URGENT | ASYNC_HP);
3126                 cfs_spin_unlock(&oap->oap_lock);
3127         }
3128         if (!cfs_list_empty(&oap->oap_pending_item)) {
3129                 cfs_list_del_init(&oap->oap_pending_item);
3130                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
3131         }
3132         loi_list_maint(cli, loi);
3133         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
3134 out:
3135         client_obd_list_unlock(&cli->cl_loi_list_lock);
3136         RETURN(rc);
3137 }
3138
3139 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
3140                                         struct ldlm_enqueue_info *einfo)
3141 {
3142         void *data = einfo->ei_cbdata;
3143         int set = 0;
3144
3145         LASSERT(lock != NULL);
3146         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
3147         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
3148         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
3149         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
3150
3151         lock_res_and_lock(lock);
3152         cfs_spin_lock(&osc_ast_guard);
3153
3154         if (lock->l_ast_data == NULL)
3155                 lock->l_ast_data = data;
3156         if (lock->l_ast_data == data)
3157                 set = 1;
3158
3159         cfs_spin_unlock(&osc_ast_guard);
3160         unlock_res_and_lock(lock);
3161
3162         return set;
3163 }
3164
3165 static int osc_set_data_with_check(struct lustre_handle *lockh,
3166                                    struct ldlm_enqueue_info *einfo)
3167 {
3168         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
3169         int set = 0;
3170
3171         if (lock != NULL) {
3172                 set = osc_set_lock_data_with_check(lock, einfo);
3173                 LDLM_LOCK_PUT(lock);
3174         } else
3175                 CERROR("lockh %p, data %p - client evicted?\n",
3176                        lockh, einfo->ei_cbdata);
3177         return set;
3178 }
3179
3180 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3181                              ldlm_iterator_t replace, void *data)
3182 {
3183         struct ldlm_res_id res_id;
3184         struct obd_device *obd = class_exp2obd(exp);
3185
3186         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3187         ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3188         return 0;
3189 }
3190
3191 /* find any ldlm lock of the inode in osc
3192  * return 0    not find
3193  *        1    find one
3194  *      < 0    error */
3195 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
3196                            ldlm_iterator_t replace, void *data)
3197 {
3198         struct ldlm_res_id res_id;
3199         struct obd_device *obd = class_exp2obd(exp);
3200         int rc = 0;
3201
3202         osc_build_res_name(lsm->lsm_object_id, lsm->lsm_object_seq, &res_id);
3203         rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
3204         if (rc == LDLM_ITER_STOP)
3205                 return(1);
3206         if (rc == LDLM_ITER_CONTINUE)
3207                 return(0);
3208         return(rc);
3209 }
3210
3211 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
3212                             obd_enqueue_update_f upcall, void *cookie,
3213                             int *flags, int rc)
3214 {
3215         int intent = *flags & LDLM_FL_HAS_INTENT;
3216         ENTRY;
3217
3218         if (intent) {
3219                 /* The request was created before ldlm_cli_enqueue call. */
3220                 if (rc == ELDLM_LOCK_ABORTED) {
3221                         struct ldlm_reply *rep;
3222                         rep = req_capsule_server_get(&req->rq_pill,
3223                                                      &RMF_DLM_REP);
3224
3225                         LASSERT(rep != NULL);
3226                         if (rep->lock_policy_res1)
3227                                 rc = rep->lock_policy_res1;
3228                 }
3229         }
3230
3231         if ((intent && rc == ELDLM_LOCK_ABORTED) || !rc) {
3232                 *flags |= LDLM_FL_LVB_READY;
3233                 CDEBUG(D_INODE,"got kms "LPU64" blocks "LPU64" mtime "LPU64"\n",
3234                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
3235         }
3236
3237         /* Call the update callback. */
3238         rc = (*upcall)(cookie, rc);
3239         RETURN(rc);
3240 }
3241
3242 static int osc_enqueue_interpret(const struct lu_env *env,
3243                                  struct ptlrpc_request *req,
3244                                  struct osc_enqueue_args *aa, int rc)
3245 {
3246         struct ldlm_lock *lock;
3247         struct lustre_handle handle;
3248         __u32 mode;
3249
3250         /* Make a local copy of a lock handle and a mode, because aa->oa_*
3251          * might be freed anytime after lock upcall has been called. */
3252         lustre_handle_copy(&handle, aa->oa_lockh);
3253         mode = aa->oa_ei->ei_mode;
3254
3255         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
3256          * be valid. */
3257         lock = ldlm_handle2lock(&handle);
3258
3259         /* Take an additional reference so that a blocking AST that
3260          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
3261          * to arrive after an upcall has been executed by
3262          * osc_enqueue_fini(). */
3263         ldlm_lock_addref(&handle, mode);
3264
3265         /* Let CP AST to grant the lock first. */
3266         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
3267
3268         /* Complete obtaining the lock procedure. */
3269         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
3270                                    mode, aa->oa_flags, aa->oa_lvb,
3271                                    sizeof(*aa->oa_lvb), &handle, rc);
3272         /* Complete osc stuff. */
3273         rc = osc_enqueue_fini(req, aa->oa_lvb,
3274                               aa->oa_upcall, aa->oa_cookie, aa->oa_flags, rc);
3275
3276         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
3277
3278         /* Release the lock for async request. */
3279         if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
3280                 /*
3281                  * Releases a reference taken by ldlm_cli_enqueue(), if it is
3282                  * not already released by
3283                  * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
3284                  */
3285                 ldlm_lock_decref(&handle, mode);
3286
3287         LASSERTF(lock != NULL, "lockh %p, req %p, aa %p - client evicted?\n",
3288                  aa->oa_lockh, req, aa);
3289         ldlm_lock_decref(&handle, mode);
3290         LDLM_LOCK_PUT(lock);
3291         return rc;
3292 }
3293
3294 void osc_update_enqueue(struct lustre_handle *lov_lockhp,
3295                         struct lov_oinfo *loi, int flags,
3296                         struct ost_lvb *lvb, __u32 mode, int rc)
3297 {
3298         if (rc == ELDLM_OK) {
3299                 struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
3300                 __u64 tmp;
3301
3302                 LASSERT(lock != NULL);
3303                 loi->loi_lvb = *lvb;
3304                 tmp = loi->loi_lvb.lvb_size;
3305                 /* Extend KMS up to the end of this lock and no further
3306                  * A lock on [x,y] means a KMS of up to y + 1 bytes! */
3307                 if (tmp > lock->l_policy_data.l_extent.end)
3308                         tmp = lock->l_policy_data.l_extent.end + 1;
3309                 if (tmp >= loi->loi_kms) {
3310                         LDLM_DEBUG(lock, "lock acquired, setting rss="LPU64
3311                                    ", kms="LPU64, loi->loi_lvb.lvb_size, tmp);
3312                         loi_kms_set(loi, tmp);
3313                 } else {
3314                         LDLM_DEBUG(lock, "lock acquired, setting rss="
3315                                    LPU64"; leaving kms="LPU64", end="LPU64,
3316                                    loi->loi_lvb.lvb_size, loi->loi_kms,
3317                                    lock->l_policy_data.l_extent.end);
3318                 }
3319                 ldlm_lock_allow_match(lock);
3320                 LDLM_LOCK_PUT(lock);
3321         } else if (rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT)) {
3322                 loi->loi_lvb = *lvb;
3323                 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
3324                        " kms="LPU64"\n", loi->loi_lvb.lvb_size, loi->loi_kms);
3325                 rc = ELDLM_OK;
3326         }
3327 }
3328 EXPORT_SYMBOL(osc_update_enqueue);
3329
3330 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
3331
3332 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
3333  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
3334  * other synchronous requests, however keeping some locks and trying to obtain
3335  * others may take a considerable amount of time in a case of ost failure; and
3336  * when other sync requests do not get released lock from a client, the client
3337  * is excluded from the cluster -- such scenarious make the life difficult, so
3338  * release locks just after they are obtained. */
3339 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3340                      int *flags, ldlm_policy_data_t *policy,
3341                      struct ost_lvb *lvb, int kms_valid,
3342                      obd_enqueue_update_f upcall, void *cookie,
3343                      struct ldlm_enqueue_info *einfo,
3344                      struct lustre_handle *lockh,
3345                      struct ptlrpc_request_set *rqset, int async)
3346 {
3347         struct obd_device *obd = exp->exp_obd;
3348         struct ptlrpc_request *req = NULL;
3349         int intent = *flags & LDLM_FL_HAS_INTENT;
3350         ldlm_mode_t mode;
3351         int rc;
3352         ENTRY;
3353
3354         /* Filesystem lock extents are extended to page boundaries so that
3355          * dealing with the page cache is a little smoother.  */
3356         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3357         policy->l_extent.end |= ~CFS_PAGE_MASK;
3358
3359         /*
3360          * kms is not valid when either object is completely fresh (so that no
3361          * locks are cached), or object was evicted. In the latter case cached
3362          * lock cannot be used, because it would prime inode state with
3363          * potentially stale LVB.
3364          */
3365         if (!kms_valid)
3366                 goto no_match;
3367
3368         /* Next, search for already existing extent locks that will cover us */
3369         /* If we're trying to read, we also search for an existing PW lock.  The
3370          * VFS and page cache already protect us locally, so lots of readers/
3371          * writers can share a single PW lock.
3372          *
3373          * There are problems with conversion deadlocks, so instead of
3374          * converting a read lock to a write lock, we'll just enqueue a new
3375          * one.
3376          *
3377          * At some point we should cancel the read lock instead of making them
3378          * send us a blocking callback, but there are problems with canceling
3379          * locks out from other users right now, too. */
3380         mode = einfo->ei_mode;
3381         if (einfo->ei_mode == LCK_PR)
3382                 mode |= LCK_PW;
3383         mode = ldlm_lock_match(obd->obd_namespace,
3384                                *flags | LDLM_FL_LVB_READY, res_id,
3385                                einfo->ei_type, policy, mode, lockh, 0);
3386         if (mode) {
3387                 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
3388
3389                 if (osc_set_lock_data_with_check(matched, einfo)) {
3390                         /* addref the lock only if not async requests and PW
3391                          * lock is matched whereas we asked for PR. */
3392                         if (!rqset && einfo->ei_mode != mode)
3393                                 ldlm_lock_addref(lockh, LCK_PR);
3394                         if (intent) {
3395                                 /* I would like to be able to ASSERT here that
3396                                  * rss <= kms, but I can't, for reasons which
3397                                  * are explained in lov_enqueue() */
3398                         }
3399
3400                         /* We already have a lock, and it's referenced */
3401                         (*upcall)(cookie, ELDLM_OK);
3402
3403                         /* For async requests, decref the lock. */
3404                         if (einfo->ei_mode != mode)
3405                                 ldlm_lock_decref(lockh, LCK_PW);
3406                         else if (rqset)
3407                                 ldlm_lock_decref(lockh, einfo->ei_mode);
3408                         LDLM_LOCK_PUT(matched);
3409                         RETURN(ELDLM_OK);
3410                 } else
3411                         ldlm_lock_decref(lockh, mode);
3412                 LDLM_LOCK_PUT(matched);
3413         }
3414
3415  no_match:
3416         if (intent) {
3417                 CFS_LIST_HEAD(cancels);
3418                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3419                                            &RQF_LDLM_ENQUEUE_LVB);
3420                 if (req == NULL)
3421                         RETURN(-ENOMEM);
3422
3423                 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
3424                 if (rc) {
3425                         ptlrpc_request_free(req);
3426                         RETURN(rc);
3427                 }
3428
3429                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
3430                                      sizeof *lvb);
3431                 ptlrpc_request_set_replen(req);
3432         }
3433
3434         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
3435         *flags &= ~LDLM_FL_BLOCK_GRANTED;
3436
3437         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
3438                               sizeof(*lvb), lockh, async);
3439         if (rqset) {
3440                 if (!rc) {
3441                         struct osc_enqueue_args *aa;
3442                         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3443                         aa = ptlrpc_req_async_args(req);
3444                         aa->oa_ei = einfo;
3445                         aa->oa_exp = exp;
3446                         aa->oa_flags  = flags;
3447                         aa->oa_upcall = upcall;
3448                         aa->oa_cookie = cookie;
3449                         aa->oa_lvb    = lvb;
3450                         aa->oa_lockh  = lockh;
3451
3452                         req->rq_interpret_reply =
3453                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
3454                         if (rqset == PTLRPCD_SET)
3455                                 ptlrpcd_add_req(req, PSCOPE_OTHER);
3456                         else
3457                                 ptlrpc_set_add_req(rqset, req);
3458                 } else if (intent) {
3459                         ptlrpc_req_finished(req);
3460                 }
3461                 RETURN(rc);
3462         }
3463
3464         rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, rc);
3465         if (intent)
3466                 ptlrpc_req_finished(req);
3467
3468         RETURN(rc);
3469 }
3470
3471 static int osc_enqueue(struct obd_export *exp, struct obd_info *oinfo,
3472                        struct ldlm_enqueue_info *einfo,
3473                        struct ptlrpc_request_set *rqset)
3474 {
3475         struct ldlm_res_id res_id;
3476         int rc;
3477         ENTRY;
3478
3479         osc_build_res_name(oinfo->oi_md->lsm_object_id,
3480                            oinfo->oi_md->lsm_object_seq, &res_id);
3481
3482         rc = osc_enqueue_base(exp, &res_id, &oinfo->oi_flags, &oinfo->oi_policy,
3483                               &oinfo->oi_md->lsm_oinfo[0]->loi_lvb,
3484                               oinfo->oi_md->lsm_oinfo[0]->loi_kms_valid,
3485                               oinfo->oi_cb_up, oinfo, einfo, oinfo->oi_lockh,
3486                               rqset, rqset != NULL);
3487         RETURN(rc);
3488 }
3489
3490 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
3491                    __u32 type, ldlm_policy_data_t *policy, __u32 mode,
3492                    int *flags, void *data, struct lustre_handle *lockh,
3493                    int unref)
3494 {
3495         struct obd_device *obd = exp->exp_obd;
3496         int lflags = *flags;
3497         ldlm_mode_t rc;
3498         ENTRY;
3499
3500         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
3501                 RETURN(-EIO);
3502
3503         /* Filesystem lock extents are extended to page boundaries so that
3504          * dealing with the page cache is a little smoother */
3505         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
3506         policy->l_extent.end |= ~CFS_PAGE_MASK;
3507
3508         /* Next, search for already existing extent locks that will cover us */
3509         /* If we're trying to read, we also search for an existing PW lock.  The
3510          * VFS and page cache already protect us locally, so lots of readers/
3511          * writers can share a single PW lock. */
3512         rc = mode;
3513         if (mode == LCK_PR)
3514                 rc |= LCK_PW;
3515         rc = ldlm_lock_match(obd->obd_namespace, lflags,
3516                              res_id, type, policy, rc, lockh, unref);
3517         if (rc) {
3518                 if (data != NULL) {
3519                         if (!osc_set_data_with_check(lockh, data)) {
3520                                 if (!(lflags & LDLM_FL_TEST_LOCK))
3521                                         ldlm_lock_decref(lockh, rc);
3522                                 RETURN(0);
3523                         }
3524                 }
3525                 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
3526                         ldlm_lock_addref(lockh, LCK_PR);
3527                         ldlm_lock_decref(lockh, LCK_PW);
3528                 }
3529                 RETURN(rc);
3530         }
3531         RETURN(rc);
3532 }
3533
3534 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
3535 {
3536         ENTRY;
3537
3538         if (unlikely(mode == LCK_GROUP))
3539                 ldlm_lock_decref_and_cancel(lockh, mode);
3540         else
3541                 ldlm_lock_decref(lockh, mode);
3542
3543         RETURN(0);
3544 }
3545
3546 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
3547                       __u32 mode, struct lustre_handle *lockh)
3548 {
3549         ENTRY;
3550         RETURN(osc_cancel_base(lockh, mode));
3551 }
3552
3553 static int osc_cancel_unused(struct obd_export *exp,
3554                              struct lov_stripe_md *lsm,
3555                              ldlm_cancel_flags_t flags,
3556                              void *opaque)
3557 {
3558         struct obd_device *obd = class_exp2obd(exp);
3559         struct ldlm_res_id res_id, *resp = NULL;
3560
3561         if (lsm != NULL) {
3562                 resp = osc_build_res_name(lsm->lsm_object_id,
3563                                           lsm->lsm_object_seq, &res_id);
3564         }
3565
3566         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
3567 }
3568
3569 static int osc_statfs_interpret(const struct lu_env *env,
3570                                 struct ptlrpc_request *req,
3571                                 struct osc_async_args *aa, int rc)
3572 {
3573         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
3574         struct obd_statfs *msfs;
3575         __u64 used;
3576         ENTRY;
3577
3578         if (rc == -EBADR)
3579                 /* The request has in fact never been sent
3580                  * due to issues at a higher level (LOV).
3581                  * Exit immediately since the caller is
3582                  * aware of the problem and takes care
3583                  * of the clean up */
3584                  RETURN(rc);
3585
3586         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3587             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3588                 GOTO(out, rc = 0);
3589
3590         if (rc != 0)
3591                 GOTO(out, rc);
3592
3593         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3594         if (msfs == NULL) {
3595                 GOTO(out, rc = -EPROTO);
3596         }
3597
3598         /* Reinitialize the RDONLY and DEGRADED flags at the client
3599          * on each statfs, so they don't stay set permanently. */
3600         cfs_spin_lock(&cli->cl_oscc.oscc_lock);
3601
3602         if (unlikely(msfs->os_state & OS_STATE_DEGRADED))
3603                 cli->cl_oscc.oscc_flags |= OSCC_FLAG_DEGRADED;
3604         else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_DEGRADED))
3605                 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_DEGRADED;
3606
3607         if (unlikely(msfs->os_state & OS_STATE_READONLY))
3608                 cli->cl_oscc.oscc_flags |= OSCC_FLAG_RDONLY;
3609         else if (unlikely(cli->cl_oscc.oscc_flags & OSCC_FLAG_RDONLY))
3610                 cli->cl_oscc.oscc_flags &= ~OSCC_FLAG_RDONLY;
3611
3612         /* Add a bit of hysteresis so this flag isn't continually flapping,
3613          * and ensure that new files don't get extremely fragmented due to
3614          * only a small amount of available space in the filesystem.
3615          * We want to set the NOSPC flag when there is less than ~0.1% free
3616          * and clear it when there is at least ~0.2% free space, so:
3617          *                   avail < ~0.1% max          max = avail + used
3618          *            1025 * avail < avail + used       used = blocks - free
3619          *            1024 * avail < used
3620          *            1024 * avail < blocks - free
3621          *                   avail < ((blocks - free) >> 10)
3622          *
3623          * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
3624          * lose that amount of space so in those cases we report no space left
3625          * if their is less than 1 GB left.                             */
3626         used = min_t(__u64,(msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30);
3627         if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) == 0) &&
3628                      ((msfs->os_ffree < 32) || (msfs->os_bavail < used))))
3629                 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC;
3630         else if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3631                           (msfs->os_ffree > 64) &&
3632                           (msfs->os_bavail > (used << 1)))) {
3633                 cli->cl_oscc.oscc_flags &= ~(OSCC_FLAG_NOSPC |
3634                                              OSCC_FLAG_NOSPC_BLK);
3635         }
3636
3637         if (unlikely(((cli->cl_oscc.oscc_flags & OSCC_FLAG_NOSPC) != 0) &&
3638                      (msfs->os_bavail < used)))
3639                 cli->cl_oscc.oscc_flags |= OSCC_FLAG_NOSPC_BLK;
3640
3641         cfs_spin_unlock(&cli->cl_oscc.oscc_lock);
3642
3643         *aa->aa_oi->oi_osfs = *msfs;
3644 out:
3645         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3646         RETURN(rc);
3647 }
3648
3649 static int osc_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
3650                             __u64 max_age, struct ptlrpc_request_set *rqset)
3651 {
3652         struct ptlrpc_request *req;
3653         struct osc_async_args *aa;
3654         int                    rc;
3655         ENTRY;
3656
3657         /* We could possibly pass max_age in the request (as an absolute
3658          * timestamp or a "seconds.usec ago") so the target can avoid doing
3659          * extra calls into the filesystem if that isn't necessary (e.g.
3660          * during mount that would help a bit).  Having relative timestamps
3661          * is not so great if request processing is slow, while absolute
3662          * timestamps are not ideal because they need time synchronization. */
3663         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3664         if (req == NULL)
3665                 RETURN(-ENOMEM);
3666
3667         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3668         if (rc) {
3669                 ptlrpc_request_free(req);
3670                 RETURN(rc);
3671         }
3672         ptlrpc_request_set_replen(req);
3673         req->rq_request_portal = OST_CREATE_PORTAL;
3674         ptlrpc_at_set_req_timeout(req);
3675
3676         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3677                 /* procfs requests not want stat in wait for avoid deadlock */
3678                 req->rq_no_resend = 1;
3679                 req->rq_no_delay = 1;
3680         }
3681
3682         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
3683         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
3684         aa = ptlrpc_req_async_args(req);
3685         aa->aa_oi = oinfo;
3686
3687         ptlrpc_set_add_req(rqset, req);
3688         RETURN(0);
3689 }
3690
3691 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
3692                       __u64 max_age, __u32 flags)
3693 {
3694         struct obd_statfs     *msfs;
3695         struct ptlrpc_request *req;
3696         struct obd_import     *imp = NULL;
3697         int rc;
3698         ENTRY;
3699
3700         /*Since the request might also come from lprocfs, so we need
3701          *sync this with client_disconnect_export Bug15684*/
3702         cfs_down_read(&obd->u.cli.cl_sem);
3703         if (obd->u.cli.cl_import)
3704                 imp = class_import_get(obd->u.cli.cl_import);
3705         cfs_up_read(&obd->u.cli.cl_sem);
3706         if (!imp)
3707                 RETURN(-ENODEV);
3708
3709         /* We could possibly pass max_age in the request (as an absolute
3710          * timestamp or a "seconds.usec ago") so the target can avoid doing
3711          * extra calls into the filesystem if that isn't necessary (e.g.
3712          * during mount that would help a bit).  Having relative timestamps
3713          * is not so great if request processing is slow, while absolute
3714          * timestamps are not ideal because they need time synchronization. */
3715         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3716
3717         class_import_put(imp);
3718
3719         if (req == NULL)
3720                 RETURN(-ENOMEM);
3721
3722         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3723         if (rc) {
3724                 ptlrpc_request_free(req);
3725                 RETURN(rc);
3726         }
3727         ptlrpc_request_set_replen(req);
3728         req->rq_request_portal = OST_CREATE_PORTAL;
3729         ptlrpc_at_set_req_timeout(req);
3730
3731         if (flags & OBD_STATFS_NODELAY) {
3732                 /* procfs requests not want stat in wait for avoid deadlock */
3733                 req->rq_no_resend = 1;
3734                 req->rq_no_delay = 1;
3735         }
3736
3737         rc = ptlrpc_queue_wait(req);
3738         if (rc)
3739                 GOTO(out, rc);
3740
3741         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3742         if (msfs == NULL) {
3743                 GOTO(out, rc = -EPROTO);
3744         }
3745
3746         *osfs = *msfs;
3747
3748         EXIT;
3749  out:
3750         ptlrpc_req_finished(req);
3751         return rc;
3752 }
3753
3754 /* Retrieve object striping information.
3755  *
3756  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
3757  * the maximum number of OST indices which will fit in the user buffer.
3758  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
3759  */
3760 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
3761 {
3762         /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
3763         struct lov_user_md_v3 lum, *lumk;
3764         struct lov_user_ost_data_v1 *lmm_objects;
3765         int rc = 0, lum_size;
3766         ENTRY;
3767
3768         if (!lsm)
3769                 RETURN(-ENODATA);
3770
3771         /* we only need the header part from user space to get lmm_magic and
3772          * lmm_stripe_count, (the header part is common to v1 and v3) */
3773         lum_size = sizeof(struct lov_user_md_v1);
3774         if (cfs_copy_from_user(&lum, lump, lum_size))
3775                 RETURN(-EFAULT);
3776
3777         if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
3778             (lum.lmm_magic != LOV_USER_MAGIC_V3))
3779                 RETURN(-EINVAL);
3780
3781         /* lov_user_md_vX and lov_mds_md_vX must have the same size */
3782         LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
3783         LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
3784         LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
3785
3786         /* we can use lov_mds_md_size() to compute lum_size
3787          * because lov_user_md_vX and lov_mds_md_vX have the same size */
3788         if (lum.lmm_stripe_count > 0) {
3789                 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
3790                 OBD_ALLOC(lumk, lum_size);
3791                 if (!lumk)
3792                         RETURN(-ENOMEM);
3793
3794                 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
3795                         lmm_objects = &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
3796                 else
3797                         lmm_objects = &(lumk->lmm_objects[0]);
3798                 lmm_objects->l_object_id = lsm->lsm_object_id;
3799         } else {
3800                 lum_size = lov_mds_md_size(0, lum.lmm_magic);
3801                 lumk = &lum;
3802         }
3803
3804         lumk->lmm_object_id = lsm->lsm_object_id;
3805         lumk->lmm_object_seq = lsm->lsm_object_seq;
3806         lumk->lmm_stripe_count = 1;
3807
3808         if (cfs_copy_to_user(lump, lumk, lum_size))
3809                 rc = -EFAULT;
3810
3811         if (lumk != &lum)
3812                 OBD_FREE(lumk, lum_size);
3813
3814         RETURN(rc);
3815 }
3816
3817
3818 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3819                          void *karg, void *uarg)
3820 {
3821         struct obd_device *obd = exp->exp_obd;
3822         struct obd_ioctl_data *data = karg;
3823         int err = 0;
3824         ENTRY;
3825
3826         if (!cfs_try_module_get(THIS_MODULE)) {
3827                 CERROR("Can't get module. Is it alive?");
3828                 return -EINVAL;
3829         }
3830         switch (cmd) {
3831         case OBD_IOC_LOV_GET_CONFIG: {
3832                 char *buf;
3833                 struct lov_desc *desc;
3834                 struct obd_uuid uuid;
3835
3836                 buf = NULL;
3837                 len = 0;
3838                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
3839                         GOTO(out, err = -EINVAL);
3840
3841                 data = (struct obd_ioctl_data *)buf;
3842
3843                 if (sizeof(*desc) > data->ioc_inllen1) {
3844                         obd_ioctl_freedata(buf, len);
3845                         GOTO(out, err = -EINVAL);
3846                 }
3847
3848                 if (data->ioc_inllen2 < sizeof(uuid)) {
3849                         obd_ioctl_freedata(buf, len);
3850                         GOTO(out, err = -EINVAL);
3851                 }
3852
3853                 desc = (struct lov_desc *)data->ioc_inlbuf1;
3854                 desc->ld_tgt_count = 1;
3855                 desc->ld_active_tgt_count = 1;
3856                 desc->ld_default_stripe_count = 1;
3857                 desc->ld_default_stripe_size = 0;
3858                 desc->ld_default_stripe_offset = 0;
3859                 desc->ld_pattern = 0;
3860                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
3861
3862                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
3863
3864                 err = cfs_copy_to_user((void *)uarg, buf, len);
3865                 if (err)
3866                         err = -EFAULT;
3867                 obd_ioctl_freedata(buf, len);
3868                 GOTO(out, err);
3869         }
3870         case LL_IOC_LOV_SETSTRIPE:
3871                 err = obd_alloc_memmd(exp, karg);
3872                 if (err > 0)
3873                         err = 0;
3874                 GOTO(out, err);
3875         case LL_IOC_LOV_GETSTRIPE:
3876                 err = osc_getstripe(karg, uarg);
3877                 GOTO(out, err);
3878         case OBD_IOC_CLIENT_RECOVER:
3879                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
3880                                             data->ioc_inlbuf1, 0);
3881                 if (err > 0)
3882                         err = 0;
3883                 GOTO(out, err);
3884         case IOC_OSC_SET_ACTIVE:
3885                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
3886                                                data->ioc_offset);
3887                 GOTO(out, err);
3888         case OBD_IOC_POLL_QUOTACHECK:
3889                 err = lquota_poll_check(quota_interface, exp,
3890                                         (struct if_quotacheck *)karg);
3891                 GOTO(out, err);
3892         case OBD_IOC_PING_TARGET:
3893                 err = ptlrpc_obd_ping(obd);
3894                 GOTO(out, err);
3895         default:
3896                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
3897                        cmd, cfs_curproc_comm());
3898                 GOTO(out, err = -ENOTTY);
3899         }
3900 out:
3901         cfs_module_put(THIS_MODULE);
3902         return err;
3903 }
3904
3905 static int osc_get_info(struct obd_export *exp, obd_count keylen,
3906                         void *key, __u32 *vallen, void *val,
3907                         struct lov_stripe_md *lsm)
3908 {
3909         ENTRY;
3910         if (!vallen || !val)
3911                 RETURN(-EFAULT);
3912
3913         if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
3914                 __u32 *stripe = val;
3915                 *vallen = sizeof(*stripe);
3916                 *stripe = 0;
3917                 RETURN(0);
3918         } else if (KEY_IS(KEY_LAST_ID)) {
3919                 struct ptlrpc_request *req;
3920                 obd_id                *reply;
3921                 char                  *tmp;
3922                 int                    rc;
3923
3924                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3925                                            &RQF_OST_GET_INFO_LAST_ID);
3926                 if (req == NULL)
3927                         RETURN(-ENOMEM);
3928
3929                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3930                                      RCL_CLIENT, keylen);
3931                 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3932                 if (rc) {
3933                         ptlrpc_request_free(req);
3934                         RETURN(rc);
3935                 }
3936
3937                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3938                 memcpy(tmp, key, keylen);
3939
3940                 req->rq_no_delay = req->rq_no_resend = 1;
3941                 ptlrpc_request_set_replen(req);
3942                 rc = ptlrpc_queue_wait(req);
3943                 if (rc)
3944                         GOTO(out, rc);
3945
3946                 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
3947                 if (reply == NULL)
3948                         GOTO(out, rc = -EPROTO);
3949
3950                 *((obd_id *)val) = *reply;
3951         out:
3952                 ptlrpc_req_finished(req);
3953                 RETURN(rc);
3954         } else if (KEY_IS(KEY_FIEMAP)) {
3955                 struct ptlrpc_request *req;
3956                 struct ll_user_fiemap *reply;
3957                 char *tmp;
3958                 int rc;
3959
3960                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
3961                                            &RQF_OST_GET_INFO_FIEMAP);
3962                 if (req == NULL)
3963                         RETURN(-ENOMEM);
3964
3965                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
3966                                      RCL_CLIENT, keylen);
3967                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3968                                      RCL_CLIENT, *vallen);
3969                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
3970                                      RCL_SERVER, *vallen);
3971
3972                 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
3973                 if (rc) {
3974                         ptlrpc_request_free(req);
3975                         RETURN(rc);
3976                 }
3977
3978                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
3979                 memcpy(tmp, key, keylen);
3980                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3981                 memcpy(tmp, val, *vallen);
3982
3983                 ptlrpc_request_set_replen(req);
3984                 rc = ptlrpc_queue_wait(req);
3985                 if (rc)
3986                         GOTO(out1, rc);
3987
3988                 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
3989                 if (reply == NULL)
3990                         GOTO(out1, rc = -EPROTO);
3991
3992                 memcpy(val, reply, *vallen);
3993         out1:
3994                 ptlrpc_req_finished(req);
3995
3996                 RETURN(rc);
3997         }
3998
3999         RETURN(-EINVAL);
4000 }
4001
4002 static int osc_setinfo_mds_connect_import(struct obd_import *imp)
4003 {
4004         struct llog_ctxt *ctxt;
4005         int rc = 0;
4006         ENTRY;
4007
4008         ctxt = llog_get_context(imp->imp_obd, LLOG_MDS_OST_ORIG_CTXT);
4009         if (ctxt) {
4010                 rc = llog_initiator_connect(ctxt);
4011                 llog_ctxt_put(ctxt);
4012         } else {
4013                 /* XXX return an error? skip setting below flags? */
4014         }
4015
4016         cfs_spin_lock(&imp->imp_lock);
4017         imp->imp_server_timeout = 1;
4018         imp->imp_pingable = 1;
4019         cfs_spin_unlock(&imp->imp_lock);
4020         CDEBUG(D_RPCTRACE, "pinging OST %s\n", obd2cli_tgt(imp->imp_obd));
4021
4022         RETURN(rc);
4023 }
4024
4025 static int osc_setinfo_mds_conn_interpret(const struct lu_env *env,
4026                                           struct ptlrpc_request *req,
4027                                           void *aa, int rc)
4028 {
4029         ENTRY;
4030         if (rc != 0)
4031                 RETURN(rc);
4032
4033         RETURN(osc_setinfo_mds_connect_import(req->rq_import));
4034 }
4035
4036 static int osc_set_info_async(struct obd_export *exp, obd_count keylen,
4037                               void *key, obd_count vallen, void *val,
4038                               struct ptlrpc_request_set *set)
4039 {
4040         struct ptlrpc_request *req;
4041         struct obd_device     *obd = exp->exp_obd;
4042         struct obd_import     *imp = class_exp2cliimp(exp);
4043         char                  *tmp;
4044         int                    rc;
4045         ENTRY;
4046
4047         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
4048
4049         if (KEY_IS(KEY_NEXT_ID)) {
4050                 obd_id new_val;
4051                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4052
4053                 if (vallen != sizeof(obd_id))
4054                         RETURN(-ERANGE);
4055                 if (val == NULL)
4056                         RETURN(-EINVAL);
4057
4058                 if (vallen != sizeof(obd_id))
4059                         RETURN(-EINVAL);
4060
4061                 /* avoid race between allocate new object and set next id
4062                  * from ll_sync thread */
4063                 cfs_spin_lock(&oscc->oscc_lock);
4064                 new_val = *((obd_id*)val) + 1;
4065                 if (new_val > oscc->oscc_next_id)
4066                         oscc->oscc_next_id = new_val;
4067                 cfs_spin_unlock(&oscc->oscc_lock);
4068                 CDEBUG(D_HA, "%s: set oscc_next_id = "LPU64"\n",
4069                        exp->exp_obd->obd_name,
4070                        obd->u.cli.cl_oscc.oscc_next_id);
4071
4072                 RETURN(0);
4073         }
4074
4075         if (KEY_IS(KEY_CHECKSUM)) {
4076                 if (vallen != sizeof(int))
4077                         RETURN(-EINVAL);
4078                 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
4079                 RETURN(0);
4080         }
4081
4082         if (KEY_IS(KEY_SPTLRPC_CONF)) {
4083                 sptlrpc_conf_client_adapt(obd);
4084                 RETURN(0);
4085         }
4086
4087         if (KEY_IS(KEY_FLUSH_CTX)) {
4088                 sptlrpc_import_flush_my_ctx(imp);
4089                 RETURN(0);
4090         }
4091
4092         if (!set && !KEY_IS(KEY_GRANT_SHRINK))
4093                 RETURN(-EINVAL);
4094
4095         /* We pass all other commands directly to OST. Since nobody calls osc
4096            methods directly and everybody is supposed to go through LOV, we
4097            assume lov checked invalid values for us.
4098            The only recognised values so far are evict_by_nid and mds_conn.
4099            Even if something bad goes through, we'd get a -EINVAL from OST
4100            anyway. */
4101
4102         if (KEY_IS(KEY_GRANT_SHRINK))
4103                 req = ptlrpc_request_alloc(imp, &RQF_OST_SET_GRANT_INFO);
4104         else
4105                 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
4106
4107         if (req == NULL)
4108                 RETURN(-ENOMEM);
4109
4110         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
4111                              RCL_CLIENT, keylen);
4112         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
4113                              RCL_CLIENT, vallen);
4114         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
4115         if (rc) {
4116                 ptlrpc_request_free(req);
4117                 RETURN(rc);
4118         }
4119
4120         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
4121         memcpy(tmp, key, keylen);
4122         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
4123         memcpy(tmp, val, vallen);
4124
4125         if (KEY_IS(KEY_MDS_CONN)) {
4126                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4127
4128                 oscc->oscc_oa.o_seq = (*(__u32 *)val);
4129                 oscc->oscc_oa.o_valid |= OBD_MD_FLGROUP;
4130                 LASSERT_SEQ_IS_MDT(oscc->oscc_oa.o_seq);
4131                 req->rq_no_delay = req->rq_no_resend = 1;
4132                 req->rq_interpret_reply = osc_setinfo_mds_conn_interpret;
4133         } else if (KEY_IS(KEY_GRANT_SHRINK)) {
4134                 struct osc_grant_args *aa;
4135                 struct obdo *oa;
4136
4137                 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
4138                 aa = ptlrpc_req_async_args(req);
4139                 OBDO_ALLOC(oa);
4140                 if (!oa) {
4141                         ptlrpc_req_finished(req);
4142                         RETURN(-ENOMEM);
4143                 }
4144                 *oa = ((struct ost_body *)val)->oa;
4145                 aa->aa_oa = oa;
4146                 req->rq_interpret_reply = osc_shrink_grant_interpret;
4147         }
4148
4149         ptlrpc_request_set_replen(req);
4150         if (!KEY_IS(KEY_GRANT_SHRINK)) {
4151                 LASSERT(set != NULL);
4152                 ptlrpc_set_add_req(set, req);
4153                 ptlrpc_check_set(NULL, set);
4154         } else
4155                 ptlrpcd_add_req(req, PSCOPE_OTHER);
4156
4157         RETURN(0);
4158 }
4159
4160
4161 static struct llog_operations osc_size_repl_logops = {
4162         lop_cancel: llog_obd_repl_cancel
4163 };
4164
4165 static struct llog_operations osc_mds_ost_orig_logops;
4166
4167 static int __osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4168                            struct obd_device *tgt, struct llog_catid *catid)
4169 {
4170         int rc;
4171         ENTRY;
4172
4173         rc = llog_setup(obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, tgt, 1,
4174                         &catid->lci_logid, &osc_mds_ost_orig_logops);
4175         if (rc) {
4176                 CERROR("failed LLOG_MDS_OST_ORIG_CTXT\n");
4177                 GOTO(out, rc);
4178         }
4179
4180         rc = llog_setup(obd, &obd->obd_olg, LLOG_SIZE_REPL_CTXT, tgt, 1,
4181                         NULL, &osc_size_repl_logops);
4182         if (rc) {
4183                 struct llog_ctxt *ctxt =
4184                         llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4185                 if (ctxt)
4186                         llog_cleanup(ctxt);
4187                 CERROR("failed LLOG_SIZE_REPL_CTXT\n");
4188         }
4189         GOTO(out, rc);
4190 out:
4191         if (rc) {
4192                 CERROR("osc '%s' tgt '%s' catid %p rc=%d\n",
4193                        obd->obd_name, tgt->obd_name, catid, rc);
4194                 CERROR("logid "LPX64":0x%x\n",
4195                        catid->lci_logid.lgl_oid, catid->lci_logid.lgl_ogen);
4196         }
4197         return rc;
4198 }
4199
4200 static int osc_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
4201                          struct obd_device *disk_obd, int *index)
4202 {
4203         struct llog_catid catid;
4204         static char name[32] = CATLIST;
4205         int rc;
4206         ENTRY;
4207
4208         LASSERT(olg == &obd->obd_olg);
4209
4210         cfs_mutex_down(&olg->olg_cat_processing);
4211         rc = llog_get_cat_list(disk_obd, name, *index, 1, &catid);
4212         if (rc) {
4213                 CERROR("rc: %d\n", rc);
4214                 GOTO(out, rc);
4215         }
4216
4217         CDEBUG(D_INFO, "%s: Init llog for %d - catid "LPX64"/"LPX64":%x\n",
4218                obd->obd_name, *index, catid.lci_logid.lgl_oid,
4219                catid.lci_logid.lgl_oseq, catid.lci_logid.lgl_ogen);
4220
4221         rc = __osc_llog_init(obd, olg, disk_obd, &catid);
4222         if (rc) {
4223                 CERROR("rc: %d\n", rc);
4224                 GOTO(out, rc);
4225         }
4226
4227         rc = llog_put_cat_list(disk_obd, name, *index, 1, &catid);
4228         if (rc) {
4229                 CERROR("rc: %d\n", rc);
4230                 GOTO(out, rc);
4231         }
4232
4233  out:
4234         cfs_mutex_up(&olg->olg_cat_processing);
4235
4236         return rc;
4237 }
4238
4239 static int osc_llog_finish(struct obd_device *obd, int count)
4240 {
4241         struct llog_ctxt *ctxt;
4242         int rc = 0, rc2 = 0;
4243         ENTRY;
4244
4245         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
4246         if (ctxt)
4247                 rc = llog_cleanup(ctxt);
4248
4249         ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4250         if (ctxt)
4251                 rc2 = llog_cleanup(ctxt);
4252         if (!rc)
4253                 rc = rc2;
4254
4255         RETURN(rc);
4256 }
4257
4258 static int osc_reconnect(const struct lu_env *env,
4259                          struct obd_export *exp, struct obd_device *obd,
4260                          struct obd_uuid *cluuid,
4261                          struct obd_connect_data *data,
4262                          void *localdata)
4263 {
4264         struct client_obd *cli = &obd->u.cli;
4265
4266         if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
4267                 long lost_grant;
4268
4269                 client_obd_list_lock(&cli->cl_loi_list_lock);
4270                 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
4271                                 2 * cli->cl_max_pages_per_rpc << CFS_PAGE_SHIFT;
4272                 lost_grant = cli->cl_lost_grant;
4273                 cli->cl_lost_grant = 0;
4274                 client_obd_list_unlock(&cli->cl_loi_list_lock);
4275
4276                 CDEBUG(D_CACHE, "request ocd_grant: %d cl_avail_grant: %ld "
4277                        "cl_dirty: %ld cl_lost_grant: %ld\n", data->ocd_grant,
4278                        cli->cl_avail_grant, cli->cl_dirty, lost_grant);
4279                 CDEBUG(D_RPCTRACE, "ocd_connect_flags: "LPX64" ocd_version: %d"
4280                        " ocd_grant: %d\n", data->ocd_connect_flags,
4281                        data->ocd_version, data->ocd_grant);
4282         }
4283
4284         RETURN(0);
4285 }
4286
4287 static int osc_disconnect(struct obd_export *exp)
4288 {
4289         struct obd_device *obd = class_exp2obd(exp);
4290         struct llog_ctxt  *ctxt;
4291         int rc;
4292
4293         ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
4294         if (ctxt) {
4295                 if (obd->u.cli.cl_conn_count == 1) {
4296                         /* Flush any remaining cancel messages out to the
4297                          * target */
4298                         llog_sync(ctxt, exp);
4299                 }
4300                 llog_ctxt_put(ctxt);
4301         } else {
4302                 CDEBUG(D_HA, "No LLOG_SIZE_REPL_CTXT found in obd %p\n",
4303                        obd);
4304         }
4305
4306         rc = client_disconnect_export(exp);
4307         /**
4308          * Initially we put del_shrink_grant before disconnect_export, but it
4309          * causes the following problem if setup (connect) and cleanup
4310          * (disconnect) are tangled together.
4311          *      connect p1                     disconnect p2
4312          *   ptlrpc_connect_import
4313          *     ...............               class_manual_cleanup
4314          *                                     osc_disconnect
4315          *                                     del_shrink_grant
4316          *   ptlrpc_connect_interrupt
4317          *     init_grant_shrink
4318          *   add this client to shrink list
4319          *                                      cleanup_osc
4320          * Bang! pinger trigger the shrink.
4321          * So the osc should be disconnected from the shrink list, after we
4322          * are sure the import has been destroyed. BUG18662
4323          */
4324         if (obd->u.cli.cl_import == NULL)
4325                 osc_del_shrink_grant(&obd->u.cli);
4326         return rc;
4327 }
4328
4329 static int osc_import_event(struct obd_device *obd,
4330                             struct obd_import *imp,
4331                             enum obd_import_event event)
4332 {
4333         struct client_obd *cli;
4334         int rc = 0;
4335
4336         ENTRY;
4337         LASSERT(imp->imp_obd == obd);
4338
4339         switch (event) {
4340         case IMP_EVENT_DISCON: {
4341                 /* Only do this on the MDS OSC's */
4342                 if (imp->imp_server_timeout) {
4343                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4344
4345                         cfs_spin_lock(&oscc->oscc_lock);
4346                         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
4347                         cfs_spin_unlock(&oscc->oscc_lock);
4348                 }
4349                 cli = &obd->u.cli;
4350                 client_obd_list_lock(&cli->cl_loi_list_lock);
4351                 cli->cl_avail_grant = 0;
4352                 cli->cl_lost_grant = 0;
4353                 client_obd_list_unlock(&cli->cl_loi_list_lock);
4354                 break;
4355         }
4356         case IMP_EVENT_INACTIVE: {
4357                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
4358                 break;
4359         }
4360         case IMP_EVENT_INVALIDATE: {
4361                 struct ldlm_namespace *ns = obd->obd_namespace;
4362                 struct lu_env         *env;
4363                 int                    refcheck;
4364
4365                 env = cl_env_get(&refcheck);
4366                 if (!IS_ERR(env)) {
4367                         /* Reset grants */
4368                         cli = &obd->u.cli;
4369                         client_obd_list_lock(&cli->cl_loi_list_lock);
4370                         /* all pages go to failing rpcs due to the invalid
4371                          * import */
4372                         osc_check_rpcs(env, cli);
4373                         client_obd_list_unlock(&cli->cl_loi_list_lock);
4374
4375                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
4376                         cl_env_put(env, &refcheck);
4377                 } else
4378                         rc = PTR_ERR(env);
4379                 break;
4380         }
4381         case IMP_EVENT_ACTIVE: {
4382                 /* Only do this on the MDS OSC's */
4383                 if (imp->imp_server_timeout) {
4384                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
4385
4386                         cfs_spin_lock(&oscc->oscc_lock);
4387                         oscc->oscc_flags &= ~(OSCC_FLAG_NOSPC |
4388                                               OSCC_FLAG_NOSPC_BLK);
4389                         cfs_spin_unlock(&oscc->oscc_lock);
4390                 }
4391                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
4392                 break;
4393         }
4394         case IMP_EVENT_OCD: {
4395                 struct obd_connect_data *ocd = &imp->imp_connect_data;
4396
4397                 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
4398                         osc_init_grant(&obd->u.cli, ocd);
4399
4400                 /* See bug 7198 */
4401                 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
4402                         imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
4403
4404                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
4405                 break;
4406         }
4407         case IMP_EVENT_DEACTIVATE: {
4408                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
4409                 break;
4410         }
4411         case IMP_EVENT_ACTIVATE: {
4412                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
4413                 break;
4414         }
4415         default:
4416                 CERROR("Unknown import event %d\n", event);
4417                 LBUG();
4418         }
4419         RETURN(rc);
4420 }
4421
4422 /**
4423  * Determine whether the lock can be canceled before replaying the lock
4424  * during recovery, see bug16774 for detailed information.
4425  *
4426  * \retval zero the lock can't be canceled
4427  * \retval other ok to cancel
4428  */
4429 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
4430 {
4431         check_res_locked(lock->l_resource);
4432
4433         /*
4434          * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
4435          *
4436          * XXX as a future improvement, we can also cancel unused write lock
4437          * if it doesn't have dirty data and active mmaps.
4438          */
4439         if (lock->l_resource->lr_type == LDLM_EXTENT &&
4440             (lock->l_granted_mode == LCK_PR ||
4441              lock->l_granted_mode == LCK_CR) &&
4442             (osc_dlm_lock_pageref(lock) == 0))
4443                 RETURN(1);
4444
4445         RETURN(0);
4446 }
4447
4448 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
4449 {
4450         int rc;
4451         ENTRY;
4452
4453         ENTRY;
4454         rc = ptlrpcd_addref();
4455         if (rc)
4456                 RETURN(rc);
4457
4458         rc = client_obd_setup(obd, lcfg);
4459         if (rc) {
4460                 ptlrpcd_decref();
4461         } else {
4462                 struct lprocfs_static_vars lvars = { 0 };
4463                 struct client_obd *cli = &obd->u.cli;
4464
4465                 cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
4466                 lprocfs_osc_init_vars(&lvars);
4467                 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0) {
4468                         lproc_osc_attach_seqstat(obd);
4469                         sptlrpc_lprocfs_cliobd_attach(obd);
4470                         ptlrpc_lprocfs_register_obd(obd);
4471                 }
4472
4473                 oscc_init(obd);
4474                 /* We need to allocate a few requests more, because
4475                    brw_interpret tries to create new requests before freeing
4476                    previous ones. Ideally we want to have 2x max_rpcs_in_flight
4477                    reserved, but I afraid that might be too much wasted RAM
4478                    in fact, so 2 is just my guess and still should work. */
4479                 cli->cl_import->imp_rq_pool =
4480                         ptlrpc_init_rq_pool(cli->cl_max_rpcs_in_flight + 2,
4481                                             OST_MAXREQSIZE,
4482                                             ptlrpc_add_rqs_to_pool);
4483
4484                 CFS_INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
4485                 cfs_sema_init(&cli->cl_grant_sem, 1);
4486
4487                 ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
4488         }
4489
4490         RETURN(rc);
4491 }
4492
4493 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
4494 {
4495         int rc = 0;
4496         ENTRY;
4497
4498         switch (stage) {
4499         case OBD_CLEANUP_EARLY: {
4500                 struct obd_import *imp;
4501                 imp = obd->u.cli.cl_import;
4502                 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
4503                 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
4504                 ptlrpc_deactivate_import(imp);
4505                 cfs_spin_lock(&imp->imp_lock);
4506                 imp->imp_pingable = 0;
4507                 cfs_spin_unlock(&imp->imp_lock);
4508                 break;
4509         }
4510         case OBD_CLEANUP_EXPORTS: {
4511                 /* LU-464
4512                  * for echo client, export may be on zombie list, wait for
4513                  * zombie thread to cull it, because cli.cl_import will be
4514                  * cleared in client_disconnect_export():
4515                  *   class_export_destroy() -> obd_cleanup() ->
4516                  *   echo_device_free() -> echo_client_cleanup() ->
4517                  *   obd_disconnect() -> osc_disconnect() ->
4518                  *   client_disconnect_export()
4519                  */
4520                 obd_zombie_barrier();
4521                 obd_cleanup_client_import(obd);
4522                 rc = obd_llog_finish(obd, 0);
4523                 if (rc != 0)
4524                         CERROR("failed to cleanup llogging subsystems\n");
4525                 break;
4526                 }
4527         }
4528         RETURN(rc);
4529 }
4530
4531 int osc_cleanup(struct obd_device *obd)
4532 {
4533         int rc;
4534
4535         ENTRY;
4536         ptlrpc_lprocfs_unregister_obd(obd);
4537         lprocfs_obd_cleanup(obd);
4538
4539         /* free memory of osc quota cache */
4540         lquota_cleanup(quota_interface, obd);
4541
4542         rc = client_obd_cleanup(obd);
4543
4544         ptlrpcd_decref();
4545         RETURN(rc);
4546 }
4547
4548 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
4549 {
4550         struct lprocfs_static_vars lvars = { 0 };
4551         int rc = 0;
4552
4553         lprocfs_osc_init_vars(&lvars);
4554
4555         switch (lcfg->lcfg_command) {
4556         default:
4557                 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
4558                                               lcfg, obd);
4559                 if (rc > 0)
4560                         rc = 0;
4561                 break;
4562         }
4563
4564         return(rc);
4565 }
4566
4567 static int osc_process_config(struct obd_device *obd, obd_count len, void *buf)
4568 {
4569         return osc_process_config_base(obd, buf);
4570 }
4571
4572 struct obd_ops osc_obd_ops = {
4573         .o_owner                = THIS_MODULE,
4574         .o_setup                = osc_setup,
4575         .o_precleanup           = osc_precleanup,
4576         .o_cleanup              = osc_cleanup,
4577         .o_add_conn             = client_import_add_conn,
4578         .o_del_conn             = client_import_del_conn,
4579         .o_connect              = client_connect_import,
4580         .o_reconnect            = osc_reconnect,
4581         .o_disconnect           = osc_disconnect,
4582         .o_statfs               = osc_statfs,
4583         .o_statfs_async         = osc_statfs_async,
4584         .o_packmd               = osc_packmd,
4585         .o_unpackmd             = osc_unpackmd,
4586         .o_precreate            = osc_precreate,
4587         .o_create               = osc_create,
4588         .o_create_async         = osc_create_async,
4589         .o_destroy              = osc_destroy,
4590         .o_getattr              = osc_getattr,
4591         .o_getattr_async        = osc_getattr_async,
4592         .o_setattr              = osc_setattr,
4593         .o_setattr_async        = osc_setattr_async,
4594         .o_brw                  = osc_brw,
4595         .o_punch                = osc_punch,
4596         .o_sync                 = osc_sync,
4597         .o_enqueue              = osc_enqueue,
4598         .o_change_cbdata        = osc_change_cbdata,
4599         .o_find_cbdata          = osc_find_cbdata,
4600         .o_cancel               = osc_cancel,
4601         .o_cancel_unused        = osc_cancel_unused,
4602         .o_iocontrol            = osc_iocontrol,
4603         .o_get_info             = osc_get_info,
4604         .o_set_info_async       = osc_set_info_async,
4605         .o_import_event         = osc_import_event,
4606         .o_llog_init            = osc_llog_init,
4607         .o_llog_finish          = osc_llog_finish,
4608         .o_process_config       = osc_process_config,
4609 };
4610
4611 extern struct lu_kmem_descr osc_caches[];
4612 extern cfs_spinlock_t       osc_ast_guard;
4613 extern cfs_lock_class_key_t osc_ast_guard_class;
4614
4615 int __init osc_init(void)
4616 {
4617         struct lprocfs_static_vars lvars = { 0 };
4618         int rc;
4619         ENTRY;
4620
4621         /* print an address of _any_ initialized kernel symbol from this
4622          * module, to allow debugging with gdb that doesn't support data
4623          * symbols from modules.*/
4624         CDEBUG(D_CONSOLE, "Lustre OSC module (%p).\n", &osc_caches);
4625
4626         rc = lu_kmem_init(osc_caches);
4627
4628         lprocfs_osc_init_vars(&lvars);
4629
4630         cfs_request_module("lquota");
4631         quota_interface = PORTAL_SYMBOL_GET(osc_quota_interface);
4632         lquota_init(quota_interface);
4633         init_obd_quota_ops(quota_interface, &osc_obd_ops);
4634
4635         rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
4636                                  LUSTRE_OSC_NAME, &osc_device_type);
4637         if (rc) {
4638                 if (quota_interface)
4639                         PORTAL_SYMBOL_PUT(osc_quota_interface);
4640                 lu_kmem_fini(osc_caches);
4641                 RETURN(rc);
4642         }
4643
4644         cfs_spin_lock_init(&osc_ast_guard);
4645         cfs_lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
4646
4647         osc_mds_ost_orig_logops = llog_lvfs_ops;
4648         osc_mds_ost_orig_logops.lop_setup = llog_obd_origin_setup;
4649         osc_mds_ost_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
4650         osc_mds_ost_orig_logops.lop_add = llog_obd_origin_add;
4651         osc_mds_ost_orig_logops.lop_connect = llog_origin_connect;
4652
4653         RETURN(rc);
4654 }
4655
4656 #ifdef __KERNEL__
4657 static void /*__exit*/ osc_exit(void)
4658 {
4659         lu_device_type_fini(&osc_device_type);
4660
4661         lquota_exit(quota_interface);
4662         if (quota_interface)
4663                 PORTAL_SYMBOL_PUT(osc_quota_interface);
4664
4665         class_unregister_type(LUSTRE_OSC_NAME);
4666         lu_kmem_fini(osc_caches);
4667 }
4668
4669 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
4670 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
4671 MODULE_LICENSE("GPL");
4672
4673 cfs_module(osc, LUSTRE_VERSION_STRING, osc_init, osc_exit);
4674 #endif