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  *  Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *   Author Peter Braam <braam@clusterfs.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  *  For testing and management it is treated as an obd_device,
  23  *  although * it does not export a full OBD method table (the
  24  *  requests are coming * in over the wire, so object target modules
  25  *  do not have a full * method table.)
  26  *
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_OSC
  33
  34 #ifdef __KERNEL__
  35 # include <linux/version.h>
  36 # include <linux/module.h>
  37 # include <linux/mm.h>
  38 # include <linux/highmem.h>
  39 # include <linux/ctype.h>
  40 # include <linux/init.h>
  41 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  42 #  include <linux/workqueue.h>
  43 #  include <linux/smp_lock.h>
  44 # else
  45 #  include <linux/locks.h>
  46 # endif
  47 #else /* __KERNEL__ */
  48 # include <liblustre.h>
  49 #endif
  50
  51 #include <linux/lustre_dlm.h>
  52 #include <libcfs/kp30.h>
  53 #include <linux/lustre_net.h>
  54 #include <linux/lustre_sec.h>
  55 #include <lustre/lustre_user.h>
  56 #include <linux/obd_ost.h>
  57 #include <linux/obd_lov.h>
  58
  59 #ifdef  __CYGWIN__
  60 # include <ctype.h>
  61 #endif
  62
  63 #include <linux/lustre_ha.h>
  64 #include <linux/lprocfs_status.h>
  65 #include <linux/lustre_log.h>
  66 #include <linux/lustre_audit.h>
  67 #include <linux/lustre_gs.h>
  68 #include "osc_internal.h"
  69
  70 /* Pack OSC object metadata for disk storage (LE byte order). */
  71 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
  72                       struct lov_stripe_md *lsm)
  73 {
  74         int lmm_size;
  75         ENTRY;
  76
  77         lmm_size = sizeof(**lmmp);
  78         if (!lmmp)
  79                 RETURN(lmm_size);
  80
  81         if (*lmmp && !lsm) {
  82                 OBD_FREE(*lmmp, lmm_size);
  83                 *lmmp = NULL;
  84                 RETURN(0);
  85         }
  86
  87         if (!*lmmp) {
  88                 OBD_ALLOC(*lmmp, lmm_size);
  89                 if (!*lmmp)
  90                         RETURN(-ENOMEM);
  91         }
  92
  93         if (lsm) {
  94                 LASSERT(lsm->lsm_object_id);
  95                 LASSERT(lsm->lsm_object_gr);
  96                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
  97                 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
  98         }
  99
 100         RETURN(lmm_size);
 101 }
 102
 103 /* Unpack OSC object metadata from disk storage (LE byte order). */
 104 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
 105                         struct lov_mds_md *lmm, int lmm_bytes)
 106 {
 107         int lsm_size;
 108         ENTRY;
 109
 110         if (lmm != NULL) {
 111                 if (lmm_bytes < sizeof (*lmm)) {
 112                         CERROR("lov_mds_md too small: %d, need %d\n",
 113                                lmm_bytes, (int)sizeof(*lmm));
 114                         RETURN(-EINVAL);
 115                 }
 116                 /* XXX LOV_MAGIC etc check? */
 117
 118                 if (lmm->lmm_object_id == 0) {
 119                         CERROR("lov_mds_md: zero lmm_object_id\n");
 120                         RETURN(-EINVAL);
 121                 }
 122         }
 123
 124         lsm_size = lov_stripe_md_size(1);
 125         if (lsmp == NULL)
 126                 RETURN(lsm_size);
 127
 128         if (*lsmp != NULL && lmm == NULL) {
 129                 OBD_FREE(*lsmp, lsm_size);
 130                 *lsmp = NULL;
 131                 RETURN(0);
 132         }
 133
 134         if (*lsmp == NULL) {
 135                 OBD_ALLOC(*lsmp, lsm_size);
 136                 if (*lsmp == NULL)
 137                         RETURN(-ENOMEM);
 138                 loi_init((*lsmp)->lsm_oinfo);
 139         }
 140
 141         if (lmm != NULL) {
 142                 /* XXX zero *lsmp? */
 143                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
 144                 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
 145                 LASSERT((*lsmp)->lsm_object_id);
 146                 LASSERT((*lsmp)->lsm_object_gr);
 147         }
 148
 149         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
 150
 151         RETURN(lsm_size);
 152 }
 153
 154 static int osc_getattr_interpret(struct ptlrpc_request *req,
 155                                  struct osc_getattr_async_args *aa, int rc)
 156 {
 157         struct ost_body *body;
 158         ENTRY;
 159
 160         if (rc != 0)
 161                 RETURN(rc);
 162
 163         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 164         if (body) {
 165                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 166                 memcpy(aa->aa_oa, &body->oa, sizeof(*aa->aa_oa));
 167
 168                 /* This should really be sent by the OST */
 169                 aa->aa_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 170                 aa->aa_oa->o_valid |= OBD_MD_FLBLKSZ;
 171         } else {
 172                 CERROR("can't unpack ost_body\n");
 173                 rc = -EPROTO;
 174                 aa->aa_oa->o_valid = 0;
 175         }
 176
 177         RETURN(rc);
 178 }
 179
 180 static int osc_getattr_async(struct obd_export *exp, struct obdo *oa,
 181                              struct lov_stripe_md *md,
 182                              struct ptlrpc_request_set *set)
 183 {
 184         struct ptlrpc_request *request;
 185         struct ost_body *body;
 186         int size = sizeof(*body);
 187         struct osc_getattr_async_args *aa;
 188         ENTRY;
 189
 190         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 191                                   OST_GETATTR, 1, &size, NULL);
 192         if (!request)
 193                 RETURN(-ENOMEM);
 194
 195         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 196         memcpy(&body->oa, oa, sizeof(*oa));
 197
 198         request->rq_replen = lustre_msg_size(1, &size);
 199         request->rq_interpret_reply = osc_getattr_interpret;
 200
 201         LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
 202         aa = (struct osc_getattr_async_args *)&request->rq_async_args;
 203         aa->aa_oa = oa;
 204
 205         ptlrpc_set_add_req (set, request);
 206         RETURN (0);
 207 }
 208
 209 static int osc_getattr(struct obd_export *exp, struct obdo *oa,
 210                        struct lov_stripe_md *md)
 211 {
 212         struct ptlrpc_request *request;
 213         struct ost_body *body;
 214         int rc, size = sizeof(*body);
 215         ENTRY;
 216
 217         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 218                                   OST_GETATTR, 1, &size, NULL);
 219         if (!request)
 220                 RETURN(-ENOMEM);
 221
 222         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 223         memcpy(&body->oa, oa, sizeof(*oa));
 224
 225         request->rq_replen = lustre_msg_size(1, &size);
 226
 227         rc = ptlrpc_queue_wait(request);
 228         if (rc) {
 229                 CERROR("%s failed: rc = %d\n", __FUNCTION__, rc);
 230                 GOTO(out, rc);
 231         }
 232
 233         body = lustre_swab_repbuf(request, 0, sizeof (*body),
 234                                   lustre_swab_ost_body);
 235         if (body == NULL) {
 236                 CERROR ("can't unpack ost_body\n");
 237                 GOTO (out, rc = -EPROTO);
 238         }
 239
 240         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 241         memcpy(oa, &body->oa, sizeof(*oa));
 242
 243         /* This should really be sent by the OST */
 244         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 245         oa->o_valid |= OBD_MD_FLBLKSZ;
 246
 247         EXIT;
 248  out:
 249         ptlrpc_req_finished(request);
 250         return rc;
 251 }
 252
 253 static int osc_setattr(struct obd_export *exp, struct obdo *oa,
 254                        struct lov_stripe_md *md, struct obd_trans_info *oti)
 255 {
 256         struct ptlrpc_request *request;
 257         struct ost_body *body;
 258         int rc, size = sizeof(*body);
 259         ENTRY;
 260
 261         LASSERT(!(oa->o_valid & OBD_MD_FLGROUP) || oa->o_gr > 0);
 262
 263         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 264                                   OST_SETATTR, 1, &size, NULL);
 265         if (!request)
 266                 RETURN(-ENOMEM);
 267
 268         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
 269         memcpy(&body->oa, oa, sizeof(*oa));
 270
 271         request->rq_replen = lustre_msg_size(1, &size);
 272
 273         if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
 274                 ptlrpcd_add_req(request);
 275                 rc = 0;
 276         } else {
 277                 rc = ptlrpc_queue_wait(request);
 278                 if (rc)
 279                         GOTO(out, rc);
 280
 281                 body = lustre_swab_repbuf(request, 0, sizeof(*body),
 282                                           lustre_swab_ost_body);
 283                 if (body == NULL)
 284                         GOTO(out, rc = -EPROTO);
 285
 286                 memcpy(oa, &body->oa, sizeof(*oa));
 287         }
 288         EXIT;
 289 out:
 290         ptlrpc_req_finished(request);
 291         RETURN(0);
 292 }
 293
 294 int osc_real_create(struct obd_export *exp, struct obdo *oa,
 295                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
 296 {
 297         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 298         struct ptlrpc_request *request;
 299         struct ost_body *body;
 300         struct lov_stripe_md *lsm;
 301         int rc, size = sizeof(*body);
 302         ENTRY;
 303
 304         LASSERT(oa);
 305         LASSERT(ea);
 306
 307         lsm = *ea;
 308         if (!lsm) {
 309                 rc = obd_alloc_memmd(exp, &lsm);
 310                 if (rc < 0)
 311                         RETURN(rc);
 312         }
 313
 314         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 315                                   OST_CREATE, 1, &size, NULL);
 316         if (!request)
 317                 GOTO(out, rc = -ENOMEM);
 318
 319         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 320         memcpy(&body->oa, oa, sizeof(body->oa));
 321
 322         request->rq_replen = lustre_msg_size(1, &size);
 323         if (oa->o_valid & OBD_MD_FLINLINE) {
 324                 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
 325                         oa->o_flags == OBD_FL_DELORPHAN);
 326                 DEBUG_REQ(D_HA, request,
 327                           "delorphan from OST integration");
 328                 /* Don't resend the delorphan request */
 329                 request->rq_no_resend = request->rq_no_delay = 1;
 330         }
 331
 332         rc = ptlrpc_queue_wait(request);
 333         if (rc)
 334                 GOTO(out_req, rc);
 335
 336         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 337                                   lustre_swab_ost_body);
 338         if (body == NULL) {
 339                 CERROR ("can't unpack ost_body\n");
 340                 GOTO (out_req, rc = -EPROTO);
 341         }
 342
 343         if ((oa->o_valid & OBD_MD_FLFLAGS) && oa->o_flags == OBD_FL_DELORPHAN) {
 344                 struct obd_import *imp = class_exp2cliimp(exp);
 345                 /* MDS declares last known object, OSS responses
 346                  * with next possible object -bzzz */
 347                 spin_lock(&oscc->oscc_lock);
 348                 oscc->oscc_next_id = body->oa.o_id;
 349                 spin_unlock(&oscc->oscc_lock);
 350                 CDEBUG(D_HA, "%s: set nextid "LPD64" after recovery\n",
 351                        imp->imp_target_uuid.uuid, oa->o_id);
 352         }
 353         memcpy(oa, &body->oa, sizeof(*oa));
 354
 355         /* This should really be sent by the OST */
 356         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 357         oa->o_valid |= OBD_MD_FLBLKSZ;
 358
 359         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
 360          * have valid lsm_oinfo data structs, so don't go touching that.
 361          * This needs to be fixed in a big way.
 362          */
 363         lsm->lsm_object_id = oa->o_id;
 364         lsm->lsm_object_gr = oa->o_gr;
 365         *ea = lsm;
 366
 367         if (oti != NULL) {
 368                 oti->oti_transno = request->rq_repmsg->transno;
 369
 370                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
 371                         if (!oti->oti_logcookies)
 372                                 oti_alloc_cookies(oti, 1);
 373                         memcpy(oti->oti_logcookies, obdo_logcookie(oa),
 374                                sizeof(oti->oti_onecookie));
 375                 }
 376         }
 377
 378         CDEBUG(D_HA, "transno: "LPD64"\n", request->rq_repmsg->transno);
 379         EXIT;
 380 out_req:
 381         ptlrpc_req_finished(request);
 382 out:
 383         if (rc && !*ea)
 384                 obd_free_memmd(exp, &lsm);
 385         return rc;
 386 }
 387
 388 static int osc_punch(struct obd_export *exp, struct obdo *oa,
 389                      struct lov_stripe_md *md, obd_size start,
 390                      obd_size end, struct obd_trans_info *oti)
 391 {
 392         struct ptlrpc_request *request;
 393         struct ost_body *body;
 394         int rc, size = sizeof(*body);
 395         ENTRY;
 396
 397         if (!oa) {
 398                 CERROR("oa NULL\n");
 399                 RETURN(-EINVAL);
 400         }
 401
 402         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 403                                   OST_PUNCH, 1, &size, NULL);
 404         if (!request)
 405                 RETURN(-ENOMEM);
 406
 407         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 408         memcpy(&body->oa, oa, sizeof(*oa));
 409
 410         /* overload the size and blocks fields in the oa with start/end */
 411         body->oa.o_size = start;
 412         body->oa.o_blocks = end;
 413         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 414
 415         request->rq_replen = lustre_msg_size(1, &size);
 416
 417         rc = ptlrpc_queue_wait(request);
 418         if (rc)
 419                 GOTO(out, rc);
 420
 421         body = lustre_swab_repbuf (request, 0, sizeof (*body),
 422                                    lustre_swab_ost_body);
 423         if (body == NULL) {
 424                 CERROR ("can't unpack ost_body\n");
 425                 GOTO (out, rc = -EPROTO);
 426         }
 427
 428         memcpy(oa, &body->oa, sizeof(*oa));
 429
 430         EXIT;
 431  out:
 432         ptlrpc_req_finished(request);
 433         return rc;
 434 }
 435
 436 static int osc_sync(struct obd_export *exp, struct obdo *oa,
 437                     struct lov_stripe_md *md, obd_size start,
 438                     obd_size end)
 439 {
 440         struct ptlrpc_request *request;
 441         struct ost_body *body;
 442         int rc, size = sizeof(*body);
 443         ENTRY;
 444
 445         if (!oa) {
 446                 CERROR("oa NULL\n");
 447                 RETURN(-EINVAL);
 448         }
 449
 450         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 451                                   OST_SYNC, 1, &size, NULL);
 452         if (!request)
 453                 RETURN(-ENOMEM);
 454
 455         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 456         memcpy(&body->oa, oa, sizeof(*oa));
 457
 458         /* overload the size and blocks fields in the oa with start/end */
 459         body->oa.o_size = start;
 460         body->oa.o_blocks = end;
 461         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 462
 463         request->rq_replen = lustre_msg_size(1, &size);
 464
 465         rc = ptlrpc_queue_wait(request);
 466         if (rc)
 467                 GOTO(out, rc);
 468
 469         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 470                                   lustre_swab_ost_body);
 471         if (body == NULL) {
 472                 CERROR ("can't unpack ost_body\n");
 473                 GOTO (out, rc = -EPROTO);
 474         }
 475
 476         memcpy(oa, &body->oa, sizeof(*oa));
 477
 478         EXIT;
 479  out:
 480         ptlrpc_req_finished(request);
 481         return rc;
 482 }
 483
 484 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
 485                        struct lov_stripe_md *ea, struct obd_trans_info *oti)
 486 {
 487         struct ptlrpc_request *request;
 488         struct ost_body *body;
 489         int rc, size = sizeof(*body);
 490         ENTRY;
 491
 492         if (!oa) {
 493                 CERROR("oa NULL\n");
 494                 RETURN(-EINVAL);
 495         }
 496
 497         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 498                                   OST_DESTROY, 1, &size, NULL);
 499         if (!request)
 500                 RETURN(-ENOMEM);
 501         request->rq_request_portal = OST_DESTROY_PORTAL;
 502
 503         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 504
 505         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
 506                 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
 507                        sizeof(*oti->oti_logcookies));
 508                 oti->oti_logcookies++;
 509         }
 510
 511         memcpy(&body->oa, oa, sizeof(*oa));
 512         request->rq_replen = lustre_msg_size(1, &size);
 513
 514         if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
 515                 ptlrpcd_add_req(request);
 516                 rc = 0;
 517         } else {
 518                 rc = ptlrpc_queue_wait(request);
 519
 520                 if (rc == -ENOENT)
 521                         rc = 0;
 522
 523                 if (rc) {
 524                         ptlrpc_req_finished(request);
 525                         RETURN(rc);
 526                 }
 527
 528                 body = lustre_swab_repbuf(request, 0, sizeof(*body),
 529                                           lustre_swab_ost_body);
 530                 if (body == NULL) {
 531                         CERROR ("Can't unpack body\n");
 532                         ptlrpc_req_finished(request);
 533                         RETURN(-EPROTO);
 534                 }
 535
 536                 memcpy(oa, &body->oa, sizeof(*oa));
 537                 ptlrpc_req_finished(request);
 538         }
 539         RETURN(rc);
 540 }
 541
 542 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 543                                 long writing_bytes)
 544 {
 545         obd_valid bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
 546
 547         LASSERT(!(oa->o_valid & bits));
 548
 549         oa->o_valid |= bits;
 550         spin_lock(&cli->cl_loi_list_lock);
 551         oa->o_dirty = cli->cl_dirty;
 552         oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
 553         oa->o_grant = cli->cl_avail_grant;
 554         oa->o_dropped = cli->cl_lost_grant;
 555         cli->cl_lost_grant = 0;
 556         spin_unlock(&cli->cl_loi_list_lock);
 557         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
 558                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 559 }
 560
 561 /* caller must hold loi_list_lock */
 562 static void osc_consume_write_grant(struct client_obd *cli,
 563                                     struct osc_async_page *oap)
 564 {
 565         cli->cl_dirty += PAGE_SIZE;
 566         cli->cl_avail_grant -= PAGE_SIZE;
 567         oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
 568         CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
 569         LASSERT(cli->cl_avail_grant >= 0);
 570 }
 571
 572 static unsigned long rpcs_in_flight(struct client_obd *cli)
 573 {
 574         return cli->cl_r_in_flight + cli->cl_w_in_flight;
 575 }
 576
 577 /* caller must hold loi_list_lock */
 578 void osc_wake_cache_waiters(struct client_obd *cli)
 579 {
 580         struct list_head *l, *tmp;
 581         struct osc_cache_waiter *ocw;
 582
 583         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 584                 /* if we can't dirty more, we must wait until some is written */
 585                 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
 586                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
 587                                cli->cl_dirty, cli->cl_dirty_max);
 588                         return;
 589                 }
 590
 591                 /* if still dirty cache but no grant wait for pending RPCs that
 592                  * may yet return us some grant before doing sync writes */
 593                 if (cli->cl_w_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
 594                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
 595                                cli->cl_w_in_flight);
 596                 }
 597                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
 598                 list_del_init(&ocw->ocw_entry);
 599                 if (cli->cl_avail_grant < PAGE_SIZE) {
 600                         /* no more RPCs in flight to return grant, do sync IO */
 601                         ocw->ocw_rc = -EDQUOT;
 602                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
 603                 } else {
 604                         osc_consume_write_grant(cli, ocw->ocw_oap);
 605                 }
 606
 607                 wake_up(&ocw->ocw_waitq);
 608         }
 609
 610         EXIT;
 611 }
 612
 613 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 614 {
 615         spin_lock(&cli->cl_loi_list_lock);
 616         CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
 617         cli->cl_avail_grant += body->oa.o_grant;
 618         /* waiters are woken in brw_interpret_oap */
 619         spin_unlock(&cli->cl_loi_list_lock);
 620 }
 621
 622 /* We assume that the reason this OSC got a short read is because it read
 623  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
 624  * via the LOV, and it _knows_ it's reading inside the file, it's just that
 625  * this stripe never got written at or beyond this stripe offset yet. */
 626 static void handle_short_read(int nob_read, obd_count page_count,
 627                               struct brw_page *pga)
 628 {
 629         char *ptr;
 630
 631         /* skip bytes read OK */
 632         while (nob_read > 0) {
 633                 LASSERT (page_count > 0);
 634
 635                 if (pga->count > nob_read) {
 636                         /* EOF inside this page */
 637                         ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
 638                         memset(ptr + nob_read, 0, pga->count - nob_read);
 639                         kunmap(pga->pg);
 640                         page_count--;
 641                         pga++;
 642                         break;
 643                 }
 644
 645                 nob_read -= pga->count;
 646                 page_count--;
 647                 pga++;
 648         }
 649
 650         /* zero remaining pages */
 651         while (page_count-- > 0) {
 652                 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
 653                 memset(ptr, 0, pga->count);
 654                 kunmap(pga->pg);
 655                 pga++;
 656         }
 657 }
 658
 659 static int check_write_rcs(struct ptlrpc_request *request,
 660                            int requested_nob, int niocount,
 661                            obd_count page_count, struct brw_page *pga)
 662 {
 663         int *remote_rcs, i;
 664
 665         /* return error if any niobuf was in error */
 666         remote_rcs = lustre_swab_repbuf(request, 1,
 667                                         sizeof(*remote_rcs) * niocount, NULL);
 668         if (remote_rcs == NULL) {
 669                 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
 670                 return(-EPROTO);
 671         }
 672         if (lustre_msg_swabbed(request->rq_repmsg))
 673                 for (i = 0; i < niocount; i++)
 674                         __swab32s((__u32 *)&remote_rcs[i]);
 675
 676         for (i = 0; i < niocount; i++) {
 677                 if (remote_rcs[i] < 0)
 678                         return(remote_rcs[i]);
 679
 680                 if (remote_rcs[i] != 0) {
 681                         CERROR("rc[%d] invalid (%d) req %p\n",
 682                                 i, remote_rcs[i], request);
 683                         return(-EPROTO);
 684                 }
 685         }
 686
 687         if (request->rq_bulk->bd_nob_transferred != requested_nob) {
 688                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
 689                        requested_nob, request->rq_bulk->bd_nob_transferred);
 690                 return(-EPROTO);
 691         }
 692
 693         return (0);
 694 }
 695
 696 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
 697 {
 698         if (p1->flag != p2->flag) {
 699                 unsigned mask = ~OBD_BRW_FROM_GRANT;
 700
 701                 /* warn if we try to combine flags that we don't know to be
 702                  * safe to combine */
 703                 if ((p1->flag & mask) != (p2->flag & mask))
 704                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
 705                                "same brw?\n", p1->flag, p2->flag);
 706                 return 0;
 707         }
 708
 709         return (p1->disk_offset + p1->count == p2->disk_offset);
 710 }
 711
 712 #if CHECKSUM_BULK
 713 static obd_count cksum_pages(int nob, obd_count page_count,
 714                              struct brw_page *pga)
 715 {
 716         obd_count cksum = 0;
 717         char *ptr;
 718
 719         while (nob > 0) {
 720                 LASSERT (page_count > 0);
 721
 722                 ptr = kmap(pga->pg);
 723                 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
 724                              pga->count > nob ? nob : pga->count);
 725                 kunmap(pga->pg);
 726
 727                 nob -= pga->count;
 728                 page_count--;
 729                 pga++;
 730         }
 731
 732         return (cksum);
 733 }
 734 #endif
 735
 736 #define osc_encrypt_page(page, off, count)  \
 737         osc_crypt_page(page, off, count, ENCRYPT_DATA)
 738 #define osc_decrypt_page(page, off, count)  \
 739         osc_crypt_page(page, off, count, DECRYPT_DATA)
 740
 741 /*Put a global call back var here is Ugly, but put it to client_obd
 742  *also seems not a good idea, WangDi*/
 743 crypt_cb_t  osc_crypt_cb = NULL;
 744
 745 static int osc_crypt_page(struct page *page, obd_off page_off, obd_off count,
 746                           int flags)
 747 {
 748         int rc = 0;
 749         ENTRY;
 750
 751         if (osc_crypt_cb != NULL)
 752                 rc = osc_crypt_cb(page, page_off, count, flags);
 753         if (rc != 0)
 754                 CERROR("crypt page error %d \n", rc);
 755         RETURN(rc);
 756 }
 757
 758 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
 759                                 struct lov_stripe_md *lsm, obd_count page_count,
 760                                 struct brw_page *pga, int *requested_nobp,
 761                                 int *niocountp, struct ptlrpc_request **reqp)
 762 {
 763         struct ptlrpc_request   *req;
 764         struct ptlrpc_bulk_desc *desc;
 765         struct client_obd       *cli = &imp->imp_obd->u.cli;
 766         struct ost_body         *body;
 767         struct obd_ioobj        *ioobj;
 768         struct niobuf_remote    *niobuf;
 769         int                      niocount;
 770         int                      size[3];
 771         int                      i;
 772         int                      requested_nob;
 773         int                      opc;
 774         int                      rc;
 775
 776         opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
 777
 778         for (niocount = i = 1; i < page_count; i++)
 779                 if (!can_merge_pages(&pga[i - 1], &pga[i]))
 780                         niocount++;
 781
 782         size[0] = sizeof(*body);
 783         size[1] = sizeof(*ioobj);
 784         size[2] = niocount * sizeof(*niobuf);
 785
 786         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, opc, 3, size, NULL);
 787         if (req == NULL)
 788                 return (-ENOMEM);
 789
 790         if (opc == OST_WRITE)
 791                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 792                                              BULK_GET_SOURCE, OST_BULK_PORTAL);
 793         else
 794                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 795                                              BULK_PUT_SINK, OST_BULK_PORTAL);
 796         if (desc == NULL)
 797                 GOTO(out, rc = -ENOMEM);
 798         /* NB request now owns desc and will free it when it gets freed */
 799
 800         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 801         ioobj = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*ioobj));
 802         niobuf = lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf));
 803
 804         memcpy(&body->oa, oa, sizeof(*oa));
 805
 806         obdo_to_ioobj(oa, ioobj);
 807         ioobj->ioo_bufcnt = niocount;
 808
 809         LASSERT (page_count > 0);
 810
 811         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
 812                 struct brw_page *pg = &pga[i];
 813                 struct brw_page *pg_prev = pg - 1;
 814
 815                 LASSERT(pg->count > 0);
 816                 LASSERTF((pg->page_offset & ~PAGE_MASK)+ pg->count <= PAGE_SIZE,
 817                          "i: %d pg: %p pg_off: "LPU64", count: %u\n", i, pg,
 818                          pg->page_offset, pg->count);
 819                 LASSERTF(i == 0 || pg->disk_offset > pg_prev->disk_offset,
 820                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
 821                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
 822                          i, page_count,
 823                          pg->pg, pg->pg->private, pg->pg->index, pg->disk_offset,
 824                          pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
 825                          pg_prev->disk_offset);
 826
 827                 if (opc == OST_WRITE) {
 828                         osc_encrypt_page(pg->pg, pg->page_offset, pg->count);
 829                 }
 830
 831                 ptlrpc_prep_bulk_page(desc, pg->pg,
 832                                       pg->page_offset & ~PAGE_MASK, pg->count);
 833                 requested_nob += pg->count;
 834
 835                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
 836                         niobuf--;
 837                         niobuf->len += pg->count;
 838                 } else {
 839                         niobuf->offset = pg->disk_offset;
 840                         niobuf->len    = pg->count;
 841                         niobuf->flags  = pg->flag;
 842                 }
 843         }
 844
 845         LASSERT((void *)(niobuf - niocount) ==
 846                 lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf)));
 847         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
 848
 849         /* size[0] still sizeof (*body) */
 850         if (opc == OST_WRITE) {
 851 #if CHECKSUM_BULK
 852                 body->oa.o_valid |= OBD_MD_FLCKSUM;
 853                 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
 854 #endif
 855                 /* 1 RC per niobuf */
 856                 size[1] = sizeof(__u32) * niocount;
 857                 req->rq_replen = lustre_msg_size(2, size);
 858         } else {
 859                 /* 1 RC for the whole I/O */
 860                 req->rq_replen = lustre_msg_size(1, size);
 861         }
 862
 863         *niocountp = niocount;
 864         *requested_nobp = requested_nob;
 865         *reqp = req;
 866         return (0);
 867
 868  out:
 869         ptlrpc_req_finished (req);
 870         return (rc);
 871 }
 872
 873 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
 874                                 int requested_nob, int niocount,
 875                                 obd_count page_count, struct brw_page *pga,
 876                                 int rc)
 877 {
 878         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 879         struct ost_body *body;
 880         ENTRY;
 881
 882         if (rc < 0)
 883                 RETURN(rc);
 884
 885         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 886         if (body == NULL) {
 887                 CERROR ("Can't unpack body\n");
 888                 RETURN(-EPROTO);
 889         }
 890
 891         osc_update_grant(cli, body);
 892         memcpy(oa, &body->oa, sizeof(*oa));
 893
 894         if (req->rq_reqmsg->opc == OST_WRITE) {
 895                 if (rc > 0) {
 896                         CERROR ("Unexpected +ve rc %d\n", rc);
 897                         RETURN(-EPROTO);
 898                 }
 899                 LASSERT (req->rq_bulk->bd_nob == requested_nob);
 900                 osc_decrypt_page(pga->pg, pga->page_offset,
 901                                  pga->count);
 902                 RETURN(check_write_rcs(req, requested_nob, niocount,
 903                                        page_count, pga));
 904         }
 905
 906         if (rc > requested_nob) {
 907                 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
 908                 RETURN(-EPROTO);
 909         }
 910
 911         if (rc != req->rq_bulk->bd_nob_transferred) {
 912                 CERROR ("Unexpected rc %d (%d transferred)\n",
 913                         rc, req->rq_bulk->bd_nob_transferred);
 914                 return (-EPROTO);
 915         }
 916
 917         if (rc < requested_nob)
 918                 handle_short_read(rc, page_count, pga);
 919
 920 #if CHECKSUM_BULK
 921         if (oa->o_valid & OBD_MD_FLCKSUM) {
 922                 const struct ptlrpc_peer *peer =
 923                         &req->rq_import->imp_connection->c_peer;
 924                 static int cksum_counter;
 925                 obd_count server_cksum = oa->o_cksum;
 926                 obd_count cksum = cksum_pages(rc, page_count, pga);
 927                 char str[PTL_NALFMT_SIZE];
 928
 929                 ptlrpc_peernid2str(peer, str);
 930
 931                 cksum_counter++;
 932                 if (server_cksum != cksum) {
 933                         CERROR("Bad checksum: server %x, client %x, server NID "
 934                                LPX64" (%s)\n", server_cksum, cksum,
 935                                peer->peer_id.nid, str);
 936                         cksum_counter = 0;
 937                         oa->o_cksum = cksum;
 938                 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
 939                         CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
 940                               cksum_counter, peer->peer_id.nid, str, cksum);
 941                 }
 942         } else {
 943                 static int cksum_missed;
 944
 945                 cksum_missed++;
 946                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
 947                         CERROR("Request checksum %u from "LPX64", no reply\n",
 948                                cksum_missed,
 949                                req->rq_import->imp_connection->c_peer.peer_id.nid);
 950         }
 951 #endif
 952         osc_decrypt_page(pga->pg, pga->page_offset, pga->count);
 953         RETURN(0);
 954 }
 955
 956 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
 957                             struct lov_stripe_md *lsm,
 958                             obd_count page_count, struct brw_page *pga)
 959 {
 960         int                    requested_nob;
 961         int                    niocount;
 962         struct ptlrpc_request *request;
 963         int                    rc;
 964         ENTRY;
 965
 966 restart_bulk:
 967         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
 968                                   page_count, pga, &requested_nob, &niocount,
 969                                   &request);
 970         if (rc != 0)
 971                 return (rc);
 972
 973         rc = ptlrpc_queue_wait(request);
 974
 975         if (rc == -ETIMEDOUT && request->rq_resend) {
 976                 DEBUG_REQ(D_HA, request,  "BULK TIMEOUT");
 977                 ptlrpc_req_finished(request);
 978                 goto restart_bulk;
 979         }
 980
 981         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
 982                                   page_count, pga, rc);
 983
 984         ptlrpc_req_finished(request);
 985         RETURN (rc);
 986 }
 987
 988 static int brw_interpret(struct ptlrpc_request *request,
 989                          struct osc_brw_async_args *aa, int rc)
 990 {
 991         struct obdo *oa      = aa->aa_oa;
 992         int requested_nob    = aa->aa_requested_nob;
 993         int niocount         = aa->aa_nio_count;
 994         obd_count page_count = aa->aa_page_count;
 995         struct brw_page *pga = aa->aa_pga;
 996         ENTRY;
 997
 998         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
 999                                   page_count, pga, rc);
1000         RETURN (rc);
1001 }
1002
1003 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1004                           struct lov_stripe_md *lsm, obd_count page_count,
1005                           struct brw_page *pga, struct ptlrpc_request_set *set)
1006 {
1007         struct ptlrpc_request     *request;
1008         int                        requested_nob;
1009         int                        nio_count;
1010         struct osc_brw_async_args *aa;
1011         int                        rc;
1012         ENTRY;
1013
1014         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1015                                   page_count, pga, &requested_nob, &nio_count,
1016                                   &request);
1017         if (rc == 0) {
1018                 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1019                 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1020                 aa->aa_oa = oa;
1021                 aa->aa_requested_nob = requested_nob;
1022                 aa->aa_nio_count = nio_count;
1023                 aa->aa_page_count = page_count;
1024                 aa->aa_pga = pga;
1025
1026                 request->rq_interpret_reply = brw_interpret;
1027                 ptlrpc_set_add_req(set, request);
1028         }
1029         RETURN (rc);
1030 }
1031
1032 #ifndef min_t
1033 #define min_t(type,x,y) \
1034         ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
1035 #endif
1036
1037 /*
1038  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1039  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1040  * fine for our small page arrays and doesn't require allocation.  its an
1041  * insertion sort that swaps elements that are strides apart, shrinking the
1042  * stride down until its '1' and the array is sorted.
1043  */
1044 static void sort_brw_pages(struct brw_page *array, int num)
1045 {
1046         int stride, i, j;
1047         struct brw_page tmp;
1048
1049         if (num == 1)
1050                 return;
1051         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1052                 ;
1053
1054         do {
1055                 stride /= 3;
1056                 for (i = stride ; i < num ; i++) {
1057                         tmp = array[i];
1058                         j = i;
1059                         while (j >= stride && array[j - stride].disk_offset >
1060                                 tmp.disk_offset) {
1061                                 array[j] = array[j - stride];
1062                                 j -= stride;
1063                         }
1064                         array[j] = tmp;
1065                 }
1066         } while (stride > 1);
1067 }
1068
1069 /* make sure we the regions we're passing to elan don't violate its '4
1070  * fragments' constraint.  portal headers are a fragment, all full
1071  * PAGE_SIZE long pages count as 1 fragment, and each partial page
1072  * counts as a fragment.  I think.  see bug 934. */
1073 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1074 {
1075         int frags_left = 3;
1076         int saw_whole_frag = 0;
1077         int i;
1078
1079         for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1080                 if (pg->count == PAGE_SIZE) {
1081                         if (!saw_whole_frag) {
1082                                 saw_whole_frag = 1;
1083                                 frags_left--;
1084                         }
1085                 } else {
1086                         frags_left--;
1087                 }
1088         }
1089         return i;
1090 }
1091
1092 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1093                    struct lov_stripe_md *lsm, obd_count page_count,
1094                    struct brw_page *pga, struct obd_trans_info *oti)
1095 {
1096         ENTRY;
1097
1098         if (cmd == OBD_BRW_CHECK) {
1099                 /* The caller just wants to know if there's a chance that this
1100                  * I/O can succeed */
1101                 struct obd_import *imp = class_exp2cliimp(exp);
1102
1103                 if (imp == NULL || imp->imp_invalid)
1104                         RETURN(-EIO);
1105                 RETURN(0);
1106         }
1107
1108         while (page_count) {
1109                 obd_count pages_per_brw;
1110                 int rc;
1111
1112                 if (page_count > PTLRPC_MAX_BRW_PAGES)
1113                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1114                 else
1115                         pages_per_brw = page_count;
1116
1117                 sort_brw_pages(pga, pages_per_brw);
1118                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1119
1120                 rc = osc_brw_internal(cmd, exp, oa, lsm, pages_per_brw, pga);
1121
1122                 if (rc != 0)
1123                         RETURN(rc);
1124
1125                 page_count -= pages_per_brw;
1126                 pga += pages_per_brw;
1127         }
1128         RETURN(0);
1129 }
1130
1131 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1132                          struct lov_stripe_md *lsm, obd_count page_count,
1133                          struct brw_page *pga, struct ptlrpc_request_set *set,
1134                          struct obd_trans_info *oti)
1135 {
1136         ENTRY;
1137
1138         if (cmd == OBD_BRW_CHECK) {
1139                 /* The caller just wants to know if there's a chance that this
1140                  * I/O can succeed */
1141                 struct obd_import *imp = class_exp2cliimp(exp);
1142
1143                 if (imp == NULL || imp->imp_invalid)
1144                         RETURN(-EIO);
1145                 RETURN(0);
1146         }
1147
1148         while (page_count) {
1149                 obd_count pages_per_brw;
1150                 int rc;
1151
1152                 if (page_count > PTLRPC_MAX_BRW_PAGES)
1153                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1154                 else
1155                         pages_per_brw = page_count;
1156
1157                 sort_brw_pages(pga, pages_per_brw);
1158                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1159
1160                 rc = async_internal(cmd, exp, oa, lsm, pages_per_brw, pga, set);
1161
1162                 if (rc != 0)
1163                         RETURN(rc);
1164
1165                 page_count -= pages_per_brw;
1166                 pga += pages_per_brw;
1167         }
1168         RETURN(0);
1169 }
1170
1171 static void osc_check_rpcs(struct client_obd *cli);
1172 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1173                            int sent);
1174 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1175 static void lop_update_pending(struct client_obd *cli,
1176                                struct loi_oap_pages *lop, int cmd, int delta);
1177
1178 /* this is called when a sync waiter receives an interruption.  Its job is to
1179  * get the caller woken as soon as possible.  If its page hasn't been put in an
1180  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
1181  * desiring interruption which will forcefully complete the rpc once the rpc
1182  * has timed out */
1183 static void osc_occ_interrupted(struct oig_callback_context *occ)
1184 {
1185         struct osc_async_page *oap;
1186         struct loi_oap_pages *lop;
1187         struct lov_oinfo *loi;
1188         ENTRY;
1189
1190         /* XXX member_of() */
1191         oap = list_entry(occ, struct osc_async_page, oap_occ);
1192
1193         spin_lock(&oap->oap_cli->cl_loi_list_lock);
1194
1195         oap->oap_interrupted = 1;
1196
1197         /* ok, it's been put in an rpc. */
1198         if (oap->oap_request != NULL) {
1199                 ptlrpc_mark_interrupted(oap->oap_request);
1200                 ptlrpcd_wake(oap->oap_request);
1201                 GOTO(unlock, 0);
1202         }
1203
1204         /* we don't get interruption callbacks until osc_trigger_sync_io()
1205          * has been called and put the sync oaps in the pending/urgent lists.*/
1206         if (!list_empty(&oap->oap_pending_item)) {
1207                 list_del_init(&oap->oap_pending_item);
1208                 if (oap->oap_async_flags & ASYNC_URGENT)
1209                         list_del_init(&oap->oap_urgent_item);
1210
1211                 loi = oap->oap_loi;
1212                 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1213                         &loi->loi_write_lop : &loi->loi_read_lop;
1214                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1215                 loi_list_maint(oap->oap_cli, oap->oap_loi);
1216
1217                 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1218                 oap->oap_oig = NULL;
1219         }
1220
1221 unlock:
1222         spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1223 }
1224
1225 /* this must be called holding the loi list lock to give coverage to exit_cache,
1226  * async_flag maintenance, and oap_request */
1227 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1228                               struct osc_async_page *oap, int sent, int rc)
1229 {
1230         osc_exit_cache(cli, oap, sent);
1231         oap->oap_async_flags = 0;
1232         oap->oap_interrupted = 0;
1233
1234         if (oap->oap_request != NULL) {
1235                 ptlrpc_req_finished(oap->oap_request);
1236                 oap->oap_request = NULL;
1237         }
1238
1239         if (rc == 0 && oa != NULL)
1240                 oap->oap_loi->loi_blocks = oa->o_blocks;
1241
1242         if (oap->oap_oig) {
1243                 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1244                 oap->oap_oig = NULL;
1245                 EXIT;
1246                 return;
1247         }
1248
1249         oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1250                                            oa, rc);
1251 }
1252
1253 static int brw_interpret_oap(struct ptlrpc_request *request,
1254                              struct osc_brw_async_args *aa, int rc)
1255 {
1256         struct osc_async_page *oap;
1257         struct client_obd *cli;
1258         struct list_head *pos, *n;
1259         struct timeval now;
1260         ENTRY;
1261
1262         do_gettimeofday(&now);
1263         rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1264                                   aa->aa_nio_count, aa->aa_page_count,
1265                                   aa->aa_pga, rc);
1266
1267         CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1268
1269         cli = aa->aa_cli;
1270         /* in failout recovery we ignore writeback failure and want
1271          * to just tell llite to unlock the page and continue */
1272         if (request->rq_reqmsg->opc == OST_WRITE &&
1273             (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1274                 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1275                        cli->cl_import,
1276                        cli->cl_import ? cli->cl_import->imp_invalid : -1);
1277                 rc = 0;
1278         }
1279
1280         spin_lock(&cli->cl_loi_list_lock);
1281
1282         if (request->rq_reqmsg->opc == OST_WRITE)
1283                 lprocfs_stime_record(&cli->cl_write_stime, &now,
1284                                      &request->rq_rpcd_start);
1285         else
1286                 lprocfs_stime_record(&cli->cl_read_stime, &now,
1287                                      &request->rq_rpcd_start);
1288
1289
1290
1291         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1292          * is called so we know whether to go to sync BRWs or wait for more
1293          * RPCs to complete */
1294         if (request->rq_reqmsg->opc == OST_WRITE)
1295                 cli->cl_w_in_flight--;
1296         else
1297                 cli->cl_r_in_flight--;
1298
1299         /* the caller may re-use the oap after the completion call so
1300          * we need to clean it up a little */
1301         list_for_each_safe(pos, n, &aa->aa_oaps) {
1302                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1303
1304                 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1305                        //oap->oap_page, oap->oap_page->index, oap);
1306
1307                 list_del_init(&oap->oap_rpc_item);
1308                 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1309         }
1310
1311         osc_wake_cache_waiters(cli);
1312         osc_check_rpcs(cli);
1313         spin_unlock(&cli->cl_loi_list_lock);
1314
1315         obdo_free(aa->aa_oa);
1316         OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1317
1318         RETURN(0);
1319 }
1320
1321 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1322                                             struct list_head *rpc_list,
1323                                             int page_count, int cmd)
1324 {
1325         struct ptlrpc_request *req;
1326         struct brw_page *pga = NULL;
1327         int requested_nob, nio_count;
1328         struct osc_brw_async_args *aa;
1329         struct obdo *oa = NULL;
1330         struct obd_async_page_ops *ops = NULL;
1331         void *caller_data = NULL;
1332         struct list_head *pos;
1333         int i, rc;
1334
1335         LASSERT(!list_empty(rpc_list));
1336
1337         OBD_ALLOC(pga, sizeof(*pga) * page_count);
1338         if (pga == NULL)
1339                 RETURN(ERR_PTR(-ENOMEM));
1340
1341         oa = obdo_alloc();
1342         if (oa == NULL)
1343                 GOTO(out, req = ERR_PTR(-ENOMEM));
1344
1345         i = 0;
1346         list_for_each(pos, rpc_list) {
1347                 struct osc_async_page *oap;
1348
1349                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1350                 if (ops == NULL) {
1351                         ops = oap->oap_caller_ops;
1352                         caller_data = oap->oap_caller_data;
1353                 }
1354                 pga[i].disk_offset = oap->oap_obj_off + oap->oap_page_off;
1355                 pga[i].page_offset = pga[i].disk_offset;
1356                 pga[i].pg = oap->oap_page;
1357                 pga[i].count = oap->oap_count;
1358                 pga[i].flag = oap->oap_brw_flags;
1359                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1360                        pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1361                 i++;
1362         }
1363
1364         /* always get the data for the obdo for the rpc */
1365         LASSERT(ops != NULL);
1366         ops->ap_fill_obdo(caller_data, cmd, oa);
1367
1368         sort_brw_pages(pga, page_count);
1369         rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1370                                   pga, &requested_nob, &nio_count, &req);
1371         if (rc != 0) {
1372                 CERROR("prep_req failed: %d\n", rc);
1373                 GOTO(out, req = ERR_PTR(rc));
1374         }
1375
1376         LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1377         aa = (struct osc_brw_async_args *)&req->rq_async_args;
1378         aa->aa_oa = oa;
1379         aa->aa_requested_nob = requested_nob;
1380         aa->aa_nio_count = nio_count;
1381         aa->aa_page_count = page_count;
1382         aa->aa_pga = pga;
1383         aa->aa_cli = cli;
1384
1385 out:
1386         if (IS_ERR(req)) {
1387                 if (oa)
1388                         obdo_free(oa);
1389                 if (pga)
1390                         OBD_FREE(pga, sizeof(*pga) * page_count);
1391         }
1392         RETURN(req);
1393 }
1394
1395 static void lop_update_pending(struct client_obd *cli,
1396                                struct loi_oap_pages *lop, int cmd, int delta)
1397 {
1398         lop->lop_num_pending += delta;
1399         if (cmd == OBD_BRW_WRITE)
1400                 cli->cl_pending_w_pages += delta;
1401         else
1402                 cli->cl_pending_r_pages += delta;
1403 }
1404
1405 /* the loi lock is held across this function but it's allowed to release
1406  * and reacquire it during its work */
1407 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1408                             int cmd, struct loi_oap_pages *lop)
1409 {
1410         struct ptlrpc_request *request;
1411         obd_count page_count = 0;
1412         struct list_head *tmp, *pos;
1413         struct osc_async_page *oap = NULL;
1414         struct osc_brw_async_args *aa;
1415         struct obd_async_page_ops *ops;
1416         LIST_HEAD(rpc_list);
1417         ENTRY;
1418
1419         /* first we find the pages we're allowed to work with */
1420         list_for_each_safe(pos, tmp, &lop->lop_pending) {
1421                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1422                 ops = oap->oap_caller_ops;
1423
1424                 LASSERT(oap->oap_magic == OAP_MAGIC);
1425
1426                 /* in llite being 'ready' equates to the page being locked
1427                  * until completion unlocks it.  commit_write submits a page
1428                  * as not ready because its unlock will happen unconditionally
1429                  * as the call returns.  if we race with commit_write giving
1430                  * us that page we dont' want to create a hole in the page
1431                  * stream, so we stop and leave the rpc to be fired by
1432                  * another dirtier or kupdated interval (the not ready page
1433                  * will still be on the dirty list).  we could call in
1434                  * at the end of ll_file_write to process the queue again. */
1435                 if (!(oap->oap_async_flags & ASYNC_READY)) {
1436                         int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1437                         if (rc < 0)
1438                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
1439                                                 "instead of ready\n", oap,
1440                                                 oap->oap_page, rc);
1441                         switch (rc) {
1442                         case -EAGAIN:
1443                                 /* llite is telling us that the page is still
1444                                  * in commit_write and that we should try
1445                                  * and put it in an rpc again later.  we
1446                                  * break out of the loop so we don't create
1447                                  * a hole in the sequence of pages in the rpc
1448                                  * stream.*/
1449                                 pos = NULL;
1450                                 break;
1451                         case -EINTR:
1452                                 /* the io isn't needed.. tell the checks
1453                                  * below to complete the rpc with EINTR */
1454                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1455                                 oap->oap_count = -EINTR;
1456                                 break;
1457                         case 0:
1458                                 oap->oap_async_flags |= ASYNC_READY;
1459                                 break;
1460                         default:
1461                                 LASSERTF(0, "oap %p page %p returned %d "
1462                                             "from make_ready\n", oap,
1463                                             oap->oap_page, rc);
1464                                 break;
1465                         }
1466                 }
1467                 if (pos == NULL)
1468                         break;
1469                 /*
1470                  * Page submitted for IO has to be locked. Either by
1471                  * ->ap_make_ready() or by higher layers.
1472                  *
1473                  * XXX nikita: this assertion should be adjusted when lustre
1474                  * starts using PG_writeback for pages being written out.
1475                  */
1476                 LASSERT(PageLocked(oap->oap_page));
1477
1478                 /* take the page out of our book-keeping */
1479                 list_del_init(&oap->oap_pending_item);
1480                 lop_update_pending(cli, lop, cmd, -1);
1481                 list_del_init(&oap->oap_urgent_item);
1482
1483                 /* ask the caller for the size of the io as the rpc leaves. */
1484                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1485                         oap->oap_count =
1486                                 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1487                 if (oap->oap_count <= 0) {
1488                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1489                                oap->oap_count);
1490                         osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1491                         continue;
1492                 }
1493
1494                 /* now put the page back in our accounting */
1495                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1496                 if (++page_count >= cli->cl_max_pages_per_rpc)
1497                         break;
1498         }
1499
1500         osc_wake_cache_waiters(cli);
1501
1502         if (page_count == 0)
1503                 RETURN(0);
1504
1505         loi_list_maint(cli, loi);
1506         spin_unlock(&cli->cl_loi_list_lock);
1507
1508         request = osc_build_req(cli, &rpc_list, page_count, cmd);
1509         if (IS_ERR(request)) {
1510                 /* this should happen rarely and is pretty bad, it makes the
1511                  * pending list not follow the dirty order */
1512                 spin_lock(&cli->cl_loi_list_lock);
1513                 list_for_each_safe(pos, tmp, &rpc_list) {
1514                         oap = list_entry(pos, struct osc_async_page,
1515                                          oap_rpc_item);
1516                         list_del_init(&oap->oap_rpc_item);
1517
1518                         /* queued sync pages can be torn down while the pages
1519                          * were between the pending list and the rpc */
1520                         if (oap->oap_interrupted) {
1521                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1522                                 osc_ap_completion(cli, NULL, oap, 0,
1523                                                   oap->oap_count);
1524                                 continue;
1525                         }
1526
1527                         /* put the page back in the loi/lop lists */
1528                         list_add_tail(&oap->oap_pending_item,
1529                                       &lop->lop_pending);
1530                         lop_update_pending(cli, lop, cmd, 1);
1531                         if (oap->oap_async_flags & ASYNC_URGENT)
1532                                 list_add(&oap->oap_urgent_item,
1533                                          &lop->lop_urgent);
1534                 }
1535                 loi_list_maint(cli, loi);
1536                 RETURN(PTR_ERR(request));
1537         }
1538
1539         LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1540         aa = (struct osc_brw_async_args *)&request->rq_async_args;
1541         INIT_LIST_HEAD(&aa->aa_oaps);
1542         list_splice(&rpc_list, &aa->aa_oaps);
1543         INIT_LIST_HEAD(&rpc_list);
1544
1545 #ifdef __KERNEL__
1546         if (cmd == OBD_BRW_READ) {
1547                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1548                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1549         } else {
1550                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1551                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1552                                  cli->cl_w_in_flight);
1553         }
1554 #endif
1555
1556         spin_lock(&cli->cl_loi_list_lock);
1557
1558         if (cmd == OBD_BRW_READ)
1559                 cli->cl_r_in_flight++;
1560         else
1561                 cli->cl_w_in_flight++;
1562         /* queued sync pages can be torn down while the pages
1563          * were between the pending list and the rpc */
1564         list_for_each(pos, &aa->aa_oaps) {
1565                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1566                 if (oap->oap_interrupted) {
1567                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1568                                oap, request);
1569                         ptlrpc_mark_interrupted(request);
1570                         break;
1571                 }
1572         }
1573
1574         CDEBUG(D_INODE, "req %p: %d pages, aa %p.  now %dr/%dw in flight\n",
1575                         request, page_count, aa, cli->cl_r_in_flight,
1576                         cli->cl_w_in_flight);
1577
1578         oap->oap_request = ptlrpc_request_addref(request);
1579         request->rq_interpret_reply = brw_interpret_oap;
1580         ptlrpcd_add_req(request);
1581         RETURN(1);
1582 }
1583
1584 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1585                          int cmd)
1586 {
1587         int optimal;
1588         ENTRY;
1589
1590         if (lop->lop_num_pending == 0)
1591                 RETURN(0);
1592
1593         /* if we have an invalid import we want to drain the queued pages
1594          * by forcing them through rpcs that immediately fail and complete
1595          * the pages.  recovery relies on this to empty the queued pages
1596          * before canceling the locks and evicting down the llite pages */
1597         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1598                 RETURN(1);
1599
1600         /* stream rpcs in queue order as long as as there is an urgent page
1601          * queued.  this is our cheap solution for good batching in the case
1602          * where writepage marks some random page in the middle of the file as
1603          * urgent because of, say, memory pressure */
1604         if (!list_empty(&lop->lop_urgent))
1605                 RETURN(1);
1606
1607         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1608         optimal = cli->cl_max_pages_per_rpc;
1609         if (cmd == OBD_BRW_WRITE) {
1610                 /* trigger a write rpc stream as long as there are dirtiers
1611                  * waiting for space.  as they're waiting, they're not going to
1612                  * create more pages to coallesce with what's waiting.. */
1613                 if (!list_empty(&cli->cl_cache_waiters))
1614                         RETURN(1);
1615
1616                 /* *2 to avoid triggering rpcs that would want to include pages
1617                  * that are being queued but which can't be made ready until
1618                  * the queuer finishes with the page. this is a wart for
1619                  * llite::commit_write() */
1620                 optimal += 16;
1621         }
1622         if (lop->lop_num_pending >= optimal)
1623                 RETURN(1);
1624
1625         RETURN(0);
1626 }
1627
1628 static void on_list(struct list_head *item, struct list_head *list,
1629                     int should_be_on)
1630 {
1631         if (list_empty(item) && should_be_on)
1632                 list_add_tail(item, list);
1633         else if (!list_empty(item) && !should_be_on)
1634                 list_del_init(item);
1635 }
1636
1637 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1638  * can find pages to build into rpcs quickly */
1639 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1640 {
1641         on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1642                 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1643                 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1644
1645         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1646                 loi->loi_write_lop.lop_num_pending);
1647
1648         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1649                 loi->loi_read_lop.lop_num_pending);
1650 }
1651
1652 #define LOI_DEBUG(LOI, STR, args...)                                     \
1653         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
1654                !list_empty(&(LOI)->loi_cli_item),                        \
1655                (LOI)->loi_write_lop.lop_num_pending,                     \
1656                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
1657                (LOI)->loi_read_lop.lop_num_pending,                      \
1658                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
1659                args)                                                     \
1660
1661 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1662 {
1663         ENTRY;
1664         /* first return all objects which we already know to have
1665          * pages ready to be stuffed into rpcs */
1666         if (!list_empty(&cli->cl_loi_ready_list))
1667                 RETURN(list_entry(cli->cl_loi_ready_list.next,
1668                                   struct lov_oinfo, loi_cli_item));
1669
1670         /* then if we have cache waiters, return all objects with queued
1671          * writes.  This is especially important when many small files
1672          * have filled up the cache and not been fired into rpcs because
1673          * they don't pass the nr_pending/object threshhold */
1674         if (!list_empty(&cli->cl_cache_waiters) &&
1675             !list_empty(&cli->cl_loi_write_list))
1676                 RETURN(list_entry(cli->cl_loi_write_list.next,
1677                                   struct lov_oinfo, loi_write_item));
1678
1679         /* then return all queued objects when we have an invalid import
1680          * so that they get flushed */
1681         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1682                 if (!list_empty(&cli->cl_loi_write_list))
1683                         RETURN(list_entry(cli->cl_loi_write_list.next,
1684                                           struct lov_oinfo, loi_write_item));
1685                 if (!list_empty(&cli->cl_loi_read_list))
1686                         RETURN(list_entry(cli->cl_loi_read_list.next,
1687                                           struct lov_oinfo, loi_read_item));
1688         }
1689         RETURN(NULL);
1690 }
1691
1692 /* called with the loi list lock held */
1693 static void osc_check_rpcs(struct client_obd *cli)
1694 {
1695         struct lov_oinfo *loi;
1696         int rc = 0, race_counter = 0;
1697         ENTRY;
1698
1699         while ((loi = osc_next_loi(cli)) != NULL) {
1700                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
1701
1702                 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
1703                         break;
1704
1705                 /* attempt some read/write balancing by alternating between
1706                  * reads and writes in an object.  The makes_rpc checks here
1707                  * would be redundant if we were getting read/write work items
1708                  * instead of objects.  we don't want send_oap_rpc to drain a
1709                  * partial read pending queue when we're given this object to
1710                  * do io on writes while there are cache waiters */
1711                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1712                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1713                                               &loi->loi_write_lop);
1714                         if (rc < 0)
1715                                 break;
1716                         if (rc > 0)
1717                                 race_counter = 0;
1718                         else
1719                                 race_counter++;
1720                 }
1721                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1722                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1723                                               &loi->loi_read_lop);
1724                         if (rc < 0)
1725                                 break;
1726                         if (rc > 0)
1727                                 race_counter = 0;
1728                         else
1729                                 race_counter++;
1730                 }
1731
1732                 /* attempt some inter-object balancing by issueing rpcs
1733                  * for each object in turn */
1734                 if (!list_empty(&loi->loi_cli_item))
1735                         list_del_init(&loi->loi_cli_item);
1736                 if (!list_empty(&loi->loi_write_item))
1737                         list_del_init(&loi->loi_write_item);
1738                 if (!list_empty(&loi->loi_read_item))
1739                         list_del_init(&loi->loi_read_item);
1740
1741                 loi_list_maint(cli, loi);
1742
1743                 /* send_oap_rpc fails with 0 when make_ready tells it to
1744                  * back off.  llite's make_ready does this when it tries
1745                  * to lock a page queued for write that is already locked.
1746                  * we want to try sending rpcs from many objects, but we
1747                  * don't want to spin failing with 0.  */
1748                 if (race_counter == 10)
1749                         break;
1750         }
1751         EXIT;
1752 }
1753
1754 /* we're trying to queue a page in the osc so we're subject to the
1755  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1756  * If the osc's queued pages are already at that limit, then we want to sleep
1757  * until there is space in the osc's queue for us.  We also may be waiting for
1758  * write credits from the OST if there are RPCs in flight that may return some
1759  * before we fall back to sync writes.
1760  *
1761  * We need this know our allocation was granted in the presence of signals */
1762 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1763 {
1764         int rc;
1765         ENTRY;
1766         spin_lock(&cli->cl_loi_list_lock);
1767         rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
1768         spin_unlock(&cli->cl_loi_list_lock);
1769         RETURN(rc);
1770 };
1771
1772 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1773  * grant or cache space. */
1774 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1775                            struct osc_async_page *oap)
1776 {
1777         struct osc_cache_waiter ocw;
1778         struct l_wait_info lwi = { 0 };
1779         struct timeval start, stop;
1780
1781         CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1782                cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1783                cli->cl_avail_grant);
1784
1785         if (cli->cl_dirty_max < PAGE_SIZE)
1786                 return(-EDQUOT);
1787
1788         /* Hopefully normal case - cache space and write credits available */
1789         if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1790             cli->cl_avail_grant >= PAGE_SIZE) {
1791                 /* account for ourselves */
1792                 osc_consume_write_grant(cli, oap);
1793                 return(0);
1794         }
1795
1796         /* Make sure that there are write rpcs in flight to wait for.  This
1797          * is a little silly as this object may not have any pending but
1798          * other objects sure might. */
1799         if (cli->cl_w_in_flight) {
1800                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1801                 init_waitqueue_head(&ocw.ocw_waitq);
1802                 ocw.ocw_oap = oap;
1803                 ocw.ocw_rc = 0;
1804
1805                 loi_list_maint(cli, loi);
1806                 osc_check_rpcs(cli);
1807                 spin_unlock(&cli->cl_loi_list_lock);
1808
1809                 CDEBUG(0, "sleeping for cache space\n");
1810                 do_gettimeofday(&start);
1811                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1812                 do_gettimeofday(&stop);
1813                 spin_lock(&cli->cl_loi_list_lock);
1814                 lprocfs_stime_record(&cli->cl_enter_stime, &stop, &start);
1815                 if (!list_empty(&ocw.ocw_entry)) {
1816                         list_del(&ocw.ocw_entry);
1817                         RETURN(-EINTR);
1818                 }
1819                 RETURN(ocw.ocw_rc);
1820         }
1821
1822         RETURN(-EDQUOT);
1823 }
1824
1825 /* the companion to enter_cache, called when an oap is no longer part of the
1826  * dirty accounting.. so writeback completes or truncate happens before writing
1827  * starts.  must be called with the loi lock held. */
1828 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1829                            int sent)
1830 {
1831         ENTRY;
1832
1833         if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1834                 EXIT;
1835                 return;
1836         }
1837
1838         oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1839         cli->cl_dirty -= PAGE_SIZE;
1840         if (!sent) {
1841                 cli->cl_lost_grant += PAGE_SIZE;
1842                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1843                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1844         }
1845
1846         EXIT;
1847 }
1848
1849 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1850                         struct lov_oinfo *loi, struct page *page,
1851                         obd_off offset, struct obd_async_page_ops *ops,
1852                         void *data, void **res)
1853 {
1854         struct osc_async_page *oap;
1855         ENTRY;
1856
1857         OBD_ALLOC(oap, sizeof(*oap));
1858         if (oap == NULL)
1859                 return -ENOMEM;
1860
1861         oap->oap_magic = OAP_MAGIC;
1862         oap->oap_cli = &exp->exp_obd->u.cli;
1863         oap->oap_loi = loi;
1864
1865         oap->oap_caller_ops = ops;
1866         oap->oap_caller_data = data;
1867
1868         oap->oap_page = page;
1869         oap->oap_obj_off = offset;
1870
1871         INIT_LIST_HEAD(&oap->oap_pending_item);
1872         INIT_LIST_HEAD(&oap->oap_urgent_item);
1873         INIT_LIST_HEAD(&oap->oap_rpc_item);
1874
1875         oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1876
1877         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1878         *res = oap;
1879         RETURN(0);
1880 }
1881
1882 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1883                               struct lov_oinfo *loi, void *cookie,
1884                               int cmd, obd_off off, int count,
1885                               obd_flags brw_flags, enum async_flags async_flags)
1886 {
1887         struct client_obd *cli = &exp->exp_obd->u.cli;
1888         struct osc_async_page *oap;
1889         struct loi_oap_pages *lop;
1890         int rc;
1891         ENTRY;
1892
1893         oap = OAP_FROM_COOKIE(cookie);
1894
1895         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1896                 RETURN(-EIO);
1897
1898         if (!list_empty(&oap->oap_pending_item) ||
1899             !list_empty(&oap->oap_urgent_item) ||
1900             !list_empty(&oap->oap_rpc_item))
1901                 RETURN(-EBUSY);
1902
1903         if (loi == NULL)
1904                 loi = &lsm->lsm_oinfo[0];
1905
1906         spin_lock(&cli->cl_loi_list_lock);
1907
1908         oap->oap_cmd = cmd;
1909         oap->oap_async_flags = async_flags;
1910         oap->oap_page_off = off;
1911         oap->oap_count = count;
1912         oap->oap_brw_flags = brw_flags;
1913
1914         if (cmd == OBD_BRW_WRITE) {
1915                 rc = osc_enter_cache(cli, loi, oap);
1916                 if (rc) {
1917                         spin_unlock(&cli->cl_loi_list_lock);
1918                         RETURN(rc);
1919                 }
1920                 lop = &loi->loi_write_lop;
1921         } else {
1922                 lop = &loi->loi_read_lop;
1923         }
1924
1925         if (oap->oap_async_flags & ASYNC_URGENT)
1926                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1927         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1928         lop_update_pending(cli, lop, cmd, 1);
1929
1930         loi_list_maint(cli, loi);
1931
1932         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
1933                   cmd);
1934
1935         osc_check_rpcs(cli);
1936         spin_unlock(&cli->cl_loi_list_lock);
1937
1938         RETURN(0);
1939 }
1940
1941 /* aka (~was & now & flag), but this is more clear :) */
1942 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
1943
1944 static int osc_set_async_flags(struct obd_export *exp,
1945                                struct lov_stripe_md *lsm,
1946                                struct lov_oinfo *loi, void *cookie,
1947                                obd_flags async_flags)
1948 {
1949         struct client_obd *cli = &exp->exp_obd->u.cli;
1950         struct loi_oap_pages *lop;
1951         struct osc_async_page *oap;
1952         int rc = 0;
1953         ENTRY;
1954
1955         oap = OAP_FROM_COOKIE(cookie);
1956
1957         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1958                 RETURN(-EIO);
1959
1960         if (loi == NULL)
1961                 loi = &lsm->lsm_oinfo[0];
1962
1963         if (oap->oap_cmd == OBD_BRW_WRITE) {
1964                 lop = &loi->loi_write_lop;
1965         } else {
1966                 lop = &loi->loi_read_lop;
1967         }
1968
1969         spin_lock(&cli->cl_loi_list_lock);
1970
1971         if (list_empty(&oap->oap_pending_item))
1972                 GOTO(out, rc = -EINVAL);
1973
1974         if ((oap->oap_async_flags & async_flags) == async_flags)
1975                 GOTO(out, rc = 0);
1976
1977         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
1978                 oap->oap_async_flags |= ASYNC_READY;
1979
1980         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
1981                 if (list_empty(&oap->oap_rpc_item)) {
1982                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1983                         loi_list_maint(cli, loi);
1984                 }
1985         }
1986
1987         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
1988                         oap->oap_async_flags);
1989 out:
1990         osc_check_rpcs(cli);
1991         spin_unlock(&cli->cl_loi_list_lock);
1992         RETURN(rc);
1993 }
1994
1995 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1996                              struct lov_oinfo *loi,
1997                              struct obd_io_group *oig, void *cookie,
1998                              int cmd, obd_off off, int count,
1999                              obd_flags brw_flags,
2000                              obd_flags async_flags)
2001 {
2002         struct client_obd *cli = &exp->exp_obd->u.cli;
2003         struct osc_async_page *oap;
2004         struct loi_oap_pages *lop;
2005         ENTRY;
2006
2007         oap = OAP_FROM_COOKIE(cookie);
2008
2009         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2010                 RETURN(-EIO);
2011
2012         if (!list_empty(&oap->oap_pending_item) ||
2013             !list_empty(&oap->oap_urgent_item) ||
2014             !list_empty(&oap->oap_rpc_item))
2015                 RETURN(-EBUSY);
2016
2017         if (loi == NULL)
2018                 loi = &lsm->lsm_oinfo[0];
2019
2020         spin_lock(&cli->cl_loi_list_lock);
2021
2022         oap->oap_cmd = cmd;
2023         oap->oap_page_off = off;
2024         oap->oap_count = count;
2025         oap->oap_brw_flags = brw_flags;
2026         oap->oap_async_flags = async_flags;
2027
2028         if (cmd == OBD_BRW_WRITE)
2029                 lop = &loi->loi_write_lop;
2030         else
2031                 lop = &loi->loi_read_lop;
2032
2033         list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2034         if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2035                 oap->oap_oig = oig;
2036                 oig_add_one(oig, &oap->oap_occ);
2037         }
2038
2039         LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
2040
2041         spin_unlock(&cli->cl_loi_list_lock);
2042
2043         RETURN(0);
2044 }
2045
2046 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2047                                  struct loi_oap_pages *lop, int cmd)
2048 {
2049         struct list_head *pos, *tmp;
2050         struct osc_async_page *oap;
2051
2052         list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2053                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2054                 list_del(&oap->oap_pending_item);
2055                 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2056                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2057                 lop_update_pending(cli, lop, cmd, 1);
2058         }
2059         loi_list_maint(cli, loi);
2060 }
2061
2062 static int osc_trigger_group_io(struct obd_export *exp,
2063                                 struct lov_stripe_md *lsm,
2064                                 struct lov_oinfo *loi,
2065                                 struct obd_io_group *oig)
2066 {
2067         struct client_obd *cli = &exp->exp_obd->u.cli;
2068         ENTRY;
2069
2070         if (loi == NULL)
2071                 loi = &lsm->lsm_oinfo[0];
2072
2073         spin_lock(&cli->cl_loi_list_lock);
2074
2075         osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2076         osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2077
2078         osc_check_rpcs(cli);
2079         spin_unlock(&cli->cl_loi_list_lock);
2080
2081         RETURN(0);
2082 }
2083
2084 static int osc_teardown_async_page(struct obd_export *exp,
2085                                    struct lov_stripe_md *lsm,
2086                                    struct lov_oinfo *loi, void *cookie)
2087 {
2088         struct client_obd *cli = &exp->exp_obd->u.cli;
2089         struct loi_oap_pages *lop;
2090         struct osc_async_page *oap;
2091         int rc = 0;
2092         ENTRY;
2093
2094         oap = OAP_FROM_COOKIE(cookie);
2095
2096         if (loi == NULL)
2097                 loi = &lsm->lsm_oinfo[0];
2098
2099         if (oap->oap_cmd == OBD_BRW_WRITE) {
2100                 lop = &loi->loi_write_lop;
2101         } else {
2102                 lop = &loi->loi_read_lop;
2103         }
2104
2105         spin_lock(&cli->cl_loi_list_lock);
2106
2107         if (!list_empty(&oap->oap_rpc_item))
2108                 GOTO(out, rc = -EBUSY);
2109
2110         osc_exit_cache(cli, oap, 0);
2111         osc_wake_cache_waiters(cli);
2112
2113         if (!list_empty(&oap->oap_urgent_item)) {
2114                 list_del_init(&oap->oap_urgent_item);
2115                 oap->oap_async_flags &= ~ASYNC_URGENT;
2116         }
2117         if (!list_empty(&oap->oap_pending_item)) {
2118                 list_del_init(&oap->oap_pending_item);
2119                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2120         }
2121         loi_list_maint(cli, loi);
2122
2123         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2124 out:
2125         spin_unlock(&cli->cl_loi_list_lock);
2126         if (rc == 0)
2127                 OBD_FREE(oap, sizeof(*oap));
2128         RETURN(rc);
2129 }
2130
2131 #ifdef __KERNEL__
2132 /* Note: caller will lock/unlock, and set uptodate on the pages */
2133 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2134 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2135                            struct lov_stripe_md *lsm, obd_count page_count,
2136                            struct brw_page *pga)
2137 {
2138         struct ptlrpc_request *request = NULL;
2139         struct ost_body *body;
2140         struct niobuf_remote *nioptr;
2141         struct obd_ioobj *iooptr;
2142         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2143         int swab;
2144         ENTRY;
2145
2146         /* XXX does not handle 'new' brw protocol */
2147
2148         size[1] = sizeof(struct obd_ioobj);
2149         size[2] = page_count * sizeof(*nioptr);
2150
2151         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2152                                   OST_SAN_READ, 3, size, NULL);
2153         if (!request)
2154                 RETURN(-ENOMEM);
2155
2156         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2157         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2158         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2159                                 sizeof(*nioptr) * page_count);
2160
2161         memcpy(&body->oa, oa, sizeof(body->oa));
2162
2163         obdo_to_ioobj(oa, iooptr);
2164         iooptr->ioo_bufcnt = page_count;
2165
2166         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2167                 LASSERT(PageLocked(pga[mapped].pg));
2168                 LASSERT(mapped == 0 ||
2169                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2170
2171                 nioptr->offset = pga[mapped].disk_offset;
2172                 nioptr->len    = pga[mapped].count;
2173                 nioptr->flags  = pga[mapped].flag;
2174         }
2175
2176         size[1] = page_count * sizeof(*nioptr);
2177         request->rq_replen = lustre_msg_size(2, size);
2178
2179         rc = ptlrpc_queue_wait(request);
2180         if (rc)
2181                 GOTO(out_req, rc);
2182
2183         body = lustre_swab_repbuf(request, 0, sizeof(*body),
2184                                   lustre_swab_ost_body);
2185         if (body == NULL) {
2186                 CERROR("Can't unpack body\n");
2187                 GOTO(out_req, rc = -EPROTO);
2188         }
2189
2190         memcpy(oa, &body->oa, sizeof(*oa));
2191
2192         swab = lustre_msg_swabbed(request->rq_repmsg);
2193         LASSERT_REPSWAB(request, 1);
2194         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2195         if (!nioptr) {
2196                 /* nioptr missing or short */
2197                 GOTO(out_req, rc = -EPROTO);
2198         }
2199
2200         /* actual read */
2201         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2202                 struct page *page = pga[mapped].pg;
2203                 struct buffer_head *bh;
2204                 kdev_t dev;
2205
2206                 if (swab)
2207                         lustre_swab_niobuf_remote (nioptr);
2208
2209                 /* got san device associated */
2210                 LASSERT(exp->exp_obd != NULL);
2211                 dev = exp->exp_obd->u.cli.cl_sandev;
2212
2213                 /* hole */
2214                 if (!nioptr->offset) {
2215                         CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2216                                         page->mapping->host->i_ino,
2217                                         page->index);
2218                         memset(page_address(page), 0, PAGE_SIZE);
2219                         continue;
2220                 }
2221
2222                 if (!page->buffers) {
2223                         create_empty_buffers(page, dev, PAGE_SIZE);
2224                         bh = page->buffers;
2225
2226                         clear_bit(BH_New, &bh->b_state);
2227                         set_bit(BH_Mapped, &bh->b_state);
2228                         bh->b_blocknr = (unsigned long)nioptr->offset;
2229
2230                         clear_bit(BH_Uptodate, &bh->b_state);
2231
2232                         ll_rw_block(READ, 1, &bh);
2233                 } else {
2234                         bh = page->buffers;
2235
2236                         /* if buffer already existed, it must be the
2237                          * one we mapped before, check it */
2238                         LASSERT(!test_bit(BH_New, &bh->b_state));
2239                         LASSERT(test_bit(BH_Mapped, &bh->b_state));
2240                         LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2241
2242                         /* wait it's io completion */
2243                         if (test_bit(BH_Lock, &bh->b_state))
2244                                 wait_on_buffer(bh);
2245
2246                         if (!test_bit(BH_Uptodate, &bh->b_state))
2247                                 ll_rw_block(READ, 1, &bh);
2248                 }
2249
2250
2251                 /* must do syncronous write here */
2252                 wait_on_buffer(bh);
2253                 if (!buffer_uptodate(bh)) {
2254                         /* I/O error */
2255                         rc = -EIO;
2256                         goto out_req;
2257                 }
2258         }
2259
2260 out_req:
2261         ptlrpc_req_finished(request);
2262         RETURN(rc);
2263 }
2264
2265 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2266                             struct lov_stripe_md *lsm, obd_count page_count,
2267                             struct brw_page *pga)
2268 {
2269         struct ptlrpc_request *request = NULL;
2270         struct ost_body *body;
2271         struct niobuf_remote *nioptr;
2272         struct obd_ioobj *iooptr;
2273         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2274         int swab;
2275         ENTRY;
2276
2277         size[1] = sizeof(struct obd_ioobj);
2278         size[2] = page_count * sizeof(*nioptr);
2279
2280         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2281                                   OST_SAN_WRITE, 3, size, NULL);
2282         if (!request)
2283                 RETURN(-ENOMEM);
2284
2285         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2286         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2287         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2288                                 sizeof (*nioptr) * page_count);
2289
2290         memcpy(&body->oa, oa, sizeof(body->oa));
2291
2292         obdo_to_ioobj(oa, iooptr);
2293         iooptr->ioo_bufcnt = page_count;
2294
2295         /* pack request */
2296         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2297                 LASSERT(PageLocked(pga[mapped].pg));
2298                 LASSERT(mapped == 0 ||
2299                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2300
2301                 nioptr->offset = pga[mapped].disk_offset;
2302                 nioptr->len    = pga[mapped].count;
2303                 nioptr->flags  = pga[mapped].flag;
2304         }
2305
2306         size[1] = page_count * sizeof(*nioptr);
2307         request->rq_replen = lustre_msg_size(2, size);
2308
2309         rc = ptlrpc_queue_wait(request);
2310         if (rc)
2311                 GOTO(out_req, rc);
2312
2313         swab = lustre_msg_swabbed (request->rq_repmsg);
2314         LASSERT_REPSWAB (request, 1);
2315         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2316         if (!nioptr) {
2317                 CERROR("absent/short niobuf array\n");
2318                 GOTO(out_req, rc = -EPROTO);
2319         }
2320
2321         /* actual write */
2322         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2323                 struct page *page = pga[mapped].pg;
2324                 struct buffer_head *bh;
2325                 kdev_t dev;
2326
2327                 if (swab)
2328                         lustre_swab_niobuf_remote (nioptr);
2329
2330                 /* got san device associated */
2331                 LASSERT(exp->exp_obd != NULL);
2332                 dev = exp->exp_obd->u.cli.cl_sandev;
2333
2334                 if (!page->buffers) {
2335                         create_empty_buffers(page, dev, PAGE_SIZE);
2336                 } else {
2337                         /* checking */
2338                         LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2339                         LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2340                         LASSERT(page->buffers->b_blocknr ==
2341                                 (unsigned long)nioptr->offset);
2342                 }
2343                 bh = page->buffers;
2344
2345                 LASSERT(bh);
2346
2347                 /* if buffer locked, wait it's io completion */
2348                 if (test_bit(BH_Lock, &bh->b_state))
2349                         wait_on_buffer(bh);
2350
2351                 clear_bit(BH_New, &bh->b_state);
2352                 set_bit(BH_Mapped, &bh->b_state);
2353
2354                 /* override the block nr */
2355                 bh->b_blocknr = (unsigned long)nioptr->offset;
2356
2357                 /* we are about to write it, so set it
2358                  * uptodate/dirty
2359                  * page lock should garentee no race condition here */
2360                 set_bit(BH_Uptodate, &bh->b_state);
2361                 set_bit(BH_Dirty, &bh->b_state);
2362
2363                 ll_rw_block(WRITE, 1, &bh);
2364
2365                 /* must do syncronous write here */
2366                 wait_on_buffer(bh);
2367                 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2368                         /* I/O error */
2369                         rc = -EIO;
2370                         goto out_req;
2371                 }
2372         }
2373
2374 out_req:
2375         ptlrpc_req_finished(request);
2376         RETURN(rc);
2377 }
2378
2379 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2380                       struct lov_stripe_md *lsm, obd_count page_count,
2381                       struct brw_page *pga, struct obd_trans_info *oti)
2382 {
2383         ENTRY;
2384
2385         while (page_count) {
2386                 obd_count pages_per_brw;
2387                 int rc;
2388
2389                 if (page_count > PTLRPC_MAX_BRW_PAGES)
2390                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2391                 else
2392                         pages_per_brw = page_count;
2393
2394                 if (cmd & OBD_BRW_WRITE)
2395                         rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2396                 else
2397                         rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2398
2399                 if (rc != 0)
2400                         RETURN(rc);
2401
2402                 page_count -= pages_per_brw;
2403                 pga += pages_per_brw;
2404         }
2405         RETURN(0);
2406 }
2407 #endif
2408 #endif
2409
2410 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2411 {
2412         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2413
2414         if (lock == NULL) {
2415                 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2416                 return;
2417         }
2418
2419         lock_res_and_lock(lock);
2420 #ifdef __KERNEL__
2421         if (lock->l_ast_data && lock->l_ast_data != data) {
2422                 struct inode *new_inode = data;
2423                 struct inode *old_inode = lock->l_ast_data;
2424                 if (!(old_inode->i_state & I_FREEING))
2425                         LDLM_ERROR(lock, "inconsistent l_ast_data found");
2426                 LASSERTF(old_inode->i_state & I_FREEING,
2427                          "Found existing inode %p/%lu/%u state %lu in lock: "
2428                          "setting data to %p/%lu/%u\n", old_inode,
2429                          old_inode->i_ino, old_inode->i_generation,
2430                          old_inode->i_state,
2431                          new_inode, new_inode->i_ino, new_inode->i_generation);
2432         }
2433 #endif
2434         lock->l_ast_data = data;
2435         unlock_res_and_lock(lock);
2436         LDLM_LOCK_PUT(lock);
2437 }
2438
2439 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2440                              ldlm_iterator_t replace, void *data)
2441 {
2442         struct ldlm_res_id res_id = { .name = {0} };
2443         struct obd_device *obd = class_exp2obd(exp);
2444
2445         res_id.name[0] = lsm->lsm_object_id;
2446         res_id.name[2] = lsm->lsm_object_gr;
2447         ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2448         return 0;
2449 }
2450
2451 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2452                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2453                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2454                        void *data, __u32 lvb_len, void *lvb_swabber,
2455                        struct lustre_handle *lockh)
2456 {
2457         struct obd_device *obd = exp->exp_obd;
2458         struct ldlm_res_id res_id = { .name = {0} };
2459         struct ost_lvb lvb;
2460         struct ldlm_reply *rep;
2461         struct ptlrpc_request *req = NULL;
2462         int rc;
2463         ENTRY;
2464
2465         res_id.name[0] = lsm->lsm_object_id;
2466         res_id.name[2] = lsm->lsm_object_gr;
2467
2468         /* Filesystem lock extents are extended to page boundaries so that
2469          * dealing with the page cache is a little smoother.  */
2470         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2471         policy->l_extent.end |= ~PAGE_MASK;
2472
2473         if (lsm->lsm_oinfo->loi_kms_valid == 0)
2474                 goto no_match;
2475
2476         /* Next, search for already existing extent locks that will cover us */
2477         rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2478                              lockh);
2479         if (rc == 1) {
2480                 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2481                         /* return immediately if no credential held */
2482                         ldlm_lock_decref(lockh, mode);
2483                         RETURN(-EACCES);
2484                 }
2485
2486                 osc_set_data_with_check(lockh, data);
2487                 if (*flags & LDLM_FL_HAS_INTENT) {
2488                         /* I would like to be able to ASSERT here that rss <=
2489                          * kms, but I can't, for reasons which are explained in
2490                          * lov_enqueue() */
2491                 }
2492                 /* We already have a lock, and it's referenced */
2493                 RETURN(ELDLM_OK);
2494         }
2495
2496         /* If we're trying to read, we also search for an existing PW lock.  The
2497          * VFS and page cache already protect us locally, so lots of readers/
2498          * writers can share a single PW lock.
2499          *
2500          * There are problems with conversion deadlocks, so instead of
2501          * converting a read lock to a write lock, we'll just enqueue a new
2502          * one.
2503          *
2504          * At some point we should cancel the read lock instead of making them
2505          * send us a blocking callback, but there are problems with canceling
2506          * locks out from other users right now, too. */
2507
2508         if (mode == LCK_PR) {
2509                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2510                                      policy, LCK_PW, lockh);
2511                 if (rc == 1) {
2512                         if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2513                                 /* return immediately if no credential held */
2514                                 ldlm_lock_decref(lockh, LCK_PW);
2515                                 RETURN(-EACCES);
2516                         }
2517
2518                         /* FIXME: This is not incredibly elegant, but it might
2519                          * be more elegant than adding another parameter to
2520                          * lock_match.  I want a second opinion. */
2521                         ldlm_lock_addref(lockh, LCK_PR);
2522                         ldlm_lock_decref(lockh, LCK_PW);
2523                         osc_set_data_with_check(lockh, data);
2524                         RETURN(ELDLM_OK);
2525                 }
2526         }
2527         if (mode == LCK_PW) {
2528                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2529                                      policy, LCK_PR, lockh);
2530                 if (rc == 1) {
2531                         rc = ldlm_cli_convert(lockh, mode, flags);
2532                         if (!rc) {
2533                                 /* Update readers/writers accounting */
2534                                 ldlm_lock_addref(lockh, LCK_PW);
2535                                 ldlm_lock_decref(lockh, LCK_PR);
2536                                 osc_set_data_with_check(lockh, data);
2537                                 RETURN(ELDLM_OK);
2538                         }
2539                         /* If the conversion failed, we need to drop refcount
2540                            on matched lock before we get new one */
2541                         /* XXX Won't it save us some efforts if we cancel PR
2542                            lock here? We are going to take PW lock anyway and it
2543                            will invalidate PR lock */
2544                         ldlm_lock_decref(lockh, LCK_PR);
2545                         if (rc != EDEADLOCK) {
2546                                 RETURN(rc);
2547                         }
2548                 }
2549         }
2550
2551 no_match:
2552         if (*flags & LDLM_FL_HAS_INTENT) {
2553                 int size[2] = {0, sizeof(struct ldlm_request)};
2554
2555                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
2556                                       LDLM_ENQUEUE, 2, size, NULL);
2557                 if (req == NULL)
2558                         RETURN(-ENOMEM);
2559
2560                 size[0] = sizeof(*rep);
2561                 size[1] = sizeof(lvb);
2562                 req->rq_replen = lustre_msg_size(2, size);
2563         }
2564         rc = ldlm_cli_enqueue(exp, req, obd->obd_namespace, res_id, type,
2565                               policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2566                               &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2567         if (req != NULL) {
2568                 if (rc == ELDLM_LOCK_ABORTED) {
2569                         /* swabbed by ldlm_cli_enqueue() */
2570                         LASSERT_REPSWABBED(req, 0);
2571                         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
2572                         LASSERT(rep != NULL);
2573                         if (rep->lock_policy_res1)
2574                                 rc = rep->lock_policy_res1;
2575                 }
2576                 ptlrpc_req_finished(req);
2577         }
2578
2579         if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2580                 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2581                        lvb.lvb_size, lvb.lvb_blocks);
2582                 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2583                 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2584         }
2585
2586         RETURN(rc);
2587 }
2588
2589 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2590                      __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2591                      int *flags, void *data, struct lustre_handle *lockh)
2592 {
2593         struct ldlm_res_id res_id = { .name = {0} };
2594         struct obd_device *obd = exp->exp_obd;
2595         int rc;
2596         ENTRY;
2597
2598         res_id.name[0] = lsm->lsm_object_id;
2599         res_id.name[2] = lsm->lsm_object_gr;
2600
2601         OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2602
2603         /* Filesystem lock extents are extended to page boundaries so that
2604          * dealing with the page cache is a little smoother */
2605         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2606         policy->l_extent.end |= ~PAGE_MASK;
2607
2608         /* Next, search for already existing extent locks that will cover us */
2609         rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2610                              policy, mode, lockh);
2611         if (rc) {
2612                // if (!(*flags & LDLM_FL_TEST_LOCK))
2613                         osc_set_data_with_check(lockh, data);
2614                 RETURN(rc);
2615         }
2616         /* If we're trying to read, we also search for an existing PW lock.  The
2617          * VFS and page cache already protect us locally, so lots of readers/
2618          * writers can share a single PW lock. */
2619         if (mode == LCK_PR) {
2620                 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2621                                      policy, LCK_PW, lockh);
2622                 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2623                         /* FIXME: This is not incredibly elegant, but it might
2624                          * be more elegant than adding another parameter to
2625                          * lock_match.  I want a second opinion. */
2626                         osc_set_data_with_check(lockh, data);
2627                         ldlm_lock_addref(lockh, LCK_PR);
2628                         ldlm_lock_decref(lockh, LCK_PW);
2629                 }
2630         }
2631         RETURN(rc);
2632 }
2633
2634 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2635                       __u32 mode, struct lustre_handle *lockh)
2636 {
2637         ENTRY;
2638
2639         if (mode == LCK_GROUP)
2640                 ldlm_lock_decref_and_cancel(lockh, mode);
2641         else
2642                 ldlm_lock_decref(lockh, mode);
2643
2644         RETURN(0);
2645 }
2646
2647 static int osc_cancel_unused(struct obd_export *exp,
2648                              struct lov_stripe_md *lsm,
2649                              int flags, void *opaque)
2650 {
2651         struct obd_device *obd = class_exp2obd(exp);
2652         struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2653
2654         if (lsm != NULL) {
2655                 res_id.name[0] = lsm->lsm_object_id;
2656                 res_id.name[2] = lsm->lsm_object_gr;
2657                 resp = &res_id;
2658         }
2659
2660         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2661 }
2662
2663 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2664                       unsigned long max_age)
2665 {
2666         struct obd_statfs *msfs;
2667         struct ptlrpc_request *request;
2668         int rc, size = sizeof(*osfs);
2669         ENTRY;
2670
2671         /* We could possibly pass max_age in the request (as an absolute
2672          * timestamp or a "seconds.usec ago") so the target can avoid doing
2673          * extra calls into the filesystem if that isn't necessary (e.g.
2674          * during mount that would help a bit).  Having relative timestamps
2675          * is not so great if request processing is slow, while absolute
2676          * timestamps are not ideal because they need time synchronization. */
2677         request = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OBD_VERSION,
2678                                   OST_STATFS, 0, NULL, NULL);
2679         if (!request)
2680                 RETURN(-ENOMEM);
2681
2682         request->rq_replen = lustre_msg_size(1, &size);
2683         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2684
2685         rc = ptlrpc_queue_wait(request);
2686         if (rc)
2687                 GOTO(out, rc);
2688
2689         msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2690                                   lustre_swab_obd_statfs);
2691         if (msfs == NULL) {
2692                 CERROR("Can't unpack obd_statfs\n");
2693                 GOTO(out, rc = -EPROTO);
2694         }
2695
2696         memcpy(osfs, msfs, sizeof(*osfs));
2697
2698         EXIT;
2699  out:
2700         ptlrpc_req_finished(request);
2701         return rc;
2702 }
2703
2704 /* Retrieve object striping information.
2705  *
2706  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2707  * the maximum number of OST indices which will fit in the user buffer.
2708  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2709  */
2710 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2711 {
2712         struct lov_user_md lum, *lumk;
2713         int rc, lum_size;
2714         ENTRY;
2715
2716         if (!lsm)
2717                 RETURN(-ENODATA);
2718
2719         rc = copy_from_user(&lum, lump, sizeof(lum));
2720         if (rc)
2721                 RETURN(-EFAULT);
2722
2723         if (lum.lmm_magic != LOV_USER_MAGIC)
2724                 RETURN(-EINVAL);
2725
2726         if (lum.lmm_stripe_count > 0) {
2727                 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2728                 OBD_ALLOC(lumk, lum_size);
2729                 if (!lumk)
2730                         RETURN(-ENOMEM);
2731
2732                 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2733                 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2734         } else {
2735                 lum_size = sizeof(lum);
2736                 lumk = &lum;
2737         }
2738
2739         lumk->lmm_object_id = lsm->lsm_object_id;
2740         lumk->lmm_object_gr = lsm->lsm_object_gr;
2741         lumk->lmm_stripe_count = 1;
2742
2743         if (copy_to_user(lump, lumk, lum_size))
2744                 rc = -EFAULT;
2745
2746         if (lumk != &lum)
2747                 OBD_FREE(lumk, lum_size);
2748
2749         RETURN(rc);
2750 }
2751
2752 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2753                          void *karg, void *uarg)
2754 {
2755         struct obd_device *obd = exp->exp_obd;
2756         struct obd_ioctl_data *data = karg;
2757         int err = 0;
2758         ENTRY;
2759
2760 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2761         MOD_INC_USE_COUNT;
2762 #else
2763         if (!try_module_get(THIS_MODULE)) {
2764                 CERROR("Can't get module. Is it alive?");
2765                 return -EINVAL;
2766         }
2767 #endif
2768         switch (cmd) {
2769         case OBD_IOC_LOV_GET_CONFIG: {
2770                 char *buf;
2771                 struct lov_desc *desc;
2772                 struct obd_uuid uuid;
2773
2774                 buf = NULL;
2775                 len = 0;
2776                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2777                         GOTO(out, err = -EINVAL);
2778
2779                 data = (struct obd_ioctl_data *)buf;
2780
2781                 if (sizeof(*desc) > data->ioc_inllen1) {
2782                         OBD_FREE(buf, len);
2783                         GOTO(out, err = -EINVAL);
2784                 }
2785
2786                 if (data->ioc_inllen2 < sizeof(uuid)) {
2787                         OBD_FREE(buf, len);
2788                         GOTO(out, err = -EINVAL);
2789                 }
2790
2791                 if (data->ioc_inllen3 < sizeof(__u32)) {
2792                         OBD_FREE(buf, len);
2793                         GOTO(out, err = -EINVAL);
2794                 }
2795
2796                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2797                 desc->ld_tgt_count = 1;
2798                 desc->ld_active_tgt_count = 1;
2799                 desc->ld_default_stripe_count = 1;
2800                 desc->ld_default_stripe_size = 0;
2801                 desc->ld_default_stripe_offset = 0;
2802                 desc->ld_pattern = 0;
2803                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2804                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2805                 *((__u32 *)data->ioc_inlbuf3) = 1;
2806
2807                 err = copy_to_user((void *)uarg, buf, len);
2808                 if (err)
2809                         err = -EFAULT;
2810                 obd_ioctl_freedata(buf, len);
2811                 GOTO(out, err);
2812         }
2813         case LL_IOC_LOV_SETSTRIPE:
2814                 err = obd_alloc_memmd(exp, karg);
2815                 if (err > 0)
2816                         err = 0;
2817                 GOTO(out, err);
2818         case LL_IOC_LOV_GETSTRIPE:
2819                 err = osc_getstripe(karg, uarg);
2820                 GOTO(out, err);
2821         case OBD_IOC_CLIENT_RECOVER:
2822                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2823                                             data->ioc_inlbuf1);
2824                 if (err > 0)
2825                         err = 0;
2826                 GOTO(out, err);
2827         case IOC_OSC_SET_ACTIVE:
2828                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2829                                                data->ioc_offset);
2830                 GOTO(out, err);
2831         case IOC_OSC_CTL_RECOVERY:
2832                 err = ptlrpc_import_control_recovery(obd->u.cli.cl_import,
2833                                                      data->ioc_offset);
2834                 GOTO(out, err);
2835         default:
2836                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2837                 GOTO(out, err = -ENOTTY);
2838         }
2839 out:
2840 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2841         MOD_DEC_USE_COUNT;
2842 #else
2843         module_put(THIS_MODULE);
2844 #endif
2845         return err;
2846 }
2847
2848 static int osc_get_info(struct obd_export *exp, __u32 keylen,
2849                         void *key, __u32 *vallen, void *val)
2850 {
2851         ENTRY;
2852         if (!vallen || !val)
2853                 RETURN(-EFAULT);
2854
2855         if (keylen > strlen("lock_to_stripe") &&
2856             strcmp(key, "lock_to_stripe") == 0) {
2857                 __u32 *stripe = val;
2858                 *vallen = sizeof(*stripe);
2859                 *stripe = 0;
2860                 RETURN(0);
2861         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2862                 struct ptlrpc_request *req;
2863                 obd_id *reply;
2864                 char *bufs[1] = {key};
2865                 int rc;
2866                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2867                                       OST_GET_INFO, 1, (int *)&keylen, bufs);
2868                 if (req == NULL)
2869                         RETURN(-ENOMEM);
2870
2871                 req->rq_replen = lustre_msg_size(1, (int *)vallen);
2872                 rc = ptlrpc_queue_wait(req);
2873                 if (rc)
2874                         GOTO(out, rc);
2875
2876                 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2877                                            lustre_swab_ost_last_id);
2878                 if (reply == NULL) {
2879                         CERROR("Can't unpack OST last ID\n");
2880                         GOTO(out, rc = -EPROTO);
2881                 }
2882                 *((obd_id *)val) = *reply;
2883         out:
2884                 ptlrpc_req_finished(req);
2885                 RETURN(rc);
2886         } else if (keylen >= strlen("client_nid") &&
2887                    strcmp(key, "client_nid") == 0) {
2888                 struct ptlrpc_connection * conn;
2889                 ptl_nid_t * nid = val;
2890                 *vallen = sizeof(*nid);
2891
2892                 conn = class_exp2cliimp(exp)->imp_connection;
2893                 if (!conn)
2894                         RETURN(-ENOTCONN);
2895
2896                 nid = &conn->c_peer.peer_id.nid;
2897
2898                 RETURN(0);
2899         }
2900         RETURN(-EPROTO);
2901 }
2902
2903 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2904                         void *key, obd_count vallen, void *val)
2905 {
2906         struct obd_device  *obd = exp->exp_obd;
2907         struct obd_import *imp = class_exp2cliimp(exp);
2908         struct llog_ctxt *ctxt;
2909         int rc = 0;
2910         ENTRY;
2911
2912         if (keylen == strlen("unlinked") &&
2913             memcmp(key, "unlinked", keylen) == 0) {
2914                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2915                 spin_lock(&oscc->oscc_lock);
2916                 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2917                 spin_unlock(&oscc->oscc_lock);
2918                 RETURN(0);
2919         }
2920         if (keylen == strlen("unrecovery") &&
2921             memcmp(key, "unrecovery", keylen) == 0) {
2922                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2923                 spin_lock(&oscc->oscc_lock);
2924                 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
2925                 spin_unlock(&oscc->oscc_lock);
2926                 RETURN(0);
2927         }
2928         if (keylen == strlen("initial_recov") &&
2929             memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
2930                 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2931                 if (vallen != sizeof(int))
2932                         RETURN(-EINVAL);
2933                 imp->imp_initial_recov = *(int *)val;
2934                 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
2935                        exp->exp_obd->obd_name,
2936                        imp->imp_initial_recov);
2937                 RETURN(0);
2938         }
2939
2940         if (keylen == strlen("async") &&
2941             memcmp(key, "async", keylen) == 0) {
2942                 struct client_obd *cl = &obd->u.cli;
2943                 if (vallen != sizeof(int))
2944                         RETURN(-EINVAL);
2945                 cl->cl_async = *(int *)val;
2946                 CDEBUG(D_HA, "%s: set async = %d\n",
2947                        obd->obd_name, cl->cl_async);
2948                 RETURN(0);
2949         }
2950
2951         if (keylen == 5 && strcmp(key, "audit") == 0) {
2952                 struct ptlrpc_request *req;
2953                 char *bufs[2] = {key, val};
2954                 int size[2] = {keylen, vallen};
2955
2956                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2957                                       OST_SET_INFO, 2, size, bufs);
2958                 if (req == NULL)
2959                         RETURN(-ENOMEM);
2960
2961                 req->rq_replen = lustre_msg_size(0, size);
2962                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
2963                                    lustre_swab_audit_attr);
2964                 rc = ptlrpc_queue_wait(req);
2965
2966                 ptlrpc_req_finished(req);
2967                 RETURN(rc);
2968         }
2969
2970         if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
2971                 struct ptlrpc_request *req;
2972                 char *bufs[2] = {key, val};
2973                 int size[2] = {keylen, vallen};
2974
2975                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2976                                       OST_SET_INFO, 2, size, bufs);
2977                 if (req == NULL)
2978                         RETURN(-ENOMEM);
2979
2980                 req->rq_replen = lustre_msg_size(0, size);
2981                 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
2982                                    lustre_swab_obdo);
2983                 rc = ptlrpc_queue_wait(req);
2984
2985                 ptlrpc_req_finished(req);
2986                 RETURN(rc);
2987         }
2988
2989         if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
2990                 struct ptlrpc_request *req;
2991                 char *bufs[2] = {key, val};
2992                 int size[2] = {keylen, vallen};
2993
2994                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2995                                       OST_SET_INFO, 2, size, bufs);
2996                 if (req == NULL)
2997                         RETURN(-ENOMEM);
2998
2999                 req->rq_replen = lustre_msg_size(0, size);
3000                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
3001                                    lustre_swab_audit_msg);
3002                 rc = ptlrpc_queue_wait(req);
3003
3004                 ptlrpc_req_finished(req);
3005                 RETURN(rc);
3006         }
3007
3008         if (keylen == strlen("sec") &&
3009             memcmp(key, "sec", keylen) == 0) {
3010                 struct client_obd *cli = &exp->exp_obd->u.cli;
3011
3012                 cli->cl_sec_flavor = ptlrpcs_name2flavor(val);
3013                 if (cli->cl_sec_flavor == PTLRPCS_FLVR_INVALID) {
3014                         CERROR("unrecognized security flavor %s\n", (char*) val);
3015                         RETURN(-EINVAL);
3016                 }
3017
3018                 RETURN(0);
3019         }
3020
3021         if (keylen == strlen("sec_flags") &&
3022             memcmp(key, "sec_flags", keylen) == 0) {
3023                 struct client_obd *cli = &exp->exp_obd->u.cli;
3024
3025                 cli->cl_sec_flags = *((unsigned long *) val);
3026                 RETURN(0);
3027         }
3028
3029         if (keylen == strlen("flush_cred") &&
3030             memcmp(key, "flush_cred", keylen) == 0) {
3031                 struct client_obd *cli = &exp->exp_obd->u.cli;
3032
3033                 if (cli->cl_import)
3034                         ptlrpcs_import_flush_current_creds(cli->cl_import);
3035                 RETURN(0);
3036         }
3037         if (keylen == strlen("crypto_cb") &&
3038             memcmp(key, "crypto_cb", keylen) == 0) {
3039                 LASSERT(vallen == sizeof(crypt_cb_t));
3040                 osc_crypt_cb = (crypt_cb_t)val;
3041                 RETURN(0);
3042         }
3043
3044         if (keylen < strlen("mds_conn") ||
3045             memcmp(key, "mds_conn", keylen) != 0)
3046                 RETURN(-EINVAL);
3047
3048         ctxt = llog_get_context(&exp->exp_obd->obd_llogs,
3049                                 LLOG_UNLINK_ORIG_CTXT);
3050         if (ctxt) {
3051                 if (rc == 0)
3052                         rc = llog_initiator_connect(ctxt);
3053                 else
3054                         CERROR("cannot establish the connect for "
3055                                "ctxt %p: %d\n", ctxt, rc);
3056         }
3057
3058         imp->imp_server_timeout = 1;
3059         CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
3060         imp->imp_pingable = 1;
3061
3062         RETURN(rc);
3063 }
3064
3065
3066 static struct llog_operations osc_size_repl_logops = {
3067         lop_cancel: llog_obd_repl_cancel
3068 };
3069
3070 static struct llog_operations osc_unlink_orig_logops;
3071
3072 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
3073                          struct obd_device *tgt, int count,
3074                          struct llog_catid *catid)
3075 {
3076         int rc;
3077         ENTRY;
3078
3079         osc_unlink_orig_logops = llog_lvfs_ops;
3080         osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
3081         osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
3082         osc_unlink_orig_logops.lop_add = llog_catalog_add;
3083         osc_unlink_orig_logops.lop_connect = llog_origin_connect;
3084
3085         rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
3086                             &catid->lci_logid, &osc_unlink_orig_logops);
3087         if (rc)
3088                 RETURN(rc);
3089
3090         rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
3091                             &osc_size_repl_logops);
3092         RETURN(rc);
3093 }
3094
3095 static int osc_llog_finish(struct obd_device *obd,
3096                            struct obd_llogs *llogs, int count)
3097 {
3098         int rc;
3099         ENTRY;
3100
3101         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
3102         if (rc)
3103                 RETURN(rc);
3104
3105         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
3106         RETURN(rc);
3107 }
3108
3109 static int osc_connect(struct lustre_handle *exph,
3110                        struct obd_device *obd, struct obd_uuid *cluuid,
3111                        struct obd_connect_data *data,
3112                        unsigned long connect_flags)
3113 {
3114         int rc;
3115         ENTRY;
3116         rc = client_connect_import(exph, obd, cluuid, data, connect_flags);
3117         RETURN(rc);
3118 }
3119
3120 static int osc_disconnect(struct obd_export *exp, unsigned long flags)
3121 {
3122         struct obd_device *obd = class_exp2obd(exp);
3123         struct llog_ctxt *ctxt;
3124         int rc;
3125         ENTRY;
3126
3127         ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
3128         if (obd->u.cli.cl_conn_count == 1)
3129                 /* flush any remaining cancel messages out to the target */
3130                 llog_sync(ctxt, exp);
3131
3132         rc = client_disconnect_export(exp, flags);
3133         RETURN(rc);
3134 }
3135
3136 static int osc_import_event(struct obd_device *obd,
3137                             struct obd_import *imp,
3138                             enum obd_import_event event)
3139 {
3140         struct client_obd *cli;
3141         int rc = 0;
3142
3143         LASSERT(imp->imp_obd == obd);
3144
3145         switch (event) {
3146         case IMP_EVENT_DISCON: {
3147                 /* Only do this on the MDS OSC's */
3148                 if (imp->imp_server_timeout) {
3149                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3150
3151                         spin_lock(&oscc->oscc_lock);
3152                         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3153                         spin_unlock(&oscc->oscc_lock);
3154                 }
3155                 break;
3156         }
3157         case IMP_EVENT_INACTIVE: {
3158                 if (obd->obd_observer)
3159                         rc = obd_notify(obd->obd_observer, obd, 0, 0);
3160                 break;
3161         }
3162         case IMP_EVENT_INVALIDATE: {
3163                 struct ldlm_namespace *ns = obd->obd_namespace;
3164
3165                 /* Reset grants */
3166                 cli = &obd->u.cli;
3167                 spin_lock(&cli->cl_loi_list_lock);
3168                 cli->cl_avail_grant = 0;
3169                 cli->cl_lost_grant = 0;
3170                 /* all pages go to failing rpcs due to the invalid import */
3171                 osc_check_rpcs(cli);
3172                 spin_unlock(&cli->cl_loi_list_lock);
3173
3174                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3175
3176                 break;
3177         }
3178         case IMP_EVENT_ACTIVE: {
3179                 /* Only do this on the MDS OSC's */
3180                 if (imp->imp_server_timeout) {
3181                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3182
3183                         spin_lock(&oscc->oscc_lock);
3184                         oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3185                         spin_unlock(&oscc->oscc_lock);
3186                 }
3187
3188                 if (obd->obd_observer)
3189                         rc = obd_notify(obd->obd_observer, obd, 1, 0);
3190                 break;
3191         }
3192         default:
3193                 CERROR("Unknown import event %d\n", event);
3194                 LBUG();
3195         }
3196         RETURN(rc);
3197 }
3198
3199 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
3200 {
3201         struct lprocfs_static_vars lvars;
3202         int rc;
3203         ENTRY;
3204
3205         lprocfs_init_vars(osc,&lvars);
3206         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3207         if (rc < 0)
3208                 RETURN(rc);
3209
3210         rc = lproc_osc_attach_seqstat(dev);
3211         if (rc < 0) {
3212                 lprocfs_obd_detach(dev);
3213                 RETURN(rc);
3214         }
3215
3216         ptlrpc_lprocfs_register_obd(dev);
3217         RETURN(0);
3218 }
3219
3220 static int osc_detach(struct obd_device *dev)
3221 {
3222         ptlrpc_lprocfs_unregister_obd(dev);
3223         return lprocfs_obd_detach(dev);
3224 }
3225
3226 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3227 {
3228         int rc;
3229         ENTRY;
3230         rc = ptlrpcd_addref();
3231         if (rc)
3232                 RETURN(rc);
3233
3234         rc = client_obd_setup(obd, len, buf);
3235         if (rc)
3236                 ptlrpcd_decref();
3237         else
3238                 oscc_init(obd);
3239
3240         RETURN(rc);
3241 }
3242
3243 static int osc_cleanup(struct obd_device *obd, int flags)
3244 {
3245         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3246         int rc;
3247
3248         rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3249                                     LDLM_FL_CONFIG_CHANGE, NULL);
3250         if (rc)
3251                 RETURN(rc);
3252
3253         spin_lock(&oscc->oscc_lock);
3254         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3255         oscc->oscc_flags |= OSCC_FLAG_EXITING;
3256         spin_unlock(&oscc->oscc_lock);
3257
3258         rc = client_obd_cleanup(obd, flags);
3259         ptlrpcd_decref();
3260         RETURN(rc);
3261 }
3262
3263 struct obd_ops osc_obd_ops = {
3264         .o_owner                = THIS_MODULE,
3265         .o_attach               = osc_attach,
3266         .o_detach               = osc_detach,
3267         .o_setup                = osc_setup,
3268         .o_cleanup              = osc_cleanup,
3269         .o_add_conn             = client_import_add_conn,
3270         .o_del_conn             = client_import_del_conn,
3271         .o_connect              = osc_connect,
3272         .o_disconnect           = osc_disconnect,
3273         .o_statfs               = osc_statfs,
3274         .o_packmd               = osc_packmd,
3275         .o_unpackmd             = osc_unpackmd,
3276         .o_create               = osc_create,
3277         .o_destroy              = osc_destroy,
3278         .o_getattr              = osc_getattr,
3279         .o_getattr_async        = osc_getattr_async,
3280         .o_setattr              = osc_setattr,
3281         .o_brw                  = osc_brw,
3282         .o_brw_async            = osc_brw_async,
3283         .o_prep_async_page      = osc_prep_async_page,
3284         .o_queue_async_io       = osc_queue_async_io,
3285         .o_set_async_flags      = osc_set_async_flags,
3286         .o_queue_group_io       = osc_queue_group_io,
3287         .o_trigger_group_io     = osc_trigger_group_io,
3288         .o_teardown_async_page  = osc_teardown_async_page,
3289         .o_punch                = osc_punch,
3290         .o_sync                 = osc_sync,
3291         .o_enqueue              = osc_enqueue,
3292         .o_match                = osc_match,
3293         .o_change_cbdata        = osc_change_cbdata,
3294         .o_cancel               = osc_cancel,
3295         .o_cancel_unused        = osc_cancel_unused,
3296         .o_iocontrol            = osc_iocontrol,
3297         .o_get_info             = osc_get_info,
3298         .o_set_info             = osc_set_info,
3299         .o_import_event         = osc_import_event,
3300         .o_llog_init            = osc_llog_init,
3301         .o_llog_finish          = osc_llog_finish,
3302 };
3303
3304 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3305 struct obd_ops sanosc_obd_ops = {
3306         .o_owner                = THIS_MODULE,
3307         .o_attach               = osc_attach,
3308         .o_detach               = osc_detach,
3309         .o_cleanup              = client_obd_cleanup,
3310         .o_add_conn             = client_import_add_conn,
3311         .o_del_conn             = client_import_del_conn,
3312         .o_connect              = osc_connect,
3313         .o_disconnect           = client_disconnect_export,
3314         .o_statfs               = osc_statfs,
3315         .o_packmd               = osc_packmd,
3316         .o_unpackmd             = osc_unpackmd,
3317         .o_create               = osc_real_create,
3318         .o_destroy              = osc_destroy,
3319         .o_getattr              = osc_getattr,
3320         .o_getattr_async        = osc_getattr_async,
3321         .o_setattr              = osc_setattr,
3322         .o_setup                = client_sanobd_setup,
3323         .o_brw                  = sanosc_brw,
3324         .o_punch                = osc_punch,
3325         .o_sync                 = osc_sync,
3326         .o_enqueue              = osc_enqueue,
3327         .o_match                = osc_match,
3328         .o_change_cbdata        = osc_change_cbdata,
3329         .o_cancel               = osc_cancel,
3330         .o_cancel_unused        = osc_cancel_unused,
3331         .o_iocontrol            = osc_iocontrol,
3332         .o_import_event         = osc_import_event,
3333         .o_llog_init            = osc_llog_init,
3334         .o_llog_finish          = osc_llog_finish,
3335 };
3336 #endif
3337
3338 int __init osc_init(void)
3339 {
3340         struct lprocfs_static_vars lvars;
3341 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3342         struct lprocfs_static_vars sanlvars;
3343 #endif
3344         int rc;
3345         ENTRY;
3346
3347         lprocfs_init_vars(osc, &lvars);
3348 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3349         lprocfs_init_vars(osc, &sanlvars);
3350 #endif
3351
3352         rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3353                                  OBD_OSC_DEVICENAME);
3354         if (rc)
3355                 RETURN(rc);
3356
3357 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3358         rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3359                                  OBD_SANOSC_DEVICENAME);
3360         if (rc)
3361                 class_unregister_type(OBD_OSC_DEVICENAME);
3362 #endif
3363
3364         RETURN(rc);
3365 }
3366
3367 #ifdef __KERNEL__
3368 static void /*__exit*/ osc_exit(void)
3369 {
3370 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3371         class_unregister_type(OBD_SANOSC_DEVICENAME);
3372 #endif
3373         class_unregister_type(OBD_OSC_DEVICENAME);
3374 }
3375
3376 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3377 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3378 MODULE_LICENSE("GPL");
3379
3380 module_init(osc_init);
3381 module_exit(osc_exit);
3382 #endif