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_lite.h>
  67 #include <linux/lustre_audit.h>
  68 #include <linux/lustre_gs.h>
  69
  70 #include "osc_internal.h"
  71
  72 /* Pack OSC object metadata for disk storage (LE byte order). */
  73 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
  74                       struct lov_stripe_md *lsm)
  75 {
  76         int lmm_size;
  77         ENTRY;
  78
  79         lmm_size = sizeof(**lmmp);
  80         if (!lmmp)
  81                 RETURN(lmm_size);
  82
  83         if (*lmmp && !lsm) {
  84                 OBD_FREE(*lmmp, lmm_size);
  85                 *lmmp = NULL;
  86                 RETURN(0);
  87         }
  88
  89         if (!*lmmp) {
  90                 OBD_ALLOC(*lmmp, lmm_size);
  91                 if (!*lmmp)
  92                         RETURN(-ENOMEM);
  93         }
  94
  95         if (lsm) {
  96                 LASSERT(lsm->lsm_object_id);
  97                 LASSERT(lsm->lsm_object_gr);
  98                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
  99                 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
 100         }
 101
 102         RETURN(lmm_size);
 103 }
 104
 105 /* Unpack OSC object metadata from disk storage (LE byte order). */
 106 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
 107                         struct lov_mds_md *lmm, int lmm_bytes)
 108 {
 109         int lsm_size;
 110         ENTRY;
 111
 112         if (lmm != NULL) {
 113                 if (lmm_bytes < sizeof (*lmm)) {
 114                         CERROR("lov_mds_md too small: %d, need %d\n",
 115                                lmm_bytes, (int)sizeof(*lmm));
 116                         RETURN(-EINVAL);
 117                 }
 118                 /* XXX LOV_MAGIC etc check? */
 119
 120                 if (lmm->lmm_object_id == 0) {
 121                         CERROR("lov_mds_md: zero lmm_object_id\n");
 122                         RETURN(-EINVAL);
 123                 }
 124         }
 125
 126         lsm_size = lov_stripe_md_size(1);
 127         if (lsmp == NULL)
 128                 RETURN(lsm_size);
 129
 130         if (*lsmp != NULL && lmm == NULL) {
 131                 OBD_FREE(*lsmp, lsm_size);
 132                 *lsmp = NULL;
 133                 RETURN(0);
 134         }
 135
 136         if (*lsmp == NULL) {
 137                 OBD_ALLOC(*lsmp, lsm_size);
 138                 if (*lsmp == NULL)
 139                         RETURN(-ENOMEM);
 140                 loi_init((*lsmp)->lsm_oinfo);
 141         }
 142
 143         if (lmm != NULL) {
 144                 /* XXX zero *lsmp? */
 145                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
 146                 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
 147                 LASSERT((*lsmp)->lsm_object_id);
 148                 LASSERT((*lsmp)->lsm_object_gr);
 149         }
 150
 151         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
 152
 153         RETURN(lsm_size);
 154 }
 155
 156 static int osc_getattr_interpret(struct ptlrpc_request *req,
 157                                  struct osc_getattr_async_args *aa, int rc)
 158 {
 159         struct ost_body *body;
 160         ENTRY;
 161
 162         if (rc != 0)
 163                 RETURN(rc);
 164
 165         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 166         if (body) {
 167                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 168                 memcpy(aa->aa_oa, &body->oa, sizeof(*aa->aa_oa));
 169
 170                 /* This should really be sent by the OST */
 171                 aa->aa_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 172                 aa->aa_oa->o_valid |= OBD_MD_FLBLKSZ;
 173         } else {
 174                 CERROR("can't unpack ost_body\n");
 175                 rc = -EPROTO;
 176                 aa->aa_oa->o_valid = 0;
 177         }
 178
 179         RETURN(rc);
 180 }
 181
 182 static int osc_getattr_async(struct obd_export *exp, struct obdo *oa,
 183                              struct lov_stripe_md *md,
 184                              struct ptlrpc_request_set *set)
 185 {
 186         struct ptlrpc_request *request;
 187         struct ost_body *body;
 188         int size = sizeof(*body);
 189         struct osc_getattr_async_args *aa;
 190         ENTRY;
 191
 192         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 193                                   OST_GETATTR, 1, &size, NULL);
 194         if (!request)
 195                 RETURN(-ENOMEM);
 196
 197         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 198         memcpy(&body->oa, oa, sizeof(*oa));
 199
 200         request->rq_replen = lustre_msg_size(1, &size);
 201         request->rq_interpret_reply = osc_getattr_interpret;
 202
 203         LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
 204         aa = (struct osc_getattr_async_args *)&request->rq_async_args;
 205         aa->aa_oa = oa;
 206
 207         ptlrpc_set_add_req (set, request);
 208         RETURN (0);
 209 }
 210
 211 static int osc_getattr(struct obd_export *exp, struct obdo *oa,
 212                        struct lov_stripe_md *md)
 213 {
 214         struct ptlrpc_request *request;
 215         struct ost_body *body;
 216         int rc, size = sizeof(*body);
 217         ENTRY;
 218
 219         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 220                                   OST_GETATTR, 1, &size, NULL);
 221         if (!request)
 222                 RETURN(-ENOMEM);
 223
 224         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 225         memcpy(&body->oa, oa, sizeof(*oa));
 226
 227         request->rq_replen = lustre_msg_size(1, &size);
 228
 229         rc = ptlrpc_queue_wait(request);
 230         if (rc) {
 231                 CERROR("%s failed: rc = %d\n", __FUNCTION__, rc);
 232                 GOTO(out, rc);
 233         }
 234
 235         body = lustre_swab_repbuf(request, 0, sizeof (*body),
 236                                   lustre_swab_ost_body);
 237         if (body == NULL) {
 238                 CERROR ("can't unpack ost_body\n");
 239                 GOTO (out, rc = -EPROTO);
 240         }
 241
 242         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 243         memcpy(oa, &body->oa, sizeof(*oa));
 244
 245         /* This should really be sent by the OST */
 246         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 247         oa->o_valid |= OBD_MD_FLBLKSZ;
 248
 249         EXIT;
 250  out:
 251         ptlrpc_req_finished(request);
 252         return rc;
 253 }
 254
 255 static int osc_setattr(struct obd_export *exp, struct obdo *oa,
 256                        struct lov_stripe_md *md, struct obd_trans_info *oti)
 257 {
 258         struct ptlrpc_request *request;
 259         struct ost_body *body;
 260         int rc, size = sizeof(*body);
 261         ENTRY;
 262
 263         LASSERT(!(oa->o_valid & OBD_MD_FLGROUP) || oa->o_gr > 0);
 264
 265         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 266                                   OST_SETATTR, 1, &size, NULL);
 267         if (!request)
 268                 RETURN(-ENOMEM);
 269
 270         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
 271         memcpy(&body->oa, oa, sizeof(*oa));
 272
 273         request->rq_replen = lustre_msg_size(1, &size);
 274
 275         if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
 276                 ptlrpcd_add_req(request);
 277                 rc = 0;
 278         } else {
 279                 rc = ptlrpc_queue_wait(request);
 280                 if (rc)
 281                         GOTO(out, rc);
 282
 283                 body = lustre_swab_repbuf(request, 0, sizeof(*body),
 284                                           lustre_swab_ost_body);
 285                 if (body == NULL)
 286                         GOTO(out, rc = -EPROTO);
 287
 288                 memcpy(oa, &body->oa, sizeof(*oa));
 289         }
 290         EXIT;
 291 out:
 292         ptlrpc_req_finished(request);
 293         RETURN(0);
 294 }
 295
 296 int osc_real_create(struct obd_export *exp, struct obdo *oa,
 297                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
 298 {
 299         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 300         struct ptlrpc_request *request;
 301         struct ost_body *body;
 302         struct lov_stripe_md *lsm;
 303         int rc, size = sizeof(*body);
 304         ENTRY;
 305
 306         LASSERT(oa);
 307         LASSERT(ea);
 308
 309         lsm = *ea;
 310         if (!lsm) {
 311                 rc = obd_alloc_memmd(exp, &lsm);
 312                 if (rc < 0)
 313                         RETURN(rc);
 314         }
 315
 316         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 317                                   OST_CREATE, 1, &size, NULL);
 318         if (!request)
 319                 GOTO(out, rc = -ENOMEM);
 320
 321         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 322         memcpy(&body->oa, oa, sizeof(body->oa));
 323
 324         request->rq_replen = lustre_msg_size(1, &size);
 325         if (oa->o_valid & OBD_MD_FLINLINE) {
 326                 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
 327                         oa->o_flags == OBD_FL_DELORPHAN);
 328                 DEBUG_REQ(D_HA, request,
 329                           "delorphan from OST integration");
 330                 /* Don't resend the delorphan request */
 331                 request->rq_no_resend = request->rq_no_delay = 1;
 332         }
 333
 334         rc = ptlrpc_queue_wait(request);
 335         if (rc)
 336                 GOTO(out_req, rc);
 337
 338         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 339                                   lustre_swab_ost_body);
 340         if (body == NULL) {
 341                 CERROR ("can't unpack ost_body\n");
 342                 GOTO (out_req, rc = -EPROTO);
 343         }
 344
 345         if ((oa->o_valid & OBD_MD_FLFLAGS) && oa->o_flags == OBD_FL_DELORPHAN) {
 346                 struct obd_import *imp = class_exp2cliimp(exp);
 347                 /* MDS declares last known object, OSS responses
 348                  * with next possible object -bzzz */
 349                 spin_lock(&oscc->oscc_lock);
 350                 oscc->oscc_next_id = body->oa.o_id;
 351                 spin_unlock(&oscc->oscc_lock);
 352                 CDEBUG(D_HA, "%s: set nextid "LPD64" after recovery\n",
 353                        imp->imp_target_uuid.uuid, oa->o_id);
 354         }
 355         memcpy(oa, &body->oa, sizeof(*oa));
 356
 357         /* This should really be sent by the OST */
 358         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 359         oa->o_valid |= OBD_MD_FLBLKSZ;
 360
 361         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
 362          * have valid lsm_oinfo data structs, so don't go touching that.
 363          * This needs to be fixed in a big way.
 364          */
 365         lsm->lsm_object_id = oa->o_id;
 366         lsm->lsm_object_gr = oa->o_gr;
 367         *ea = lsm;
 368
 369         if (oti != NULL) {
 370                 oti->oti_transno = request->rq_repmsg->transno;
 371
 372                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
 373                         if (!oti->oti_logcookies)
 374                                 oti_alloc_cookies(oti, 1);
 375                         memcpy(oti->oti_logcookies, obdo_logcookie(oa),
 376                                sizeof(oti->oti_onecookie));
 377                 }
 378         }
 379
 380         CDEBUG(D_HA, "transno: "LPD64"\n", request->rq_repmsg->transno);
 381         EXIT;
 382 out_req:
 383         ptlrpc_req_finished(request);
 384 out:
 385         if (rc && !*ea)
 386                 obd_free_memmd(exp, &lsm);
 387         return rc;
 388 }
 389
 390 static int osc_punch(struct obd_export *exp, struct obdo *oa,
 391                      struct lov_stripe_md *md, obd_size start,
 392                      obd_size end, struct obd_trans_info *oti,
 393                      struct lustre_capa *capa)
 394 {
 395         struct ptlrpc_request *request;
 396         struct ost_body *body;
 397         struct lustre_capa *req_capa;
 398         int bufcnt = 0;
 399         int rc, size[2] = { sizeof(*body), sizeof(*capa) };
 400         ENTRY;
 401
 402         if (!oa) {
 403                 CERROR("oa NULL\n");
 404                 RETURN(-EINVAL);
 405         }
 406
 407         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 408                                   OST_PUNCH, capa ? 2 : 1, size, NULL);
 409         if (!request)
 410                 RETURN(-ENOMEM);
 411
 412         body = lustre_msg_buf(request->rq_reqmsg, bufcnt++, sizeof (*body));
 413
 414         memcpy(&body->oa, oa, sizeof(*oa));
 415
 416         /* overload the size and blocks fields in the oa with start/end */
 417         body->oa.o_size = start;
 418         body->oa.o_blocks = end;
 419         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 420
 421         if (capa) {
 422                 req_capa = lustre_msg_buf(request->rq_reqmsg, bufcnt++,
 423                                           sizeof(*capa));
 424                 capa_dup2(req_capa, capa);
 425                 body->oa.o_valid |= OBD_MD_CAPA;
 426         }
 427
 428         request->rq_replen = lustre_msg_size(1, size);
 429
 430         rc = ptlrpc_queue_wait(request);
 431         if (rc)
 432                 GOTO(out, rc);
 433
 434         body = lustre_swab_repbuf (request, 0, sizeof (*body),
 435                                    lustre_swab_ost_body);
 436         if (body == NULL) {
 437                 CERROR ("can't unpack ost_body\n");
 438                 GOTO (out, rc = -EPROTO);
 439         }
 440
 441         memcpy(oa, &body->oa, sizeof(*oa));
 442
 443         EXIT;
 444  out:
 445         ptlrpc_req_finished(request);
 446         return rc;
 447 }
 448
 449 static int osc_sync(struct obd_export *exp, struct obdo *oa,
 450                     struct lov_stripe_md *md, obd_size start,
 451                     obd_size end)
 452 {
 453         struct ptlrpc_request *request;
 454         struct ost_body *body;
 455         int rc, size = sizeof(*body);
 456         ENTRY;
 457
 458         if (!oa) {
 459                 CERROR("oa NULL\n");
 460                 RETURN(-EINVAL);
 461         }
 462
 463         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 464                                   OST_SYNC, 1, &size, NULL);
 465         if (!request)
 466                 RETURN(-ENOMEM);
 467
 468         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 469         memcpy(&body->oa, oa, sizeof(*oa));
 470
 471         /* overload the size and blocks fields in the oa with start/end */
 472         body->oa.o_size = start;
 473         body->oa.o_blocks = end;
 474         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 475
 476         request->rq_replen = lustre_msg_size(1, &size);
 477
 478         rc = ptlrpc_queue_wait(request);
 479         if (rc)
 480                 GOTO(out, rc);
 481
 482         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 483                                   lustre_swab_ost_body);
 484         if (body == NULL) {
 485                 CERROR ("can't unpack ost_body\n");
 486                 GOTO (out, rc = -EPROTO);
 487         }
 488
 489         memcpy(oa, &body->oa, sizeof(*oa));
 490
 491         EXIT;
 492  out:
 493         ptlrpc_req_finished(request);
 494         return rc;
 495 }
 496
 497 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
 498                        struct lov_stripe_md *ea, struct obd_trans_info *oti)
 499 {
 500         struct ptlrpc_request *request;
 501         struct ost_body *body;
 502         int rc, size = sizeof(*body);
 503         ENTRY;
 504
 505         if (!oa) {
 506                 CERROR("oa NULL\n");
 507                 RETURN(-EINVAL);
 508         }
 509
 510         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 511                                   OST_DESTROY, 1, &size, NULL);
 512         if (!request)
 513                 RETURN(-ENOMEM);
 514         request->rq_request_portal = OST_DESTROY_PORTAL;
 515
 516         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 517
 518         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
 519                 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
 520                        sizeof(*oti->oti_logcookies));
 521                 oti->oti_logcookies++;
 522         }
 523
 524         memcpy(&body->oa, oa, sizeof(*oa));
 525         request->rq_replen = lustre_msg_size(1, &size);
 526
 527         if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
 528                 ptlrpcd_add_req(request);
 529                 rc = 0;
 530         } else {
 531                 rc = ptlrpc_queue_wait(request);
 532
 533                 if (rc == -ENOENT)
 534                         rc = 0;
 535
 536                 if (rc) {
 537                         ptlrpc_req_finished(request);
 538                         RETURN(rc);
 539                 }
 540
 541                 body = lustre_swab_repbuf(request, 0, sizeof(*body),
 542                                           lustre_swab_ost_body);
 543                 if (body == NULL) {
 544                         CERROR ("Can't unpack body\n");
 545                         ptlrpc_req_finished(request);
 546                         RETURN(-EPROTO);
 547                 }
 548
 549                 memcpy(oa, &body->oa, sizeof(*oa));
 550                 ptlrpc_req_finished(request);
 551         }
 552         RETURN(rc);
 553 }
 554
 555 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 556                                 long writing_bytes)
 557 {
 558         obd_valid bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
 559
 560         LASSERT(!(oa->o_valid & bits));
 561
 562         oa->o_valid |= bits;
 563         spin_lock(&cli->cl_loi_list_lock);
 564         oa->o_dirty = cli->cl_dirty;
 565         oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
 566         oa->o_grant = cli->cl_avail_grant;
 567         oa->o_dropped = cli->cl_lost_grant;
 568         cli->cl_lost_grant = 0;
 569         spin_unlock(&cli->cl_loi_list_lock);
 570         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
 571                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 572 }
 573
 574 /* caller must hold loi_list_lock */
 575 static void osc_consume_write_grant(struct client_obd *cli,
 576                                     struct osc_async_page *oap)
 577 {
 578         cli->cl_dirty += PAGE_SIZE;
 579         cli->cl_avail_grant -= PAGE_SIZE;
 580         oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
 581         CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
 582         LASSERT(cli->cl_avail_grant >= 0);
 583 }
 584
 585 static unsigned long rpcs_in_flight(struct client_obd *cli)
 586 {
 587         return cli->cl_r_in_flight + cli->cl_w_in_flight;
 588 }
 589
 590 /* caller must hold loi_list_lock */
 591 void osc_wake_cache_waiters(struct client_obd *cli)
 592 {
 593         struct list_head *l, *tmp;
 594         struct osc_cache_waiter *ocw;
 595
 596         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 597                 /* if we can't dirty more, we must wait until some is written */
 598                 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
 599                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
 600                                cli->cl_dirty, cli->cl_dirty_max);
 601                         return;
 602                 }
 603
 604                 /* if still dirty cache but no grant wait for pending RPCs that
 605                  * may yet return us some grant before doing sync writes */
 606                 if (cli->cl_w_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
 607                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
 608                                cli->cl_w_in_flight);
 609                 }
 610                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
 611                 list_del_init(&ocw->ocw_entry);
 612                 if (cli->cl_avail_grant < PAGE_SIZE) {
 613                         /* no more RPCs in flight to return grant, do sync IO */
 614                         ocw->ocw_rc = -EDQUOT;
 615                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
 616                 } else {
 617                         osc_consume_write_grant(cli, ocw->ocw_oap);
 618                 }
 619
 620                 wake_up(&ocw->ocw_waitq);
 621         }
 622
 623         EXIT;
 624 }
 625
 626 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 627 {
 628         spin_lock(&cli->cl_loi_list_lock);
 629         CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
 630         cli->cl_avail_grant += body->oa.o_grant;
 631         /* waiters are woken in brw_interpret_oap */
 632         spin_unlock(&cli->cl_loi_list_lock);
 633 }
 634
 635 /* We assume that the reason this OSC got a short read is because it read
 636  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
 637  * via the LOV, and it _knows_ it's reading inside the file, it's just that
 638  * this stripe never got written at or beyond this stripe offset yet. */
 639 static void handle_short_read(int nob_read, obd_count page_count,
 640                               struct brw_page *pga)
 641 {
 642         char *ptr;
 643
 644         /* skip bytes read OK */
 645         while (nob_read > 0) {
 646                 LASSERT (page_count > 0);
 647
 648                 if (pga->count > nob_read) {
 649                         /* EOF inside this page */
 650                         ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
 651                         memset(ptr + nob_read, 0, pga->count - nob_read);
 652                         kunmap(pga->pg);
 653                         page_count--;
 654                         pga++;
 655                         break;
 656                 }
 657
 658                 nob_read -= pga->count;
 659                 page_count--;
 660                 pga++;
 661         }
 662
 663         /* zero remaining pages */
 664         while (page_count-- > 0) {
 665                 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
 666                 memset(ptr, 0, pga->count);
 667                 kunmap(pga->pg);
 668                 pga++;
 669         }
 670 }
 671
 672 static int check_write_rcs(struct ptlrpc_request *request,
 673                            int requested_nob, int niocount,
 674                            obd_count page_count, struct brw_page *pga)
 675 {
 676         int *remote_rcs, i;
 677
 678         /* return error if any niobuf was in error */
 679         remote_rcs = lustre_swab_repbuf(request, 1,
 680                                         sizeof(*remote_rcs) * niocount, NULL);
 681         if (remote_rcs == NULL) {
 682                 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
 683                 return(-EPROTO);
 684         }
 685         if (lustre_msg_swabbed(request->rq_repmsg))
 686                 for (i = 0; i < niocount; i++)
 687                         __swab32s((__u32 *)&remote_rcs[i]);
 688
 689         for (i = 0; i < niocount; i++) {
 690                 if (remote_rcs[i] < 0)
 691                         return(remote_rcs[i]);
 692
 693                 if (remote_rcs[i] != 0) {
 694                         CERROR("rc[%d] invalid (%d) req %p\n",
 695                                 i, remote_rcs[i], request);
 696                         return(-EPROTO);
 697                 }
 698         }
 699
 700         if (request->rq_bulk->bd_nob_transferred != requested_nob) {
 701                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
 702                        requested_nob, request->rq_bulk->bd_nob_transferred);
 703                 return(-EPROTO);
 704         }
 705
 706         return (0);
 707 }
 708
 709 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
 710 {
 711         if (p1->flag != p2->flag) {
 712                 unsigned mask = ~OBD_BRW_FROM_GRANT;
 713
 714                 /* warn if we try to combine flags that we don't know to be
 715                  * safe to combine */
 716                 if ((p1->flag & mask) != (p2->flag & mask))
 717                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
 718                                "same brw?\n", p1->flag, p2->flag);
 719                 return 0;
 720         }
 721
 722         return (p1->disk_offset + p1->count == p2->disk_offset);
 723 }
 724
 725 #if CHECKSUM_BULK
 726 static obd_count cksum_pages(int nob, obd_count page_count,
 727                              struct brw_page *pga)
 728 {
 729         obd_count cksum = 0;
 730         char *ptr;
 731
 732         while (nob > 0) {
 733                 LASSERT (page_count > 0);
 734
 735                 ptr = kmap(pga->pg);
 736                 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
 737                              pga->count > nob ? nob : pga->count);
 738                 kunmap(pga->pg);
 739
 740                 nob -= pga->count;
 741                 page_count--;
 742                 pga++;
 743         }
 744
 745         return (cksum);
 746 }
 747 #endif
 748
 749 #define osc_encrypt_page(page, off, count)  \
 750         osc_crypt_page(page, off, count, ENCRYPT_DATA)
 751 #define osc_decrypt_page(page, off, count)  \
 752         osc_crypt_page(page, off, count, DECRYPT_DATA)
 753
 754 /*Put a global call back var here is Ugly, but put it to client_obd
 755  *also seems not a good idea, WangDi*/
 756 crypt_cb_t  osc_crypt_cb = NULL;
 757
 758 static int osc_crypt_page(struct page *page, obd_off page_off, obd_off count,
 759                           int flags)
 760 {
 761         int rc = 0;
 762         ENTRY;
 763
 764         if (osc_crypt_cb != NULL)
 765                 rc = osc_crypt_cb(page, page_off, count, flags);
 766         if (rc != 0)
 767                 CERROR("crypt page error %d \n", rc);
 768         RETURN(rc);
 769 }
 770
 771 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
 772                                 struct lov_stripe_md *lsm, obd_count page_count,
 773                                 struct brw_page *pga, int *requested_nobp,
 774                                 int *niocountp, struct ptlrpc_request **reqp)
 775 {
 776         struct ptlrpc_request   *req;
 777         struct ptlrpc_bulk_desc *desc;
 778         struct client_obd       *cli = &imp->imp_obd->u.cli;
 779         struct ost_body         *body;
 780         struct lustre_id        *raw_id = obdo_id(oa);
 781         struct obd_capa         *ocapa = NULL;
 782         struct lustre_capa      *capa = NULL;
 783         struct obd_ioobj        *ioobj;
 784         struct niobuf_remote    *niobuf;
 785         int                      niocount;
 786         int                      size[4];
 787         int                      i, bufcnt = 0;
 788         int                      requested_nob;
 789         int                      opc;
 790         int                      capa_op;
 791         int                      rc;
 792
 793         opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
 794
 795         for (niocount = i = 1; i < page_count; i++)
 796                 if (!can_merge_pages(&pga[i - 1], &pga[i]))
 797                         niocount++;
 798
 799         capa_op = (opc == OST_WRITE) ? MAY_WRITE : MAY_READ;
 800 get_capa:
 801         ocapa = capa_get(oa->o_fsuid, capa_op, raw_id->li_fid.lf_group,
 802                          raw_id->li_stc.u.e3s.l3s_ino, CLIENT_CAPA,
 803                          NULL, NULL, NULL);
 804         if (!ocapa) {
 805                 if (opc == OST_READ && capa_op == MAY_READ) {
 806                         /* partial write might cause read, MAY_WRITE capability
 807                          * should be used here */
 808                         capa_op = MAY_WRITE;
 809                         goto get_capa;
 810                 }
 811         }
 812
 813         size[bufcnt++] = sizeof(*body);
 814         size[bufcnt++] = sizeof(*ioobj);
 815         if (ocapa)
 816                 size[bufcnt++] = sizeof(*capa);
 817         size[bufcnt++] = niocount * sizeof(*niobuf);
 818
 819         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, opc, bufcnt, size, NULL);
 820         if (req == NULL)
 821                 return (-ENOMEM);
 822
 823         if (opc == OST_WRITE)
 824                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 825                                              BULK_GET_SOURCE, OST_BULK_PORTAL);
 826         else
 827                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 828                                              BULK_PUT_SINK, OST_BULK_PORTAL);
 829         if (desc == NULL)
 830                 GOTO(out, rc = -ENOMEM);
 831         /* NB request now owns desc and will free it when it gets freed */
 832
 833         bufcnt = 0;
 834         body = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*body));
 835         ioobj = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*ioobj));
 836         if (ocapa)
 837                 capa = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*capa));
 838         niobuf = lustre_msg_buf(req->rq_reqmsg, bufcnt++,
 839                                 niocount * sizeof(*niobuf));
 840
 841         memcpy(&body->oa, oa, sizeof(*oa));
 842
 843         obdo_to_ioobj(oa, ioobj);
 844         ioobj->ioo_bufcnt = niocount;
 845
 846         if (ocapa) {
 847                 capa_dup(capa, ocapa);
 848                 body->oa.o_valid |= OBD_MD_CAPA;
 849                 capa_put(ocapa, CLIENT_CAPA);
 850         }
 851
 852         LASSERT (page_count > 0);
 853
 854         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
 855                 struct brw_page *pg = &pga[i];
 856                 struct brw_page *pg_prev = pg - 1;
 857
 858                 LASSERT(pg->count > 0);
 859                 LASSERTF((pg->page_offset & ~PAGE_MASK)+ pg->count <= PAGE_SIZE,
 860                          "i: %d pg: %p pg_off: "LPU64", count: %u\n", i, pg,
 861                          pg->page_offset, pg->count);
 862                 LASSERTF(i == 0 || pg->disk_offset > pg_prev->disk_offset,
 863                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
 864                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
 865                          i, page_count,
 866                          pg->pg, pg->pg->private, pg->pg->index, pg->disk_offset,
 867                          pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
 868                          pg_prev->disk_offset);
 869
 870                 if (opc == OST_WRITE) {
 871                         rc = osc_encrypt_page(pg->pg, pg->page_offset, pg->count);
 872                         if (rc)
 873                                 GOTO(out, rc);
 874                 }
 875
 876                 ptlrpc_prep_bulk_page(desc, pg->pg,
 877                                       pg->page_offset & ~PAGE_MASK, pg->count);
 878                 requested_nob += pg->count;
 879
 880                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
 881                         niobuf--;
 882                         niobuf->len += pg->count;
 883                 } else {
 884                         niobuf->offset = pg->disk_offset;
 885                         niobuf->len    = pg->count;
 886                         niobuf->flags  = pg->flag;
 887                 }
 888         }
 889
 890         LASSERT((void *)(niobuf - niocount) ==
 891                 lustre_msg_buf(req->rq_reqmsg, bufcnt - 1,
 892                                niocount * sizeof(*niobuf)));
 893         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
 894
 895         /* size[0] still sizeof (*body) */
 896         if (opc == OST_WRITE) {
 897 #if CHECKSUM_BULK
 898                 body->oa.o_valid |= OBD_MD_FLCKSUM;
 899                 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
 900 #endif
 901                 /* 1 RC per niobuf */
 902                 size[1] = sizeof(__u32) * niocount;
 903                 req->rq_replen = lustre_msg_size(2, size);
 904         } else {
 905                 /* 1 RC for the whole I/O */
 906                 req->rq_replen = lustre_msg_size(1, size);
 907         }
 908
 909         *niocountp = niocount;
 910         *requested_nobp = requested_nob;
 911         *reqp = req;
 912         return (0);
 913
 914  out:
 915         ptlrpc_req_finished (req);
 916         return (rc);
 917 }
 918
 919 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
 920                                 int requested_nob, int niocount,
 921                                 obd_count page_count, struct brw_page *pga,
 922                                 int rc)
 923 {
 924         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 925         struct ost_body *body;
 926         ENTRY;
 927
 928         if (rc < 0)
 929                 RETURN(rc);
 930
 931         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 932         if (body == NULL) {
 933                 CERROR ("Can't unpack body\n");
 934                 RETURN(-EPROTO);
 935         }
 936
 937         osc_update_grant(cli, body);
 938         memcpy(oa, &body->oa, sizeof(*oa));
 939
 940         if (req->rq_reqmsg->opc == OST_WRITE) {
 941                 if (rc > 0) {
 942                         CERROR ("Unexpected +ve rc %d\n", rc);
 943                         RETURN(-EPROTO);
 944                 }
 945                 LASSERT (req->rq_bulk->bd_nob == requested_nob);
 946                 osc_decrypt_page(pga->pg, pga->page_offset,
 947                                  pga->count);
 948                 RETURN(check_write_rcs(req, requested_nob, niocount,
 949                                        page_count, pga));
 950         }
 951
 952         if (rc > requested_nob) {
 953                 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
 954                 RETURN(-EPROTO);
 955         }
 956
 957         if (rc != req->rq_bulk->bd_nob_transferred) {
 958                 CERROR ("Unexpected rc %d (%d transferred)\n",
 959                         rc, req->rq_bulk->bd_nob_transferred);
 960                 return (-EPROTO);
 961         }
 962
 963         if (rc < requested_nob)
 964                 handle_short_read(rc, page_count, pga);
 965
 966 #if CHECKSUM_BULK
 967         if (oa->o_valid & OBD_MD_FLCKSUM) {
 968                 const struct ptlrpc_peer *peer =
 969                         &req->rq_import->imp_connection->c_peer;
 970                 static int cksum_counter;
 971                 obd_count server_cksum = oa->o_cksum;
 972                 obd_count cksum = cksum_pages(rc, page_count, pga);
 973                 char str[PTL_NALFMT_SIZE];
 974
 975                 ptlrpc_peernid2str(peer, str);
 976
 977                 cksum_counter++;
 978                 if (server_cksum != cksum) {
 979                         CERROR("Bad checksum: server %x, client %x, server NID "
 980                                LPX64" (%s)\n", server_cksum, cksum,
 981                                peer->peer_id.nid, str);
 982                         cksum_counter = 0;
 983                         oa->o_cksum = cksum;
 984                 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
 985                         CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
 986                               cksum_counter, peer->peer_id.nid, str, cksum);
 987                 }
 988         } else {
 989                 static int cksum_missed;
 990
 991                 cksum_missed++;
 992                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
 993                         CERROR("Request checksum %u from "LPX64", no reply\n",
 994                                cksum_missed,
 995                                req->rq_import->imp_connection->c_peer.peer_id.nid);
 996         }
 997 #endif
 998         osc_decrypt_page(pga->pg, pga->page_offset, pga->count);
 999         RETURN(0);
1000 }
1001
1002 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
1003                             struct lov_stripe_md *lsm,
1004                             obd_count page_count, struct brw_page *pga)
1005 {
1006         int                    requested_nob;
1007         int                    niocount;
1008         struct ptlrpc_request *request;
1009         int                    rc;
1010         ENTRY;
1011
1012 restart_bulk:
1013         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1014                                   page_count, pga, &requested_nob, &niocount,
1015                                   &request);
1016         if (rc != 0)
1017                 return (rc);
1018
1019         rc = ptlrpc_queue_wait(request);
1020
1021         if (rc == -ETIMEDOUT && request->rq_resend) {
1022                 DEBUG_REQ(D_HA, request,  "BULK TIMEOUT");
1023                 ptlrpc_req_finished(request);
1024                 goto restart_bulk;
1025         }
1026
1027         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
1028                                   page_count, pga, rc);
1029
1030         ptlrpc_req_finished(request);
1031         RETURN (rc);
1032 }
1033
1034 static int brw_interpret(struct ptlrpc_request *request,
1035                          struct osc_brw_async_args *aa, int rc)
1036 {
1037         struct obdo *oa      = aa->aa_oa;
1038         int requested_nob    = aa->aa_requested_nob;
1039         int niocount         = aa->aa_nio_count;
1040         obd_count page_count = aa->aa_page_count;
1041         struct brw_page *pga = aa->aa_pga;
1042         ENTRY;
1043
1044         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
1045                                   page_count, pga, rc);
1046         RETURN (rc);
1047 }
1048
1049 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1050                           struct lov_stripe_md *lsm, obd_count page_count,
1051                           struct brw_page *pga, struct ptlrpc_request_set *set)
1052 {
1053         struct ptlrpc_request     *request;
1054         int                        requested_nob;
1055         int                        nio_count;
1056         struct osc_brw_async_args *aa;
1057         int                        rc;
1058         ENTRY;
1059
1060         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1061                                   page_count, pga, &requested_nob, &nio_count,
1062                                   &request);
1063         if (rc == 0) {
1064                 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1065                 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1066                 aa->aa_oa = oa;
1067                 aa->aa_requested_nob = requested_nob;
1068                 aa->aa_nio_count = nio_count;
1069                 aa->aa_page_count = page_count;
1070                 aa->aa_pga = pga;
1071
1072                 request->rq_interpret_reply = brw_interpret;
1073                 ptlrpc_set_add_req(set, request);
1074         }
1075         RETURN (rc);
1076 }
1077
1078 #ifndef min_t
1079 #define min_t(type,x,y) \
1080         ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
1081 #endif
1082
1083 /*
1084  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1085  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1086  * fine for our small page arrays and doesn't require allocation.  its an
1087  * insertion sort that swaps elements that are strides apart, shrinking the
1088  * stride down until its '1' and the array is sorted.
1089  */
1090 static void sort_brw_pages(struct brw_page *array, int num)
1091 {
1092         int stride, i, j;
1093         struct brw_page tmp;
1094
1095         if (num == 1)
1096                 return;
1097         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1098                 ;
1099
1100         do {
1101                 stride /= 3;
1102                 for (i = stride ; i < num ; i++) {
1103                         tmp = array[i];
1104                         j = i;
1105                         while (j >= stride && array[j - stride].disk_offset >
1106                                 tmp.disk_offset) {
1107                                 array[j] = array[j - stride];
1108                                 j -= stride;
1109                         }
1110                         array[j] = tmp;
1111                 }
1112         } while (stride > 1);
1113 }
1114
1115 /* make sure we the regions we're passing to elan don't violate its '4
1116  * fragments' constraint.  portal headers are a fragment, all full
1117  * PAGE_SIZE long pages count as 1 fragment, and each partial page
1118  * counts as a fragment.  I think.  see bug 934. */
1119 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1120 {
1121         int frags_left = 3;
1122         int saw_whole_frag = 0;
1123         int i;
1124
1125         for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1126                 if (pg->count == PAGE_SIZE) {
1127                         if (!saw_whole_frag) {
1128                                 saw_whole_frag = 1;
1129                                 frags_left--;
1130                         }
1131                 } else {
1132                         frags_left--;
1133                 }
1134         }
1135         return i;
1136 }
1137
1138 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1139                    struct lov_stripe_md *lsm, obd_count page_count,
1140                    struct brw_page *pga, struct obd_trans_info *oti)
1141 {
1142         ENTRY;
1143
1144         if (cmd == OBD_BRW_CHECK) {
1145                 /* The caller just wants to know if there's a chance that this
1146                  * I/O can succeed */
1147                 struct obd_import *imp = class_exp2cliimp(exp);
1148
1149                 if (imp == NULL || imp->imp_invalid)
1150                         RETURN(-EIO);
1151                 RETURN(0);
1152         }
1153
1154         while (page_count) {
1155                 obd_count pages_per_brw;
1156                 int rc;
1157
1158                 if (page_count > PTLRPC_MAX_BRW_PAGES)
1159                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1160                 else
1161                         pages_per_brw = page_count;
1162
1163                 sort_brw_pages(pga, pages_per_brw);
1164                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1165
1166                 rc = osc_brw_internal(cmd, exp, oa, lsm, pages_per_brw, pga);
1167
1168                 if (rc != 0)
1169                         RETURN(rc);
1170
1171                 page_count -= pages_per_brw;
1172                 pga += pages_per_brw;
1173         }
1174         RETURN(0);
1175 }
1176
1177 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1178                          struct lov_stripe_md *lsm, obd_count page_count,
1179                          struct brw_page *pga, struct ptlrpc_request_set *set,
1180                          struct obd_trans_info *oti)
1181 {
1182         ENTRY;
1183
1184         if (cmd == OBD_BRW_CHECK) {
1185                 /* The caller just wants to know if there's a chance that this
1186                  * I/O can succeed */
1187                 struct obd_import *imp = class_exp2cliimp(exp);
1188
1189                 if (imp == NULL || imp->imp_invalid)
1190                         RETURN(-EIO);
1191                 RETURN(0);
1192         }
1193
1194         while (page_count) {
1195                 obd_count pages_per_brw;
1196                 int rc;
1197
1198                 if (page_count > PTLRPC_MAX_BRW_PAGES)
1199                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1200                 else
1201                         pages_per_brw = page_count;
1202
1203                 sort_brw_pages(pga, pages_per_brw);
1204                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1205
1206                 rc = async_internal(cmd, exp, oa, lsm, pages_per_brw, pga, set);
1207
1208                 if (rc != 0)
1209                         RETURN(rc);
1210
1211                 page_count -= pages_per_brw;
1212                 pga += pages_per_brw;
1213         }
1214         RETURN(0);
1215 }
1216
1217 static void osc_check_rpcs(struct client_obd *cli);
1218 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1219                            int sent);
1220 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1221 static void lop_update_pending(struct client_obd *cli,
1222                                struct loi_oap_pages *lop, int cmd, int delta);
1223
1224 /* this is called when a sync waiter receives an interruption.  Its job is to
1225  * get the caller woken as soon as possible.  If its page hasn't been put in an
1226  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
1227  * desiring interruption which will forcefully complete the rpc once the rpc
1228  * has timed out */
1229 static void osc_occ_interrupted(struct oig_callback_context *occ)
1230 {
1231         struct osc_async_page *oap;
1232         struct loi_oap_pages *lop;
1233         struct lov_oinfo *loi;
1234         ENTRY;
1235
1236         /* XXX member_of() */
1237         oap = list_entry(occ, struct osc_async_page, oap_occ);
1238
1239         spin_lock(&oap->oap_cli->cl_loi_list_lock);
1240
1241         oap->oap_interrupted = 1;
1242
1243         /* ok, it's been put in an rpc. */
1244         if (oap->oap_request != NULL) {
1245                 ptlrpc_mark_interrupted(oap->oap_request);
1246                 ptlrpcd_wake(oap->oap_request);
1247                 GOTO(unlock, 0);
1248         }
1249
1250         /* we don't get interruption callbacks until osc_trigger_sync_io()
1251          * has been called and put the sync oaps in the pending/urgent lists.*/
1252         if (!list_empty(&oap->oap_pending_item)) {
1253                 list_del_init(&oap->oap_pending_item);
1254                 if (oap->oap_async_flags & ASYNC_URGENT)
1255                         list_del_init(&oap->oap_urgent_item);
1256
1257                 loi = oap->oap_loi;
1258                 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1259                         &loi->loi_write_lop : &loi->loi_read_lop;
1260                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1261                 loi_list_maint(oap->oap_cli, oap->oap_loi);
1262
1263                 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1264                 oap->oap_oig = NULL;
1265         }
1266
1267 unlock:
1268         spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1269 }
1270
1271 /* this must be called holding the loi list lock to give coverage to exit_cache,
1272  * async_flag maintenance, and oap_request */
1273 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1274                               struct osc_async_page *oap, int sent, int rc)
1275 {
1276         osc_exit_cache(cli, oap, sent);
1277         oap->oap_async_flags = 0;
1278         oap->oap_interrupted = 0;
1279
1280         if (oap->oap_request != NULL) {
1281                 ptlrpc_req_finished(oap->oap_request);
1282                 oap->oap_request = NULL;
1283         }
1284
1285         if (rc == 0 && oa != NULL)
1286                 oap->oap_loi->loi_blocks = oa->o_blocks;
1287
1288         if (oap->oap_oig) {
1289                 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1290                 oap->oap_oig = NULL;
1291                 EXIT;
1292                 return;
1293         }
1294
1295         oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1296                                            oa, rc);
1297 }
1298
1299 static int brw_interpret_oap(struct ptlrpc_request *request,
1300                              struct osc_brw_async_args *aa, int rc)
1301 {
1302         struct osc_async_page *oap;
1303         struct client_obd *cli;
1304         struct list_head *pos, *n;
1305         struct timeval now;
1306         ENTRY;
1307
1308         do_gettimeofday(&now);
1309         rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1310                                   aa->aa_nio_count, aa->aa_page_count,
1311                                   aa->aa_pga, rc);
1312
1313         CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1314
1315         cli = aa->aa_cli;
1316         /* in failout recovery we ignore writeback failure and want
1317          * to just tell llite to unlock the page and continue */
1318         if (request->rq_reqmsg->opc == OST_WRITE &&
1319             (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1320                 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1321                        cli->cl_import,
1322                        cli->cl_import ? cli->cl_import->imp_invalid : -1);
1323                 rc = 0;
1324         }
1325
1326         spin_lock(&cli->cl_loi_list_lock);
1327
1328         if (request->rq_reqmsg->opc == OST_WRITE)
1329                 lprocfs_stime_record(&cli->cl_write_stime, &now,
1330                                      &request->rq_rpcd_start);
1331         else
1332                 lprocfs_stime_record(&cli->cl_read_stime, &now,
1333                                      &request->rq_rpcd_start);
1334
1335
1336
1337         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1338          * is called so we know whether to go to sync BRWs or wait for more
1339          * RPCs to complete */
1340         if (request->rq_reqmsg->opc == OST_WRITE)
1341                 cli->cl_w_in_flight--;
1342         else
1343                 cli->cl_r_in_flight--;
1344
1345         /* the caller may re-use the oap after the completion call so
1346          * we need to clean it up a little */
1347         list_for_each_safe(pos, n, &aa->aa_oaps) {
1348                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1349
1350                 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1351                        //oap->oap_page, oap->oap_page->index, oap);
1352
1353                 list_del_init(&oap->oap_rpc_item);
1354                 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1355         }
1356
1357         /* no write RPCs in flight, reset the time */
1358         if (request->rq_reqmsg->opc == OST_WRITE && cli->cl_w_in_flight == 0)
1359                 do_gettimeofday(&cli->cl_last_write_time);
1360
1361         osc_wake_cache_waiters(cli);
1362         osc_check_rpcs(cli);
1363         spin_unlock(&cli->cl_loi_list_lock);
1364
1365         obdo_free(aa->aa_oa);
1366         OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1367
1368         RETURN(0);
1369 }
1370
1371 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1372                                             struct list_head *rpc_list,
1373                                             int page_count, int cmd)
1374 {
1375         struct ptlrpc_request *req;
1376         struct brw_page *pga = NULL;
1377         int requested_nob, nio_count;
1378         struct osc_brw_async_args *aa;
1379         struct obdo *oa = NULL;
1380         struct obd_async_page_ops *ops = NULL;
1381         void *caller_data = NULL;
1382         struct list_head *pos;
1383         int i, rc;
1384
1385         LASSERT(!list_empty(rpc_list));
1386
1387         OBD_ALLOC(pga, sizeof(*pga) * page_count);
1388         if (pga == NULL)
1389                 RETURN(ERR_PTR(-ENOMEM));
1390
1391         oa = obdo_alloc();
1392         if (oa == NULL)
1393                 GOTO(out, req = ERR_PTR(-ENOMEM));
1394
1395         i = 0;
1396         list_for_each(pos, rpc_list) {
1397                 struct osc_async_page *oap;
1398
1399                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1400                 if (ops == NULL) {
1401                         ops = oap->oap_caller_ops;
1402                         caller_data = oap->oap_caller_data;
1403                 }
1404                 pga[i].disk_offset = oap->oap_obj_off + oap->oap_page_off;
1405                 pga[i].page_offset = pga[i].disk_offset;
1406                 pga[i].pg = oap->oap_page;
1407                 pga[i].count = oap->oap_count;
1408                 pga[i].flag = oap->oap_brw_flags;
1409                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1410                        pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1411                 i++;
1412         }
1413
1414         /* always get the data for the obdo for the rpc */
1415         LASSERT(ops != NULL);
1416         ops->ap_fill_obdo(caller_data, cmd, oa);
1417
1418         sort_brw_pages(pga, page_count);
1419         rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1420                                   pga, &requested_nob, &nio_count, &req);
1421         if (rc != 0) {
1422                 CERROR("prep_req failed: %d\n", rc);
1423                 GOTO(out, req = ERR_PTR(rc));
1424         }
1425
1426         LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1427         aa = (struct osc_brw_async_args *)&req->rq_async_args;
1428         aa->aa_oa = oa;
1429         aa->aa_requested_nob = requested_nob;
1430         aa->aa_nio_count = nio_count;
1431         aa->aa_page_count = page_count;
1432         aa->aa_pga = pga;
1433         aa->aa_cli = cli;
1434
1435 out:
1436         if (IS_ERR(req)) {
1437                 if (oa)
1438                         obdo_free(oa);
1439                 if (pga)
1440                         OBD_FREE(pga, sizeof(*pga) * page_count);
1441         }
1442         RETURN(req);
1443 }
1444
1445 /* strange write gap too long (15s) */
1446 #define CLI_ODD_WRITE_GAP 15000000
1447
1448 static void lop_update_pending(struct client_obd *cli,
1449                                struct loi_oap_pages *lop, int cmd, int delta)
1450 {
1451         lop->lop_num_pending += delta;
1452         if (cmd == OBD_BRW_WRITE)
1453                 cli->cl_pending_w_pages += delta;
1454         else
1455                 cli->cl_pending_r_pages += delta;
1456 }
1457
1458 static long timeval_sub(struct timeval *large, struct timeval *small)
1459 {
1460         return (large->tv_sec - small->tv_sec) * 1000000 +
1461                 (large->tv_usec - small->tv_usec);
1462 }
1463
1464 /* the loi lock is held across this function but it's allowed to release
1465  * and reacquire it during its work */
1466 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1467                             int cmd, struct loi_oap_pages *lop)
1468 {
1469         struct ptlrpc_request *request;
1470         obd_count page_count = 0;
1471         struct list_head *tmp, *pos;
1472         struct osc_async_page *oap = NULL;
1473         struct osc_brw_async_args *aa;
1474         struct obd_async_page_ops *ops;
1475         LIST_HEAD(rpc_list);
1476         ENTRY;
1477
1478         /* first we find the pages we're allowed to work with */
1479         list_for_each_safe(pos, tmp, &lop->lop_pending) {
1480                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1481                 ops = oap->oap_caller_ops;
1482
1483                 LASSERT(oap->oap_magic == OAP_MAGIC);
1484
1485                 /* in llite being 'ready' equates to the page being locked
1486                  * until completion unlocks it.  commit_write submits a page
1487                  * as not ready because its unlock will happen unconditionally
1488                  * as the call returns.  if we race with commit_write giving
1489                  * us that page we dont' want to create a hole in the page
1490                  * stream, so we stop and leave the rpc to be fired by
1491                  * another dirtier or kupdated interval (the not ready page
1492                  * will still be on the dirty list).  we could call in
1493                  * at the end of ll_file_write to process the queue again. */
1494                 if (!(oap->oap_async_flags & ASYNC_READY)) {
1495                         int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1496                         if (rc < 0)
1497                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
1498                                                 "instead of ready\n", oap,
1499                                                 oap->oap_page, rc);
1500                         switch (rc) {
1501                         case -EAGAIN:
1502                                 /* llite is telling us that the page is still
1503                                  * in commit_write and that we should try
1504                                  * and put it in an rpc again later.  we
1505                                  * break out of the loop so we don't create
1506                                  * a hole in the sequence of pages in the rpc
1507                                  * stream.*/
1508                                 pos = NULL;
1509                                 break;
1510                         case -EINTR:
1511                                 /* the io isn't needed.. tell the checks
1512                                  * below to complete the rpc with EINTR */
1513                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1514                                 oap->oap_count = -EINTR;
1515                                 break;
1516                         case 0:
1517                                 oap->oap_async_flags |= ASYNC_READY;
1518                                 break;
1519                         default:
1520                                 LASSERTF(0, "oap %p page %p returned %d "
1521                                             "from make_ready\n", oap,
1522                                             oap->oap_page, rc);
1523                                 break;
1524                         }
1525                 }
1526                 if (pos == NULL)
1527                         break;
1528                 /*
1529                  * Page submitted for IO has to be locked. Either by
1530                  * ->ap_make_ready() or by higher layers.
1531                  *
1532                  * XXX nikita: this assertion should be adjusted when lustre
1533                  * starts using PG_writeback for pages being written out.
1534                  */
1535                 LASSERT(PageLocked(oap->oap_page));
1536
1537                 /* take the page out of our book-keeping */
1538                 list_del_init(&oap->oap_pending_item);
1539                 lop_update_pending(cli, lop, cmd, -1);
1540                 list_del_init(&oap->oap_urgent_item);
1541
1542                 /* ask the caller for the size of the io as the rpc leaves. */
1543                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1544                         oap->oap_count =
1545                                 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1546                 if (oap->oap_count <= 0) {
1547                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1548                                oap->oap_count);
1549                         osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1550                         continue;
1551                 }
1552
1553                 /* now put the page back in our accounting */
1554                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1555                 if (++page_count >= cli->cl_max_pages_per_rpc)
1556                         break;
1557         }
1558
1559         osc_wake_cache_waiters(cli);
1560
1561         if (page_count == 0)
1562                 RETURN(0);
1563
1564         loi_list_maint(cli, loi);
1565         spin_unlock(&cli->cl_loi_list_lock);
1566
1567         request = osc_build_req(cli, &rpc_list, page_count, cmd);
1568         if (IS_ERR(request)) {
1569                 /* this should happen rarely and is pretty bad, it makes the
1570                  * pending list not follow the dirty order */
1571                 spin_lock(&cli->cl_loi_list_lock);
1572                 list_for_each_safe(pos, tmp, &rpc_list) {
1573                         oap = list_entry(pos, struct osc_async_page,
1574                                          oap_rpc_item);
1575                         list_del_init(&oap->oap_rpc_item);
1576
1577                         /* queued sync pages can be torn down while the pages
1578                          * were between the pending list and the rpc */
1579                         if (oap->oap_interrupted) {
1580                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1581                                 osc_ap_completion(cli, NULL, oap, 0,
1582                                                   oap->oap_count);
1583                                 continue;
1584                         }
1585
1586                         /* put the page back in the loi/lop lists */
1587                         list_add_tail(&oap->oap_pending_item,
1588                                       &lop->lop_pending);
1589                         lop_update_pending(cli, lop, cmd, 1);
1590                         if (oap->oap_async_flags & ASYNC_URGENT)
1591                                 list_add(&oap->oap_urgent_item,
1592                                          &lop->lop_urgent);
1593                 }
1594                 loi_list_maint(cli, loi);
1595                 RETURN(PTR_ERR(request));
1596         }
1597
1598         LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1599         aa = (struct osc_brw_async_args *)&request->rq_async_args;
1600         INIT_LIST_HEAD(&aa->aa_oaps);
1601         list_splice(&rpc_list, &aa->aa_oaps);
1602         INIT_LIST_HEAD(&rpc_list);
1603
1604 #ifdef __KERNEL__
1605         if (cmd == OBD_BRW_READ) {
1606                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1607                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1608         } else {
1609                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1610                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1611                                  cli->cl_w_in_flight);
1612         }
1613 #endif
1614
1615         spin_lock(&cli->cl_loi_list_lock);
1616
1617         /* collect write gaps and sum of them */
1618         if (cmd == OBD_BRW_WRITE && cli->cl_w_in_flight == 0) {
1619                 struct timeval now;
1620                 long diff;
1621
1622                 do_gettimeofday(&now);
1623
1624                 if (cli->cl_last_write_time.tv_sec) {
1625                         diff = timeval_sub(&now, &cli->cl_last_write_time);
1626                         if (diff < CLI_ODD_WRITE_GAP) {
1627                                 cli->cl_write_gap_sum += diff;
1628                                 cli->cl_write_gaps++;
1629                         }
1630                 }
1631         }
1632
1633         if (cmd == OBD_BRW_READ)
1634                 cli->cl_r_in_flight++;
1635         else
1636                 cli->cl_w_in_flight++;
1637
1638         /* queued sync pages can be torn down while the pages
1639          * were between the pending list and the rpc */
1640         list_for_each(pos, &aa->aa_oaps) {
1641                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1642                 if (oap->oap_interrupted) {
1643                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1644                                oap, request);
1645                         ptlrpc_mark_interrupted(request);
1646                         break;
1647                 }
1648         }
1649
1650         CDEBUG(D_INODE, "req %p: %d pages, aa %p.  now %dr/%dw in flight\n",
1651                         request, page_count, aa, cli->cl_r_in_flight,
1652                         cli->cl_w_in_flight);
1653
1654         oap->oap_request = ptlrpc_request_addref(request);
1655         request->rq_interpret_reply = brw_interpret_oap;
1656
1657         ptlrpcd_add_req(request);
1658         RETURN(1);
1659 }
1660
1661 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1662                          int cmd)
1663 {
1664         int optimal;
1665         ENTRY;
1666
1667         if (lop->lop_num_pending == 0)
1668                 RETURN(0);
1669
1670         /* if we have an invalid import we want to drain the queued pages
1671          * by forcing them through rpcs that immediately fail and complete
1672          * the pages.  recovery relies on this to empty the queued pages
1673          * before canceling the locks and evicting down the llite pages */
1674         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1675                 RETURN(1);
1676
1677         /* stream rpcs in queue order as long as as there is an urgent page
1678          * queued.  this is our cheap solution for good batching in the case
1679          * where writepage marks some random page in the middle of the file as
1680          * urgent because of, say, memory pressure */
1681         if (!list_empty(&lop->lop_urgent))
1682                 RETURN(1);
1683
1684         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1685         optimal = cli->cl_max_pages_per_rpc;
1686         if (cmd == OBD_BRW_WRITE) {
1687                 /* trigger a write rpc stream as long as there are dirtiers
1688                  * waiting for space.  as they're waiting, they're not going to
1689                  * create more pages to coallesce with what's waiting.. */
1690                 if (!list_empty(&cli->cl_cache_waiters))
1691                         RETURN(1);
1692
1693                 /* *2 to avoid triggering rpcs that would want to include pages
1694                  * that are being queued but which can't be made ready until
1695                  * the queuer finishes with the page. this is a wart for
1696                  * llite::commit_write() */
1697                 optimal += 16;
1698         }
1699         if (lop->lop_num_pending >= optimal)
1700                 RETURN(1);
1701
1702         RETURN(0);
1703 }
1704
1705 static void on_list(struct list_head *item, struct list_head *list,
1706                     int should_be_on)
1707 {
1708         if (list_empty(item) && should_be_on)
1709                 list_add_tail(item, list);
1710         else if (!list_empty(item) && !should_be_on)
1711                 list_del_init(item);
1712 }
1713
1714 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1715  * can find pages to build into rpcs quickly */
1716 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1717 {
1718         on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1719                 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1720                 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1721
1722         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1723                 loi->loi_write_lop.lop_num_pending);
1724
1725         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1726                 loi->loi_read_lop.lop_num_pending);
1727 }
1728
1729 #define LOI_DEBUG(LOI, STR, args...)                                     \
1730         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
1731                !list_empty(&(LOI)->loi_cli_item),                        \
1732                (LOI)->loi_write_lop.lop_num_pending,                     \
1733                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
1734                (LOI)->loi_read_lop.lop_num_pending,                      \
1735                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
1736                args)                                                     \
1737
1738 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1739 {
1740         ENTRY;
1741         /* first return all objects which we already know to have
1742          * pages ready to be stuffed into rpcs */
1743         if (!list_empty(&cli->cl_loi_ready_list))
1744                 RETURN(list_entry(cli->cl_loi_ready_list.next,
1745                                   struct lov_oinfo, loi_cli_item));
1746
1747         /* then if we have cache waiters, return all objects with queued
1748          * writes.  This is especially important when many small files
1749          * have filled up the cache and not been fired into rpcs because
1750          * they don't pass the nr_pending/object threshhold */
1751         if (!list_empty(&cli->cl_cache_waiters) &&
1752             !list_empty(&cli->cl_loi_write_list))
1753                 RETURN(list_entry(cli->cl_loi_write_list.next,
1754                                   struct lov_oinfo, loi_write_item));
1755
1756         /* then return all queued objects when we have an invalid import
1757          * so that they get flushed */
1758         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1759                 if (!list_empty(&cli->cl_loi_write_list))
1760                         RETURN(list_entry(cli->cl_loi_write_list.next,
1761                                           struct lov_oinfo, loi_write_item));
1762                 if (!list_empty(&cli->cl_loi_read_list))
1763                         RETURN(list_entry(cli->cl_loi_read_list.next,
1764                                           struct lov_oinfo, loi_read_item));
1765         }
1766         RETURN(NULL);
1767 }
1768
1769 /* called with the loi list lock held */
1770 static void osc_check_rpcs(struct client_obd *cli)
1771 {
1772         struct lov_oinfo *loi;
1773         int rc = 0, race_counter = 0;
1774         ENTRY;
1775
1776         while ((loi = osc_next_loi(cli)) != NULL) {
1777                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
1778
1779                 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
1780                         break;
1781
1782                 /* attempt some read/write balancing by alternating between
1783                  * reads and writes in an object.  The makes_rpc checks here
1784                  * would be redundant if we were getting read/write work items
1785                  * instead of objects.  we don't want send_oap_rpc to drain a
1786                  * partial read pending queue when we're given this object to
1787                  * do io on writes while there are cache waiters */
1788                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1789                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1790                                               &loi->loi_write_lop);
1791                         if (rc < 0)
1792                                 break;
1793                         if (rc > 0)
1794                                 race_counter = 0;
1795                         else
1796                                 race_counter++;
1797                 }
1798                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1799                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1800                                               &loi->loi_read_lop);
1801                         if (rc < 0)
1802                                 break;
1803                         if (rc > 0)
1804                                 race_counter = 0;
1805                         else
1806                                 race_counter++;
1807                 }
1808
1809                 /* attempt some inter-object balancing by issueing rpcs
1810                  * for each object in turn */
1811                 if (!list_empty(&loi->loi_cli_item))
1812                         list_del_init(&loi->loi_cli_item);
1813                 if (!list_empty(&loi->loi_write_item))
1814                         list_del_init(&loi->loi_write_item);
1815                 if (!list_empty(&loi->loi_read_item))
1816                         list_del_init(&loi->loi_read_item);
1817
1818                 loi_list_maint(cli, loi);
1819
1820                 /* send_oap_rpc fails with 0 when make_ready tells it to
1821                  * back off.  llite's make_ready does this when it tries
1822                  * to lock a page queued for write that is already locked.
1823                  * we want to try sending rpcs from many objects, but we
1824                  * don't want to spin failing with 0.  */
1825                 if (race_counter == 10)
1826                         break;
1827         }
1828         EXIT;
1829 }
1830
1831 /* we're trying to queue a page in the osc so we're subject to the
1832  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1833  * If the osc's queued pages are already at that limit, then we want to sleep
1834  * until there is space in the osc's queue for us.  We also may be waiting for
1835  * write credits from the OST if there are RPCs in flight that may return some
1836  * before we fall back to sync writes.
1837  *
1838  * We need this know our allocation was granted in the presence of signals */
1839 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1840 {
1841         int rc;
1842         ENTRY;
1843         spin_lock(&cli->cl_loi_list_lock);
1844         rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
1845         spin_unlock(&cli->cl_loi_list_lock);
1846         RETURN(rc);
1847 };
1848
1849 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1850  * grant or cache space. */
1851 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1852                            struct osc_async_page *oap)
1853 {
1854         struct osc_cache_waiter ocw;
1855         struct l_wait_info lwi = { 0 };
1856         struct timeval start, stop;
1857
1858         CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1859                cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1860                cli->cl_avail_grant);
1861
1862         if (cli->cl_dirty_max < PAGE_SIZE)
1863                 return(-EDQUOT);
1864
1865         /* Hopefully normal case - cache space and write credits available */
1866         if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1867             cli->cl_avail_grant >= PAGE_SIZE) {
1868                 /* account for ourselves */
1869                 osc_consume_write_grant(cli, oap);
1870                 return(0);
1871         }
1872
1873         /* Make sure that there are write rpcs in flight to wait for.  This
1874          * is a little silly as this object may not have any pending but
1875          * other objects sure might. */
1876         if (cli->cl_w_in_flight) {
1877                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1878                 init_waitqueue_head(&ocw.ocw_waitq);
1879                 ocw.ocw_oap = oap;
1880                 ocw.ocw_rc = 0;
1881
1882                 loi_list_maint(cli, loi);
1883                 osc_check_rpcs(cli);
1884                 spin_unlock(&cli->cl_loi_list_lock);
1885
1886                 CDEBUG(0, "sleeping for cache space\n");
1887                 do_gettimeofday(&start);
1888                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1889                 do_gettimeofday(&stop);
1890                 spin_lock(&cli->cl_loi_list_lock);
1891                 lprocfs_stime_record(&cli->cl_enter_stime, &stop, &start);
1892                 if (!list_empty(&ocw.ocw_entry)) {
1893                         list_del(&ocw.ocw_entry);
1894                         RETURN(-EINTR);
1895                 }
1896                 RETURN(ocw.ocw_rc);
1897         }
1898
1899         RETURN(-EDQUOT);
1900 }
1901
1902 /* the companion to enter_cache, called when an oap is no longer part of the
1903  * dirty accounting.. so writeback completes or truncate happens before writing
1904  * starts.  must be called with the loi lock held. */
1905 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1906                            int sent)
1907 {
1908         ENTRY;
1909
1910         if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1911                 EXIT;
1912                 return;
1913         }
1914
1915         oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1916         cli->cl_dirty -= PAGE_SIZE;
1917         if (!sent) {
1918                 cli->cl_lost_grant += PAGE_SIZE;
1919                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1920                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1921         }
1922
1923         EXIT;
1924 }
1925
1926 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1927                         struct lov_oinfo *loi, struct page *page,
1928                         obd_off offset, struct obd_async_page_ops *ops,
1929                         void *data, void **res)
1930 {
1931         struct osc_async_page *oap;
1932         ENTRY;
1933
1934         OBD_ALLOC(oap, sizeof(*oap));
1935         if (oap == NULL)
1936                 return -ENOMEM;
1937
1938         oap->oap_magic = OAP_MAGIC;
1939         oap->oap_cli = &exp->exp_obd->u.cli;
1940         oap->oap_loi = loi;
1941
1942         oap->oap_caller_ops = ops;
1943         oap->oap_caller_data = data;
1944
1945         oap->oap_page = page;
1946         oap->oap_obj_off = offset;
1947
1948         INIT_LIST_HEAD(&oap->oap_pending_item);
1949         INIT_LIST_HEAD(&oap->oap_urgent_item);
1950         INIT_LIST_HEAD(&oap->oap_rpc_item);
1951
1952         oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1953
1954         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1955         *res = oap;
1956         RETURN(0);
1957 }
1958
1959 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1960                               struct lov_oinfo *loi, void *cookie,
1961                               int cmd, obd_off off, int count,
1962                               obd_flags brw_flags, enum async_flags async_flags)
1963 {
1964         struct client_obd *cli = &exp->exp_obd->u.cli;
1965         struct osc_async_page *oap;
1966         struct loi_oap_pages *lop;
1967         int rc;
1968         ENTRY;
1969
1970         oap = OAP_FROM_COOKIE(cookie);
1971
1972         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1973                 RETURN(-EIO);
1974
1975         if (!list_empty(&oap->oap_pending_item) ||
1976             !list_empty(&oap->oap_urgent_item) ||
1977             !list_empty(&oap->oap_rpc_item))
1978                 RETURN(-EBUSY);
1979
1980         if (loi == NULL)
1981                 loi = &lsm->lsm_oinfo[0];
1982
1983         spin_lock(&cli->cl_loi_list_lock);
1984
1985         oap->oap_cmd = cmd;
1986         oap->oap_async_flags = async_flags;
1987         oap->oap_page_off = off;
1988         oap->oap_count = count;
1989         oap->oap_brw_flags = brw_flags;
1990
1991         if (cmd == OBD_BRW_WRITE) {
1992                 rc = osc_enter_cache(cli, loi, oap);
1993                 if (rc) {
1994                         spin_unlock(&cli->cl_loi_list_lock);
1995                         RETURN(rc);
1996                 }
1997                 lop = &loi->loi_write_lop;
1998         } else {
1999                 lop = &loi->loi_read_lop;
2000         }
2001
2002         if (oap->oap_async_flags & ASYNC_URGENT)
2003                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2004         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2005         lop_update_pending(cli, lop, cmd, 1);
2006
2007         loi_list_maint(cli, loi);
2008
2009         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2010                   cmd);
2011
2012         osc_check_rpcs(cli);
2013         spin_unlock(&cli->cl_loi_list_lock);
2014
2015         RETURN(0);
2016 }
2017
2018 /* aka (~was & now & flag), but this is more clear :) */
2019 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2020
2021 static int osc_set_async_flags(struct obd_export *exp,
2022                                struct lov_stripe_md *lsm,
2023                                struct lov_oinfo *loi, void *cookie,
2024                                obd_flags async_flags)
2025 {
2026         struct client_obd *cli = &exp->exp_obd->u.cli;
2027         struct loi_oap_pages *lop;
2028         struct osc_async_page *oap;
2029         int rc = 0;
2030         ENTRY;
2031
2032         oap = OAP_FROM_COOKIE(cookie);
2033
2034         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2035                 RETURN(-EIO);
2036
2037         if (loi == NULL)
2038                 loi = &lsm->lsm_oinfo[0];
2039
2040         if (oap->oap_cmd == OBD_BRW_WRITE) {
2041                 lop = &loi->loi_write_lop;
2042         } else {
2043                 lop = &loi->loi_read_lop;
2044         }
2045
2046         spin_lock(&cli->cl_loi_list_lock);
2047
2048         if (list_empty(&oap->oap_pending_item))
2049                 GOTO(out, rc = -EINVAL);
2050
2051         if ((oap->oap_async_flags & async_flags) == async_flags)
2052                 GOTO(out, rc = 0);
2053
2054         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2055                 oap->oap_async_flags |= ASYNC_READY;
2056
2057         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2058                 if (list_empty(&oap->oap_rpc_item)) {
2059                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2060                         loi_list_maint(cli, loi);
2061                 }
2062         }
2063
2064         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2065                         oap->oap_async_flags);
2066 out:
2067         osc_check_rpcs(cli);
2068         spin_unlock(&cli->cl_loi_list_lock);
2069         RETURN(rc);
2070 }
2071
2072 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2073                              struct lov_oinfo *loi,
2074                              struct obd_io_group *oig, void *cookie,
2075                              int cmd, obd_off off, int count,
2076                              obd_flags brw_flags,
2077                              obd_flags async_flags)
2078 {
2079         struct client_obd *cli = &exp->exp_obd->u.cli;
2080         struct osc_async_page *oap;
2081         struct loi_oap_pages *lop;
2082         ENTRY;
2083
2084         oap = OAP_FROM_COOKIE(cookie);
2085
2086         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2087                 RETURN(-EIO);
2088
2089         if (!list_empty(&oap->oap_pending_item) ||
2090             !list_empty(&oap->oap_urgent_item) ||
2091             !list_empty(&oap->oap_rpc_item))
2092                 RETURN(-EBUSY);
2093
2094         if (loi == NULL)
2095                 loi = &lsm->lsm_oinfo[0];
2096
2097         spin_lock(&cli->cl_loi_list_lock);
2098
2099         oap->oap_cmd = cmd;
2100         oap->oap_page_off = off;
2101         oap->oap_count = count;
2102         oap->oap_brw_flags = brw_flags;
2103         oap->oap_async_flags = async_flags;
2104
2105         if (cmd == OBD_BRW_WRITE)
2106                 lop = &loi->loi_write_lop;
2107         else
2108                 lop = &loi->loi_read_lop;
2109
2110         list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2111         if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2112                 oap->oap_oig = oig;
2113                 oig_add_one(oig, &oap->oap_occ);
2114         }
2115
2116         LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
2117
2118         spin_unlock(&cli->cl_loi_list_lock);
2119
2120         RETURN(0);
2121 }
2122
2123 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2124                                  struct loi_oap_pages *lop, int cmd)
2125 {
2126         struct list_head *pos, *tmp;
2127         struct osc_async_page *oap;
2128
2129         list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2130                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2131                 list_del(&oap->oap_pending_item);
2132                 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2133                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2134                 lop_update_pending(cli, lop, cmd, 1);
2135         }
2136         loi_list_maint(cli, loi);
2137 }
2138
2139 static int osc_trigger_group_io(struct obd_export *exp,
2140                                 struct lov_stripe_md *lsm,
2141                                 struct lov_oinfo *loi,
2142                                 struct obd_io_group *oig)
2143 {
2144         struct client_obd *cli = &exp->exp_obd->u.cli;
2145         ENTRY;
2146
2147         if (loi == NULL)
2148                 loi = &lsm->lsm_oinfo[0];
2149
2150         spin_lock(&cli->cl_loi_list_lock);
2151
2152         osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2153         osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2154
2155         osc_check_rpcs(cli);
2156         spin_unlock(&cli->cl_loi_list_lock);
2157
2158         RETURN(0);
2159 }
2160
2161 static int osc_teardown_async_page(struct obd_export *exp,
2162                                    struct lov_stripe_md *lsm,
2163                                    struct lov_oinfo *loi, void *cookie)
2164 {
2165         struct client_obd *cli = &exp->exp_obd->u.cli;
2166         struct loi_oap_pages *lop;
2167         struct osc_async_page *oap;
2168         int rc = 0;
2169         ENTRY;
2170
2171         oap = OAP_FROM_COOKIE(cookie);
2172
2173         if (loi == NULL)
2174                 loi = &lsm->lsm_oinfo[0];
2175
2176         if (oap->oap_cmd == OBD_BRW_WRITE) {
2177                 lop = &loi->loi_write_lop;
2178         } else {
2179                 lop = &loi->loi_read_lop;
2180         }
2181
2182         spin_lock(&cli->cl_loi_list_lock);
2183
2184         if (!list_empty(&oap->oap_rpc_item))
2185                 GOTO(out, rc = -EBUSY);
2186
2187         osc_exit_cache(cli, oap, 0);
2188         osc_wake_cache_waiters(cli);
2189
2190         if (!list_empty(&oap->oap_urgent_item)) {
2191                 list_del_init(&oap->oap_urgent_item);
2192                 oap->oap_async_flags &= ~ASYNC_URGENT;
2193         }
2194         if (!list_empty(&oap->oap_pending_item)) {
2195                 list_del_init(&oap->oap_pending_item);
2196                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2197         }
2198         loi_list_maint(cli, loi);
2199
2200         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2201 out:
2202         spin_unlock(&cli->cl_loi_list_lock);
2203         if (rc == 0)
2204                 OBD_FREE(oap, sizeof(*oap));
2205         RETURN(rc);
2206 }
2207
2208 #ifdef __KERNEL__
2209 /* Note: caller will lock/unlock, and set uptodate on the pages */
2210 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2211 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2212                            struct lov_stripe_md *lsm, obd_count page_count,
2213                            struct brw_page *pga)
2214 {
2215         struct ptlrpc_request *request = NULL;
2216         struct ost_body *body;
2217         struct niobuf_remote *nioptr;
2218         struct obd_ioobj *iooptr;
2219         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2220         int swab;
2221         ENTRY;
2222
2223         /* XXX does not handle 'new' brw protocol */
2224
2225         size[1] = sizeof(struct obd_ioobj);
2226         size[2] = page_count * sizeof(*nioptr);
2227
2228         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2229                                   OST_SAN_READ, 3, size, NULL);
2230         if (!request)
2231                 RETURN(-ENOMEM);
2232
2233         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2234         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2235         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2236                                 sizeof(*nioptr) * page_count);
2237
2238         memcpy(&body->oa, oa, sizeof(body->oa));
2239
2240         obdo_to_ioobj(oa, iooptr);
2241         iooptr->ioo_bufcnt = page_count;
2242
2243         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2244                 LASSERT(PageLocked(pga[mapped].pg));
2245                 LASSERT(mapped == 0 ||
2246                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2247
2248                 nioptr->offset = pga[mapped].disk_offset;
2249                 nioptr->len    = pga[mapped].count;
2250                 nioptr->flags  = pga[mapped].flag;
2251         }
2252
2253         size[1] = page_count * sizeof(*nioptr);
2254         request->rq_replen = lustre_msg_size(2, size);
2255
2256         rc = ptlrpc_queue_wait(request);
2257         if (rc)
2258                 GOTO(out_req, rc);
2259
2260         body = lustre_swab_repbuf(request, 0, sizeof(*body),
2261                                   lustre_swab_ost_body);
2262         if (body == NULL) {
2263                 CERROR("Can't unpack body\n");
2264                 GOTO(out_req, rc = -EPROTO);
2265         }
2266
2267         memcpy(oa, &body->oa, sizeof(*oa));
2268
2269         swab = lustre_msg_swabbed(request->rq_repmsg);
2270         LASSERT_REPSWAB(request, 1);
2271         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2272         if (!nioptr) {
2273                 /* nioptr missing or short */
2274                 GOTO(out_req, rc = -EPROTO);
2275         }
2276
2277         /* actual read */
2278         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2279                 struct page *page = pga[mapped].pg;
2280                 struct buffer_head *bh;
2281                 kdev_t dev;
2282
2283                 if (swab)
2284                         lustre_swab_niobuf_remote (nioptr);
2285
2286                 /* got san device associated */
2287                 LASSERT(exp->exp_obd != NULL);
2288                 dev = exp->exp_obd->u.cli.cl_sandev;
2289
2290                 /* hole */
2291                 if (!nioptr->offset) {
2292                         CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2293                                         page->mapping->host->i_ino,
2294                                         page->index);
2295                         memset(page_address(page), 0, PAGE_SIZE);
2296                         continue;
2297                 }
2298
2299                 if (!page->buffers) {
2300                         create_empty_buffers(page, dev, PAGE_SIZE);
2301                         bh = page->buffers;
2302
2303                         clear_bit(BH_New, &bh->b_state);
2304                         set_bit(BH_Mapped, &bh->b_state);
2305                         bh->b_blocknr = (unsigned long)nioptr->offset;
2306
2307                         clear_bit(BH_Uptodate, &bh->b_state);
2308
2309                         ll_rw_block(READ, 1, &bh);
2310                 } else {
2311                         bh = page->buffers;
2312
2313                         /* if buffer already existed, it must be the
2314                          * one we mapped before, check it */
2315                         LASSERT(!test_bit(BH_New, &bh->b_state));
2316                         LASSERT(test_bit(BH_Mapped, &bh->b_state));
2317                         LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2318
2319                         /* wait it's io completion */
2320                         if (test_bit(BH_Lock, &bh->b_state))
2321                                 wait_on_buffer(bh);
2322
2323                         if (!test_bit(BH_Uptodate, &bh->b_state))
2324                                 ll_rw_block(READ, 1, &bh);
2325                 }
2326
2327
2328                 /* must do syncronous write here */
2329                 wait_on_buffer(bh);
2330                 if (!buffer_uptodate(bh)) {
2331                         /* I/O error */
2332                         rc = -EIO;
2333                         goto out_req;
2334                 }
2335         }
2336
2337 out_req:
2338         ptlrpc_req_finished(request);
2339         RETURN(rc);
2340 }
2341
2342 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2343                             struct lov_stripe_md *lsm, obd_count page_count,
2344                             struct brw_page *pga)
2345 {
2346         struct ptlrpc_request *request = NULL;
2347         struct ost_body *body;
2348         struct niobuf_remote *nioptr;
2349         struct obd_ioobj *iooptr;
2350         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2351         int swab;
2352         ENTRY;
2353
2354         size[1] = sizeof(struct obd_ioobj);
2355         size[2] = page_count * sizeof(*nioptr);
2356
2357         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2358                                   OST_SAN_WRITE, 3, size, NULL);
2359         if (!request)
2360                 RETURN(-ENOMEM);
2361
2362         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2363         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2364         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2365                                 sizeof (*nioptr) * page_count);
2366
2367         memcpy(&body->oa, oa, sizeof(body->oa));
2368
2369         obdo_to_ioobj(oa, iooptr);
2370         iooptr->ioo_bufcnt = page_count;
2371
2372         /* pack request */
2373         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2374                 LASSERT(PageLocked(pga[mapped].pg));
2375                 LASSERT(mapped == 0 ||
2376                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2377
2378                 nioptr->offset = pga[mapped].disk_offset;
2379                 nioptr->len    = pga[mapped].count;
2380                 nioptr->flags  = pga[mapped].flag;
2381         }
2382
2383         size[1] = page_count * sizeof(*nioptr);
2384         request->rq_replen = lustre_msg_size(2, size);
2385
2386         rc = ptlrpc_queue_wait(request);
2387         if (rc)
2388                 GOTO(out_req, rc);
2389
2390         swab = lustre_msg_swabbed (request->rq_repmsg);
2391         LASSERT_REPSWAB (request, 1);
2392         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2393         if (!nioptr) {
2394                 CERROR("absent/short niobuf array\n");
2395                 GOTO(out_req, rc = -EPROTO);
2396         }
2397
2398         /* actual write */
2399         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2400                 struct page *page = pga[mapped].pg;
2401                 struct buffer_head *bh;
2402                 kdev_t dev;
2403
2404                 if (swab)
2405                         lustre_swab_niobuf_remote (nioptr);
2406
2407                 /* got san device associated */
2408                 LASSERT(exp->exp_obd != NULL);
2409                 dev = exp->exp_obd->u.cli.cl_sandev;
2410
2411                 if (!page->buffers) {
2412                         create_empty_buffers(page, dev, PAGE_SIZE);
2413                 } else {
2414                         /* checking */
2415                         LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2416                         LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2417                         LASSERT(page->buffers->b_blocknr ==
2418                                 (unsigned long)nioptr->offset);
2419                 }
2420                 bh = page->buffers;
2421
2422                 LASSERT(bh);
2423
2424                 /* if buffer locked, wait it's io completion */
2425                 if (test_bit(BH_Lock, &bh->b_state))
2426                         wait_on_buffer(bh);
2427
2428                 clear_bit(BH_New, &bh->b_state);
2429                 set_bit(BH_Mapped, &bh->b_state);
2430
2431                 /* override the block nr */
2432                 bh->b_blocknr = (unsigned long)nioptr->offset;
2433
2434                 /* we are about to write it, so set it
2435                  * uptodate/dirty
2436                  * page lock should garentee no race condition here */
2437                 set_bit(BH_Uptodate, &bh->b_state);
2438                 set_bit(BH_Dirty, &bh->b_state);
2439
2440                 ll_rw_block(WRITE, 1, &bh);
2441
2442                 /* must do syncronous write here */
2443                 wait_on_buffer(bh);
2444                 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2445                         /* I/O error */
2446                         rc = -EIO;
2447                         goto out_req;
2448                 }
2449         }
2450
2451 out_req:
2452         ptlrpc_req_finished(request);
2453         RETURN(rc);
2454 }
2455
2456 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2457                       struct lov_stripe_md *lsm, obd_count page_count,
2458                       struct brw_page *pga, struct obd_trans_info *oti)
2459 {
2460         ENTRY;
2461
2462         while (page_count) {
2463                 obd_count pages_per_brw;
2464                 int rc;
2465
2466                 if (page_count > PTLRPC_MAX_BRW_PAGES)
2467                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2468                 else
2469                         pages_per_brw = page_count;
2470
2471                 if (cmd & OBD_BRW_WRITE)
2472                         rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2473                 else
2474                         rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2475
2476                 if (rc != 0)
2477                         RETURN(rc);
2478
2479                 page_count -= pages_per_brw;
2480                 pga += pages_per_brw;
2481         }
2482         RETURN(0);
2483 }
2484 #endif
2485 #endif
2486
2487 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2488 {
2489         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2490
2491         if (lock == NULL) {
2492                 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2493                 return;
2494         }
2495
2496         lock_res_and_lock(lock);
2497 #ifdef __KERNEL__
2498         if (lock->l_ast_data && lock->l_ast_data != data) {
2499                 struct inode *new_inode = data;
2500                 struct inode *old_inode = lock->l_ast_data;
2501                 if (!(old_inode->i_state & I_FREEING))
2502                         LDLM_ERROR(lock, "inconsistent l_ast_data found");
2503                 LASSERTF(old_inode->i_state & I_FREEING,
2504                          "Found existing inode %p/%lu/%u state %lu in lock: "
2505                          "setting data to %p/%lu/%u\n", old_inode,
2506                          old_inode->i_ino, old_inode->i_generation,
2507                          old_inode->i_state,
2508                          new_inode, new_inode->i_ino, new_inode->i_generation);
2509         }
2510 #endif
2511         lock->l_ast_data = data;
2512         unlock_res_and_lock(lock);
2513         LDLM_LOCK_PUT(lock);
2514 }
2515
2516 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2517                              ldlm_iterator_t replace, void *data)
2518 {
2519         struct ldlm_res_id res_id = { .name = {0} };
2520         struct obd_device *obd = class_exp2obd(exp);
2521
2522         res_id.name[0] = lsm->lsm_object_id;
2523         res_id.name[2] = lsm->lsm_object_gr;
2524         ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2525         return 0;
2526 }
2527
2528 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2529                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2530                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2531                        void *data, __u32 lvb_len, void *lvb_swabber,
2532                        struct lustre_handle *lockh)
2533 {
2534         struct obd_device *obd = exp->exp_obd;
2535         struct ldlm_res_id res_id = { .name = {0} };
2536         struct ost_lvb lvb;
2537         struct ldlm_reply *rep;
2538         struct ptlrpc_request *req = NULL;
2539         int rc;
2540         ENTRY;
2541
2542         res_id.name[0] = lsm->lsm_object_id;
2543         res_id.name[2] = lsm->lsm_object_gr;
2544
2545         /* Filesystem lock extents are extended to page boundaries so that
2546          * dealing with the page cache is a little smoother.  */
2547         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2548         policy->l_extent.end |= ~PAGE_MASK;
2549
2550         if (lsm->lsm_oinfo->loi_kms_valid == 0)
2551                 goto no_match;
2552
2553         /* Next, search for already existing extent locks that will cover us */
2554         rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2555                              lockh);
2556         if (rc == 1) {
2557                 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2558                         /* return immediately if no credential held */
2559                         ldlm_lock_decref(lockh, mode);
2560                         RETURN(-EACCES);
2561                 }
2562
2563                 osc_set_data_with_check(lockh, data);
2564                 if (*flags & LDLM_FL_HAS_INTENT) {
2565                         /* I would like to be able to ASSERT here that rss <=
2566                          * kms, but I can't, for reasons which are explained in
2567                          * lov_enqueue() */
2568                 }
2569                 /* We already have a lock, and it's referenced */
2570                 RETURN(ELDLM_OK);
2571         }
2572
2573         /* If we're trying to read, we also search for an existing PW lock.  The
2574          * VFS and page cache already protect us locally, so lots of readers/
2575          * writers can share a single PW lock.
2576          *
2577          * There are problems with conversion deadlocks, so instead of
2578          * converting a read lock to a write lock, we'll just enqueue a new
2579          * one.
2580          *
2581          * At some point we should cancel the read lock instead of making them
2582          * send us a blocking callback, but there are problems with canceling
2583          * locks out from other users right now, too. */
2584
2585         if (mode == LCK_PR) {
2586                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2587                                      policy, LCK_PW, lockh);
2588                 if (rc == 1) {
2589                         if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2590                                 /* return immediately if no credential held */
2591                                 ldlm_lock_decref(lockh, LCK_PW);
2592                                 RETURN(-EACCES);
2593                         }
2594
2595                         /* FIXME: This is not incredibly elegant, but it might
2596                          * be more elegant than adding another parameter to
2597                          * lock_match.  I want a second opinion. */
2598                         ldlm_lock_addref(lockh, LCK_PR);
2599                         ldlm_lock_decref(lockh, LCK_PW);
2600                         osc_set_data_with_check(lockh, data);
2601                         RETURN(ELDLM_OK);
2602                 }
2603         }
2604         if (mode == LCK_PW) {
2605                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2606                                      policy, LCK_PR, lockh);
2607                 if (rc == 1) {
2608                         rc = ldlm_cli_convert(lockh, mode, flags);
2609                         if (!rc) {
2610                                 /* Update readers/writers accounting */
2611                                 ldlm_lock_addref(lockh, LCK_PW);
2612                                 ldlm_lock_decref(lockh, LCK_PR);
2613                                 osc_set_data_with_check(lockh, data);
2614                                 RETURN(ELDLM_OK);
2615                         }
2616                         /* If the conversion failed, we need to drop refcount
2617                            on matched lock before we get new one */
2618                         /* XXX Won't it save us some efforts if we cancel PR
2619                            lock here? We are going to take PW lock anyway and it
2620                            will invalidate PR lock */
2621                         ldlm_lock_decref(lockh, LCK_PR);
2622                         if (rc != EDEADLOCK) {
2623                                 RETURN(rc);
2624                         }
2625                 }
2626         }
2627
2628 no_match:
2629         if (*flags & LDLM_FL_HAS_INTENT) {
2630                 int size[2] = {0, sizeof(struct ldlm_request)};
2631
2632                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
2633                                       LDLM_ENQUEUE, 2, size, NULL);
2634                 if (req == NULL)
2635                         RETURN(-ENOMEM);
2636
2637                 size[0] = sizeof(*rep);
2638                 size[1] = sizeof(lvb);
2639                 req->rq_replen = lustre_msg_size(2, size);
2640         }
2641         rc = ldlm_cli_enqueue(exp, req, obd->obd_namespace, res_id, type,
2642                               policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2643                               &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2644         if (req != NULL) {
2645                 if (rc == ELDLM_LOCK_ABORTED) {
2646                         /* swabbed by ldlm_cli_enqueue() */
2647                         LASSERT_REPSWABBED(req, 0);
2648                         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
2649                         LASSERT(rep != NULL);
2650                         if (rep->lock_policy_res1)
2651                                 rc = rep->lock_policy_res1;
2652                 }
2653                 ptlrpc_req_finished(req);
2654         }
2655
2656         if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2657                 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2658                        lvb.lvb_size, lvb.lvb_blocks);
2659                 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2660                 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2661         }
2662
2663         RETURN(rc);
2664 }
2665
2666 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2667                      __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2668                      int *flags, void *data, struct lustre_handle *lockh)
2669 {
2670         struct ldlm_res_id res_id = { .name = {0} };
2671         struct obd_device *obd = exp->exp_obd;
2672         int rc;
2673         ENTRY;
2674
2675         res_id.name[0] = lsm->lsm_object_id;
2676         res_id.name[2] = lsm->lsm_object_gr;
2677
2678         OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2679
2680         /* Filesystem lock extents are extended to page boundaries so that
2681          * dealing with the page cache is a little smoother */
2682         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2683         policy->l_extent.end |= ~PAGE_MASK;
2684
2685         /* Next, search for already existing extent locks that will cover us */
2686         rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2687                              policy, mode, lockh);
2688         if (rc) {
2689                // if (!(*flags & LDLM_FL_TEST_LOCK))
2690                         osc_set_data_with_check(lockh, data);
2691                 RETURN(rc);
2692         }
2693         /* If we're trying to read, we also search for an existing PW lock.  The
2694          * VFS and page cache already protect us locally, so lots of readers/
2695          * writers can share a single PW lock. */
2696         if (mode == LCK_PR) {
2697                 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2698                                      policy, LCK_PW, lockh);
2699                 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2700                         /* FIXME: This is not incredibly elegant, but it might
2701                          * be more elegant than adding another parameter to
2702                          * lock_match.  I want a second opinion. */
2703                         osc_set_data_with_check(lockh, data);
2704                         ldlm_lock_addref(lockh, LCK_PR);
2705                         ldlm_lock_decref(lockh, LCK_PW);
2706                 }
2707         }
2708         RETURN(rc);
2709 }
2710
2711 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2712                       __u32 mode, struct lustre_handle *lockh)
2713 {
2714         ENTRY;
2715
2716         if (mode == LCK_GROUP)
2717                 ldlm_lock_decref_and_cancel(lockh, mode);
2718         else
2719                 ldlm_lock_decref(lockh, mode);
2720
2721         RETURN(0);
2722 }
2723
2724 static int osc_cancel_unused(struct obd_export *exp,
2725                              struct lov_stripe_md *lsm,
2726                              int flags, void *opaque)
2727 {
2728         struct obd_device *obd = class_exp2obd(exp);
2729         struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2730
2731         if (lsm != NULL) {
2732                 res_id.name[0] = lsm->lsm_object_id;
2733                 res_id.name[2] = lsm->lsm_object_gr;
2734                 resp = &res_id;
2735         }
2736
2737         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2738 }
2739
2740 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2741                       unsigned long max_age)
2742 {
2743         struct obd_statfs *msfs;
2744         struct ptlrpc_request *request;
2745         int rc, size = sizeof(*osfs);
2746         ENTRY;
2747
2748         /* We could possibly pass max_age in the request (as an absolute
2749          * timestamp or a "seconds.usec ago") so the target can avoid doing
2750          * extra calls into the filesystem if that isn't necessary (e.g.
2751          * during mount that would help a bit).  Having relative timestamps
2752          * is not so great if request processing is slow, while absolute
2753          * timestamps are not ideal because they need time synchronization. */
2754         request = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OBD_VERSION,
2755                                   OST_STATFS, 0, NULL, NULL);
2756         if (!request)
2757                 RETURN(-ENOMEM);
2758
2759         request->rq_replen = lustre_msg_size(1, &size);
2760         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2761
2762         rc = ptlrpc_queue_wait(request);
2763         if (rc)
2764                 GOTO(out, rc);
2765
2766         msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2767                                   lustre_swab_obd_statfs);
2768         if (msfs == NULL) {
2769                 CERROR("Can't unpack obd_statfs\n");
2770                 GOTO(out, rc = -EPROTO);
2771         }
2772
2773         memcpy(osfs, msfs, sizeof(*osfs));
2774
2775         EXIT;
2776  out:
2777         ptlrpc_req_finished(request);
2778         return rc;
2779 }
2780
2781 /* Retrieve object striping information.
2782  *
2783  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2784  * the maximum number of OST indices which will fit in the user buffer.
2785  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2786  */
2787 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2788 {
2789         struct lov_user_md lum, *lumk;
2790         int rc, lum_size;
2791         ENTRY;
2792
2793         if (!lsm)
2794                 RETURN(-ENODATA);
2795
2796         rc = copy_from_user(&lum, lump, sizeof(lum));
2797         if (rc)
2798                 RETURN(-EFAULT);
2799
2800         if (lum.lmm_magic != LOV_USER_MAGIC)
2801                 RETURN(-EINVAL);
2802
2803         if (lum.lmm_stripe_count > 0) {
2804                 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2805                 OBD_ALLOC(lumk, lum_size);
2806                 if (!lumk)
2807                         RETURN(-ENOMEM);
2808
2809                 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2810                 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2811         } else {
2812                 lum_size = sizeof(lum);
2813                 lumk = &lum;
2814         }
2815
2816         lumk->lmm_object_id = lsm->lsm_object_id;
2817         lumk->lmm_object_gr = lsm->lsm_object_gr;
2818         lumk->lmm_stripe_count = 1;
2819
2820         if (copy_to_user(lump, lumk, lum_size))
2821                 rc = -EFAULT;
2822
2823         if (lumk != &lum)
2824                 OBD_FREE(lumk, lum_size);
2825
2826         RETURN(rc);
2827 }
2828
2829 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2830                          void *karg, void *uarg)
2831 {
2832         struct obd_device *obd = exp->exp_obd;
2833         struct obd_ioctl_data *data = karg;
2834         int err = 0;
2835         ENTRY;
2836
2837 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2838         MOD_INC_USE_COUNT;
2839 #else
2840         if (!try_module_get(THIS_MODULE)) {
2841                 CERROR("Can't get module. Is it alive?");
2842                 return -EINVAL;
2843         }
2844 #endif
2845         switch (cmd) {
2846         case OBD_IOC_LOV_GET_CONFIG: {
2847                 char *buf;
2848                 struct lov_desc *desc;
2849                 struct obd_uuid uuid;
2850
2851                 buf = NULL;
2852                 len = 0;
2853                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2854                         GOTO(out, err = -EINVAL);
2855
2856                 data = (struct obd_ioctl_data *)buf;
2857
2858                 if (sizeof(*desc) > data->ioc_inllen1) {
2859                         OBD_FREE(buf, len);
2860                         GOTO(out, err = -EINVAL);
2861                 }
2862
2863                 if (data->ioc_inllen2 < sizeof(uuid)) {
2864                         OBD_FREE(buf, len);
2865                         GOTO(out, err = -EINVAL);
2866                 }
2867
2868                 if (data->ioc_inllen3 < sizeof(__u32)) {
2869                         OBD_FREE(buf, len);
2870                         GOTO(out, err = -EINVAL);
2871                 }
2872
2873                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2874                 desc->ld_tgt_count = 1;
2875                 desc->ld_active_tgt_count = 1;
2876                 desc->ld_default_stripe_count = 1;
2877                 desc->ld_default_stripe_size = 0;
2878                 desc->ld_default_stripe_offset = 0;
2879                 desc->ld_pattern = 0;
2880                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2881                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2882                 *((__u32 *)data->ioc_inlbuf3) = 1;
2883
2884                 err = copy_to_user((void *)uarg, buf, len);
2885                 if (err)
2886                         err = -EFAULT;
2887                 obd_ioctl_freedata(buf, len);
2888                 GOTO(out, err);
2889         }
2890         case LL_IOC_LOV_SETSTRIPE:
2891                 err = obd_alloc_memmd(exp, karg);
2892                 if (err > 0)
2893                         err = 0;
2894                 GOTO(out, err);
2895         case LL_IOC_LOV_GETSTRIPE:
2896                 err = osc_getstripe(karg, uarg);
2897                 GOTO(out, err);
2898         case OBD_IOC_CLIENT_RECOVER:
2899                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2900                                             data->ioc_inlbuf1);
2901                 if (err > 0)
2902                         err = 0;
2903                 GOTO(out, err);
2904         case IOC_OSC_SET_ACTIVE:
2905                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2906                                                data->ioc_offset);
2907                 GOTO(out, err);
2908         case IOC_OSC_CTL_RECOVERY:
2909                 err = ptlrpc_import_control_recovery(obd->u.cli.cl_import,
2910                                                      data->ioc_offset);
2911                 GOTO(out, err);
2912         default:
2913                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2914                 GOTO(out, err = -ENOTTY);
2915         }
2916 out:
2917 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2918         MOD_DEC_USE_COUNT;
2919 #else
2920         module_put(THIS_MODULE);
2921 #endif
2922         return err;
2923 }
2924
2925 static int osc_get_info(struct obd_export *exp, __u32 keylen,
2926                         void *key, __u32 *vallen, void *val)
2927 {
2928         ENTRY;
2929         if (!vallen || !val)
2930                 RETURN(-EFAULT);
2931
2932         if (keylen > strlen("lock_to_stripe") &&
2933             strcmp(key, "lock_to_stripe") == 0) {
2934                 __u32 *stripe = val;
2935                 *vallen = sizeof(*stripe);
2936                 *stripe = 0;
2937                 RETURN(0);
2938         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2939                 struct ptlrpc_request *req;
2940                 obd_id *reply;
2941                 char *bufs[1] = {key};
2942                 int rc;
2943                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2944                                       OST_GET_INFO, 1, (int *)&keylen, bufs);
2945                 if (req == NULL)
2946                         RETURN(-ENOMEM);
2947
2948                 req->rq_replen = lustre_msg_size(1, (int *)vallen);
2949                 rc = ptlrpc_queue_wait(req);
2950                 if (rc)
2951                         GOTO(out, rc);
2952
2953                 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2954                                            lustre_swab_ost_last_id);
2955                 if (reply == NULL) {
2956                         CERROR("Can't unpack OST last ID\n");
2957                         GOTO(out, rc = -EPROTO);
2958                 }
2959                 *((obd_id *)val) = *reply;
2960         out:
2961                 ptlrpc_req_finished(req);
2962                 RETURN(rc);
2963         } else if (keylen == 10 && strcmp(key, "client_nid") == 0) {
2964                 struct ptlrpc_connection * conn;
2965                 ptl_nid_t * nid = val;
2966                 ptl_process_id_t id;
2967                 int rc;
2968
2969                 *vallen = sizeof(*nid);
2970                 conn = class_exp2cliimp(exp)->imp_connection;
2971
2972                 if (!conn || !conn->c_peer.peer_ni)
2973                         RETURN(-ENOTCONN);
2974
2975                 rc = PtlGetId(conn->c_peer.peer_ni->pni_ni_h, &id);
2976                 if (rc == PTL_OK)
2977                         *nid = id.nid;
2978
2979                 RETURN(0);
2980         }
2981         RETURN(-EPROTO);
2982 }
2983
2984 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2985                         void *key, obd_count vallen, void *val)
2986 {
2987         struct obd_device  *obd = exp->exp_obd;
2988         struct obd_import *imp = class_exp2cliimp(exp);
2989         struct llog_ctxt *ctxt;
2990         int rc = 0;
2991         ENTRY;
2992
2993         if (keylen == strlen("unlinked") &&
2994             memcmp(key, "unlinked", keylen) == 0) {
2995                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2996                 spin_lock(&oscc->oscc_lock);
2997                 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2998                 spin_unlock(&oscc->oscc_lock);
2999                 RETURN(0);
3000         }
3001
3002         if (keylen == strlen("unrecovery") &&
3003             memcmp(key, "unrecovery", keylen) == 0) {
3004                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3005                 spin_lock(&oscc->oscc_lock);
3006                 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3007                 spin_unlock(&oscc->oscc_lock);
3008                 RETURN(0);
3009         }
3010
3011         if (keylen == strlen("initial_recov") &&
3012             memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
3013                 struct obd_import *imp = class_exp2cliimp(exp);
3014                 if (vallen != sizeof(int))
3015                         RETURN(-EINVAL);
3016                 imp->imp_initial_recov = *(int *)val;
3017                 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
3018                        exp->exp_obd->obd_name,
3019                        imp->imp_initial_recov);
3020                 RETURN(0);
3021         }
3022
3023         if (keylen == strlen("async") &&
3024             memcmp(key, "async", keylen) == 0) {
3025                 struct client_obd *cl = &obd->u.cli;
3026                 if (vallen != sizeof(int))
3027                         RETURN(-EINVAL);
3028                 cl->cl_async = *(int *)val;
3029                 CDEBUG(D_HA, "%s: set async = %d\n",
3030                        obd->obd_name, cl->cl_async);
3031                 RETURN(0);
3032         }
3033
3034         if (keylen == 5 && strcmp(key, "audit") == 0) {
3035                 struct ptlrpc_request *req;
3036                 char *bufs[2] = {key, val};
3037                 int size[2] = {keylen, vallen};
3038
3039                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3040                                       OST_SET_INFO, 2, size, bufs);
3041                 if (req == NULL)
3042                         RETURN(-ENOMEM);
3043
3044                 req->rq_replen = lustre_msg_size(0, size);
3045                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
3046                                    lustre_swab_audit_attr);
3047                 rc = ptlrpc_queue_wait(req);
3048
3049                 ptlrpc_req_finished(req);
3050                 RETURN(rc);
3051         }
3052
3053         if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
3054                 struct ptlrpc_request *req;
3055                 char *bufs[2] = {key, val};
3056                 int size[2] = {keylen, vallen};
3057
3058                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3059                                       OST_SET_INFO, 2, size, bufs);
3060                 if (req == NULL)
3061                         RETURN(-ENOMEM);
3062
3063                 req->rq_replen = lustre_msg_size(0, size);
3064                 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
3065                                    lustre_swab_obdo);
3066                 rc = ptlrpc_queue_wait(req);
3067
3068                 ptlrpc_req_finished(req);
3069                 RETURN(rc);
3070         }
3071
3072         if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
3073                 struct ptlrpc_request *req;
3074                 char *bufs[2] = {key, val};
3075                 int size[2] = {keylen, vallen};
3076
3077                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3078                                       OST_SET_INFO, 2, size, bufs);
3079                 if (req == NULL)
3080                         RETURN(-ENOMEM);
3081
3082                 req->rq_replen = lustre_msg_size(0, size);
3083                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
3084                                    lustre_swab_audit_msg);
3085                 rc = ptlrpc_queue_wait(req);
3086
3087                 ptlrpc_req_finished(req);
3088                 RETURN(rc);
3089         }
3090
3091         if (keylen == strlen("sec") && memcmp(key, "sec", keylen) == 0) {
3092                 struct client_obd *cli = &exp->exp_obd->u.cli;
3093
3094                 cli->cl_sec_flavor = ptlrpcs_name2flavor(val);
3095                 if (cli->cl_sec_flavor == PTLRPCS_FLVR_INVALID) {
3096                         CERROR("unrecognized security flavor %s\n", (char*) val);
3097                         RETURN(-EINVAL);
3098                 }
3099
3100                 RETURN(0);
3101         }
3102
3103         if (keylen == strlen("sec_flags") &&
3104             memcmp(key, "sec_flags", keylen) == 0) {
3105                 struct client_obd *cli = &exp->exp_obd->u.cli;
3106
3107                 cli->cl_sec_flags = *((unsigned long *) val);
3108                 RETURN(0);
3109         }
3110
3111         if (keylen == strlen("flush_cred") &&
3112             memcmp(key, "flush_cred", keylen) == 0) {
3113                 struct client_obd *cli = &exp->exp_obd->u.cli;
3114
3115                 if (cli->cl_import)
3116                         ptlrpcs_import_flush_current_creds(cli->cl_import);
3117                 RETURN(0);
3118         }
3119         if (keylen == strlen("crypto_cb") &&
3120             memcmp(key, "crypto_cb", keylen) == 0) {
3121                 LASSERT(vallen == sizeof(crypt_cb_t));
3122                 osc_crypt_cb = (crypt_cb_t)val;
3123                 RETURN(0);
3124         }
3125
3126         if (keylen == 8 && memcmp(key, "capa_key", 8) == 0) {
3127                 struct ptlrpc_request *req;
3128                 char *bufs[2] = {key, val};
3129                 unsigned long irqflags;
3130                 int rc, size[2] = {keylen, vallen};
3131
3132                 LASSERT(vallen == sizeof(struct lustre_capa_key));
3133
3134                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3135                                       OST_SET_INFO, 2, size, bufs);
3136                 if (req == NULL)
3137                         RETURN(-ENOMEM);
3138
3139                 spin_lock_irqsave (&req->rq_lock, irqflags);
3140                 req->rq_replay = 1;
3141                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
3142
3143                 req->rq_replen = lustre_msg_size(0, NULL);
3144                 rc = ptlrpc_queue_wait(req);
3145                 ptlrpc_req_finished(req);
3146                 RETURN(rc);
3147         }
3148
3149         if (keylen == strlen("setext") &&
3150             memcmp(key, "setext", keylen) == 0) {
3151                 struct client_obd *cli = &exp->exp_obd->u.cli;
3152                 struct osc_creator *oscc = &cli->cl_oscc;
3153                 struct fid_extent *ext = val;
3154
3155                 oscc->oscc_next_id = (obd_id)ext->fe_start;
3156                 RETURN(0);
3157         }
3158
3159         if (keylen < strlen("mds_conn") ||
3160             memcmp(key, "mds_conn", keylen) != 0)
3161                 RETURN(-EINVAL);
3162
3163         ctxt = llog_get_context(&exp->exp_obd->obd_llogs,
3164                                 LLOG_UNLINK_ORIG_CTXT);
3165         if (ctxt) {
3166                 if (rc == 0)
3167                         rc = llog_initiator_connect(ctxt);
3168                 else
3169                         CERROR("cannot establish the connect for "
3170                                "ctxt %p: %d\n", ctxt, rc);
3171         }
3172
3173         imp->imp_server_timeout = 1;
3174         CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
3175         imp->imp_pingable = 1;
3176
3177         RETURN(rc);
3178 }
3179
3180
3181 static struct llog_operations osc_size_repl_logops = {
3182         lop_cancel: llog_obd_repl_cancel
3183 };
3184
3185 static struct llog_operations osc_unlink_orig_logops;
3186
3187 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
3188                          struct obd_device *tgt, int count,
3189                          struct llog_catid *catid)
3190 {
3191         int rc;
3192         ENTRY;
3193
3194         osc_unlink_orig_logops = llog_lvfs_ops;
3195         osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
3196         osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
3197         osc_unlink_orig_logops.lop_add = llog_catalog_add;
3198         osc_unlink_orig_logops.lop_connect = llog_origin_connect;
3199
3200         rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
3201                             &catid->lci_logid, &osc_unlink_orig_logops);
3202         if (rc)
3203                 RETURN(rc);
3204
3205         rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
3206                             &osc_size_repl_logops);
3207         RETURN(rc);
3208 }
3209
3210 static int osc_llog_finish(struct obd_device *obd,
3211                            struct obd_llogs *llogs, int count)
3212 {
3213         int rc;
3214         ENTRY;
3215
3216         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
3217         if (rc)
3218                 RETURN(rc);
3219
3220         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
3221         RETURN(rc);
3222 }
3223
3224 static int osc_connect(struct lustre_handle *exph,
3225                        struct obd_device *obd, struct obd_uuid *cluuid,
3226                        struct obd_connect_data *data,
3227                        unsigned long connect_flags)
3228 {
3229         int rc;
3230         ENTRY;
3231         rc = client_connect_import(exph, obd, cluuid, data, connect_flags);
3232         RETURN(rc);
3233 }
3234
3235 static int osc_disconnect(struct obd_export *exp, unsigned long flags)
3236 {
3237         struct obd_device *obd = class_exp2obd(exp);
3238         struct llog_ctxt *ctxt;
3239         int rc;
3240         ENTRY;
3241
3242         ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
3243         if (obd->u.cli.cl_conn_count == 1)
3244                 /* flush any remaining cancel messages out to the target */
3245                 llog_sync(ctxt, exp);
3246
3247         rc = client_disconnect_export(exp, flags);
3248         RETURN(rc);
3249 }
3250
3251 static int osc_import_event(struct obd_device *obd,
3252                             struct obd_import *imp,
3253                             enum obd_import_event event)
3254 {
3255         struct client_obd *cli;
3256         int rc = 0;
3257         ENTRY;
3258
3259         LASSERT(imp->imp_obd == obd);
3260
3261         switch (event) {
3262         case IMP_EVENT_DISCON: {
3263                 /* Only do this on the MDS OSC's */
3264                 if (imp->imp_server_timeout) {
3265                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3266
3267                         spin_lock(&oscc->oscc_lock);
3268                         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3269                         spin_unlock(&oscc->oscc_lock);
3270                 }
3271                 break;
3272         }
3273         case IMP_EVENT_INACTIVE: {
3274                 if (obd->obd_observer)
3275                         rc = obd_notify(obd->obd_observer, obd, 0, 0);
3276                 break;
3277         }
3278         case IMP_EVENT_INVALIDATE: {
3279                 struct ldlm_namespace *ns = obd->obd_namespace;
3280
3281                 /* Reset grants */
3282                 cli = &obd->u.cli;
3283                 spin_lock(&cli->cl_loi_list_lock);
3284                 cli->cl_avail_grant = 0;
3285                 cli->cl_lost_grant = 0;
3286                 /* all pages go to failing rpcs due to the invalid import */
3287                 osc_check_rpcs(cli);
3288                 spin_unlock(&cli->cl_loi_list_lock);
3289
3290                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3291
3292                 break;
3293         }
3294         case IMP_EVENT_ACTIVE: {
3295                 /* Only do this on the MDS OSC's */
3296                 if (imp->imp_server_timeout) {
3297                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3298
3299                         spin_lock(&oscc->oscc_lock);
3300                         oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3301                         spin_unlock(&oscc->oscc_lock);
3302                 }
3303
3304                 if (obd->obd_observer)
3305                         rc = obd_notify(obd->obd_observer, obd, 1, 0);
3306                 break;
3307         }
3308         default:
3309                 CERROR("Unknown import event %d\n", event);
3310                 LBUG();
3311         }
3312         RETURN(rc);
3313 }
3314
3315 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
3316 {
3317         struct lprocfs_static_vars lvars;
3318         int rc;
3319         ENTRY;
3320
3321         lprocfs_init_vars(osc,&lvars);
3322         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3323         if (rc < 0)
3324                 RETURN(rc);
3325
3326         rc = lproc_osc_attach_seqstat(dev);
3327         if (rc < 0) {
3328                 lprocfs_obd_detach(dev);
3329                 RETURN(rc);
3330         }
3331
3332         ptlrpc_lprocfs_register_obd(dev);
3333         RETURN(0);
3334 }
3335
3336 static int osc_detach(struct obd_device *dev)
3337 {
3338         ptlrpc_lprocfs_unregister_obd(dev);
3339         return lprocfs_obd_detach(dev);
3340 }
3341
3342 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3343 {
3344         int rc;
3345         ENTRY;
3346         rc = ptlrpcd_addref();
3347         if (rc)
3348                 RETURN(rc);
3349
3350         rc = client_obd_setup(obd, len, buf);
3351         if (rc)
3352                 ptlrpcd_decref();
3353         else
3354                 oscc_init(obd);
3355
3356         RETURN(rc);
3357 }
3358
3359 static int osc_cleanup(struct obd_device *obd, int flags)
3360 {
3361         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3362         int rc;
3363
3364         rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3365                                     LDLM_FL_CONFIG_CHANGE, NULL);
3366         if (rc)
3367                 RETURN(rc);
3368
3369         spin_lock(&oscc->oscc_lock);
3370         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3371         oscc->oscc_flags |= OSCC_FLAG_EXITING;
3372         spin_unlock(&oscc->oscc_lock);
3373
3374         rc = client_obd_cleanup(obd, flags);
3375         ptlrpcd_decref();
3376         RETURN(rc);
3377 }
3378
3379
3380 struct obd_ops osc_obd_ops = {
3381         .o_owner                = THIS_MODULE,
3382         .o_attach               = osc_attach,
3383         .o_detach               = osc_detach,
3384         .o_setup                = osc_setup,
3385         .o_cleanup              = osc_cleanup,
3386         .o_add_conn             = client_import_add_conn,
3387         .o_del_conn             = client_import_del_conn,
3388         .o_connect              = osc_connect,
3389         .o_disconnect           = osc_disconnect,
3390         .o_statfs               = osc_statfs,
3391         .o_packmd               = osc_packmd,
3392         .o_unpackmd             = osc_unpackmd,
3393         .o_create               = osc_create,
3394         .o_destroy              = osc_destroy,
3395         .o_getattr              = osc_getattr,
3396         .o_getattr_async        = osc_getattr_async,
3397         .o_setattr              = osc_setattr,
3398         .o_brw                  = osc_brw,
3399         .o_brw_async            = osc_brw_async,
3400         .o_prep_async_page      = osc_prep_async_page,
3401         .o_queue_async_io       = osc_queue_async_io,
3402         .o_set_async_flags      = osc_set_async_flags,
3403         .o_queue_group_io       = osc_queue_group_io,
3404         .o_trigger_group_io     = osc_trigger_group_io,
3405         .o_teardown_async_page  = osc_teardown_async_page,
3406         .o_punch                = osc_punch,
3407         .o_sync                 = osc_sync,
3408         .o_enqueue              = osc_enqueue,
3409         .o_match                = osc_match,
3410         .o_change_cbdata        = osc_change_cbdata,
3411         .o_cancel               = osc_cancel,
3412         .o_cancel_unused        = osc_cancel_unused,
3413         .o_iocontrol            = osc_iocontrol,
3414         .o_get_info             = osc_get_info,
3415         .o_set_info             = osc_set_info,
3416         .o_import_event         = osc_import_event,
3417         .o_llog_init            = osc_llog_init,
3418         .o_llog_finish          = osc_llog_finish,
3419 };
3420
3421 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3422 struct obd_ops sanosc_obd_ops = {
3423         .o_owner                = THIS_MODULE,
3424         .o_attach               = osc_attach,
3425         .o_detach               = osc_detach,
3426         .o_cleanup              = client_obd_cleanup,
3427         .o_add_conn             = client_import_add_conn,
3428         .o_del_conn             = client_import_del_conn,
3429         .o_connect              = osc_connect,
3430         .o_disconnect           = client_disconnect_export,
3431         .o_statfs               = osc_statfs,
3432         .o_packmd               = osc_packmd,
3433         .o_unpackmd             = osc_unpackmd,
3434         .o_create               = osc_real_create,
3435         .o_destroy              = osc_destroy,
3436         .o_getattr              = osc_getattr,
3437         .o_getattr_async        = osc_getattr_async,
3438         .o_setattr              = osc_setattr,
3439         .o_setup                = client_sanobd_setup,
3440         .o_brw                  = sanosc_brw,
3441         .o_punch                = osc_punch,
3442         .o_sync                 = osc_sync,
3443         .o_enqueue              = osc_enqueue,
3444         .o_match                = osc_match,
3445         .o_change_cbdata        = osc_change_cbdata,
3446         .o_cancel               = osc_cancel,
3447         .o_cancel_unused        = osc_cancel_unused,
3448         .o_iocontrol            = osc_iocontrol,
3449         .o_import_event         = osc_import_event,
3450         .o_llog_init            = osc_llog_init,
3451         .o_llog_finish          = osc_llog_finish,
3452 };
3453 #endif
3454
3455 int __init osc_init(void)
3456 {
3457         struct lprocfs_static_vars lvars;
3458 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3459         struct lprocfs_static_vars sanlvars;
3460 #endif
3461         int rc;
3462         ENTRY;
3463
3464         lprocfs_init_vars(osc, &lvars);
3465 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3466         lprocfs_init_vars(osc, &sanlvars);
3467 #endif
3468
3469         rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3470                                  OBD_OSC_DEVICENAME);
3471         if (rc)
3472                 RETURN(rc);
3473
3474 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3475         rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3476                                  OBD_SANOSC_DEVICENAME);
3477         if (rc)
3478                 class_unregister_type(OBD_OSC_DEVICENAME);
3479 #endif
3480
3481         RETURN(rc);
3482 }
3483
3484 #ifdef __KERNEL__
3485 static void /*__exit*/ osc_exit(void)
3486 {
3487 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3488         class_unregister_type(OBD_SANOSC_DEVICENAME);
3489 #endif
3490         class_unregister_type(OBD_OSC_DEVICENAME);
3491 }
3492
3493 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3494 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3495 MODULE_LICENSE("GPL");
3496
3497 module_init(osc_init);
3498 module_exit(osc_exit);
3499 #endif