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         LASSERT(lop != LP_POISON);
1479
1480         /* first we find the pages we're allowed to work with */
1481         list_for_each_safe(pos, tmp, &lop->lop_pending) {
1482                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1483                 ops = oap->oap_caller_ops;
1484
1485                 LASSERT(oap->oap_magic == OAP_MAGIC);
1486
1487                 /* in llite being 'ready' equates to the page being locked
1488                  * until completion unlocks it.  commit_write submits a page
1489                  * as not ready because its unlock will happen unconditionally
1490                  * as the call returns.  if we race with commit_write giving
1491                  * us that page we dont' want to create a hole in the page
1492                  * stream, so we stop and leave the rpc to be fired by
1493                  * another dirtier or kupdated interval (the not ready page
1494                  * will still be on the dirty list).  we could call in
1495                  * at the end of ll_file_write to process the queue again. */
1496                 if (!(oap->oap_async_flags & ASYNC_READY)) {
1497                         int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1498                         if (rc < 0)
1499                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
1500                                                 "instead of ready\n", oap,
1501                                                 oap->oap_page, rc);
1502                         switch (rc) {
1503                         case -EAGAIN:
1504                                 /* llite is telling us that the page is still
1505                                  * in commit_write and that we should try
1506                                  * and put it in an rpc again later.  we
1507                                  * break out of the loop so we don't create
1508                                  * a hole in the sequence of pages in the rpc
1509                                  * stream.*/
1510                                 pos = NULL;
1511                                 break;
1512                         case -EINTR:
1513                                 /* the io isn't needed.. tell the checks
1514                                  * below to complete the rpc with EINTR */
1515                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1516                                 oap->oap_count = -EINTR;
1517                                 break;
1518                         case 0:
1519                                 oap->oap_async_flags |= ASYNC_READY;
1520                                 break;
1521                         default:
1522                                 LASSERTF(0, "oap %p page %p returned %d "
1523                                             "from make_ready\n", oap,
1524                                             oap->oap_page, rc);
1525                                 break;
1526                         }
1527                 }
1528                 if (pos == NULL)
1529                         break;
1530                 /*
1531                  * Page submitted for IO has to be locked. Either by
1532                  * ->ap_make_ready() or by higher layers.
1533                  *
1534                  * XXX nikita: this assertion should be adjusted when lustre
1535                  * starts using PG_writeback for pages being written out.
1536                  */
1537                 LASSERT(PageLocked(oap->oap_page));
1538
1539                 /* take the page out of our book-keeping */
1540                 list_del_init(&oap->oap_pending_item);
1541                 lop_update_pending(cli, lop, cmd, -1);
1542                 list_del_init(&oap->oap_urgent_item);
1543
1544                 /* ask the caller for the size of the io as the rpc leaves. */
1545                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1546                         oap->oap_count =
1547                                 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1548                 if (oap->oap_count <= 0) {
1549                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1550                                oap->oap_count);
1551                         osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1552                         continue;
1553                 }
1554
1555                 /* now put the page back in our accounting */
1556                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1557                 if (++page_count >= cli->cl_max_pages_per_rpc)
1558                         break;
1559         }
1560
1561         osc_wake_cache_waiters(cli);
1562
1563         if (page_count == 0)
1564                 RETURN(0);
1565
1566         loi_list_maint(cli, loi);
1567         spin_unlock(&cli->cl_loi_list_lock);
1568
1569         request = osc_build_req(cli, &rpc_list, page_count, cmd);
1570         if (IS_ERR(request)) {
1571                 /* this should happen rarely and is pretty bad, it makes the
1572                  * pending list not follow the dirty order */
1573                 spin_lock(&cli->cl_loi_list_lock);
1574                 list_for_each_safe(pos, tmp, &rpc_list) {
1575                         oap = list_entry(pos, struct osc_async_page,
1576                                          oap_rpc_item);
1577                         list_del_init(&oap->oap_rpc_item);
1578
1579                         /* queued sync pages can be torn down while the pages
1580                          * were between the pending list and the rpc */
1581                         if (oap->oap_interrupted) {
1582                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1583                                 osc_ap_completion(cli, NULL, oap, 0,
1584                                                   oap->oap_count);
1585                                 continue;
1586                         }
1587
1588                         /* put the page back in the loi/lop lists */
1589                         list_add_tail(&oap->oap_pending_item,
1590                                       &lop->lop_pending);
1591                         lop_update_pending(cli, lop, cmd, 1);
1592                         if (oap->oap_async_flags & ASYNC_URGENT)
1593                                 list_add(&oap->oap_urgent_item,
1594                                          &lop->lop_urgent);
1595                 }
1596                 loi_list_maint(cli, loi);
1597                 RETURN(PTR_ERR(request));
1598         }
1599
1600         LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1601         aa = (struct osc_brw_async_args *)&request->rq_async_args;
1602         INIT_LIST_HEAD(&aa->aa_oaps);
1603         list_splice(&rpc_list, &aa->aa_oaps);
1604         INIT_LIST_HEAD(&rpc_list);
1605
1606 #ifdef __KERNEL__
1607         if (cmd == OBD_BRW_READ) {
1608                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1609                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1610         } else {
1611                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1612                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1613                                  cli->cl_w_in_flight);
1614         }
1615 #endif
1616
1617         spin_lock(&cli->cl_loi_list_lock);
1618
1619         /* collect write gaps and sum of them */
1620         if (cmd == OBD_BRW_WRITE && cli->cl_w_in_flight == 0) {
1621                 struct timeval now;
1622                 long diff;
1623
1624                 do_gettimeofday(&now);
1625
1626                 if (cli->cl_last_write_time.tv_sec) {
1627                         diff = timeval_sub(&now, &cli->cl_last_write_time);
1628                         if (diff < CLI_ODD_WRITE_GAP) {
1629                                 cli->cl_write_gap_sum += diff;
1630                                 cli->cl_write_gaps++;
1631                         }
1632                 }
1633         }
1634
1635         if (cmd == OBD_BRW_READ) {
1636                 cli->cl_r_in_flight++;
1637                 cli->cl_read_num++;
1638         } else {
1639                 cli->cl_w_in_flight++;
1640                 cli->cl_write_num++;
1641         }
1642
1643         /* queued sync pages can be torn down while the pages
1644          * were between the pending list and the rpc */
1645         list_for_each(pos, &aa->aa_oaps) {
1646                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1647                 if (oap->oap_interrupted) {
1648                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1649                                oap, request);
1650                         ptlrpc_mark_interrupted(request);
1651                         break;
1652                 }
1653         }
1654
1655         CDEBUG(D_INODE, "req %p: %d pages, aa %p.  now %dr/%dw in flight\n",
1656                         request, page_count, aa, cli->cl_r_in_flight,
1657                         cli->cl_w_in_flight);
1658
1659         oap->oap_request = ptlrpc_request_addref(request);
1660         request->rq_interpret_reply = brw_interpret_oap;
1661
1662         ptlrpcd_add_req(request);
1663         RETURN(1);
1664 }
1665
1666 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1667                          int cmd)
1668 {
1669         int optimal;
1670         ENTRY;
1671
1672         if (lop->lop_num_pending == 0)
1673                 RETURN(0);
1674
1675         /* if we have an invalid import we want to drain the queued pages
1676          * by forcing them through rpcs that immediately fail and complete
1677          * the pages.  recovery relies on this to empty the queued pages
1678          * before canceling the locks and evicting down the llite pages */
1679         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1680                 RETURN(1);
1681
1682         /* stream rpcs in queue order as long as as there is an urgent page
1683          * queued.  this is our cheap solution for good batching in the case
1684          * where writepage marks some random page in the middle of the file as
1685          * urgent because of, say, memory pressure */
1686         if (!list_empty(&lop->lop_urgent))
1687                 RETURN(1);
1688
1689         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1690         optimal = cli->cl_max_pages_per_rpc;
1691         if (cmd == OBD_BRW_WRITE) {
1692                 /* trigger a write rpc stream as long as there are dirtiers
1693                  * waiting for space.  as they're waiting, they're not going to
1694                  * create more pages to coallesce with what's waiting.. */
1695                 if (!list_empty(&cli->cl_cache_waiters))
1696                         RETURN(1);
1697
1698                 /* *2 to avoid triggering rpcs that would want to include pages
1699                  * that are being queued but which can't be made ready until
1700                  * the queuer finishes with the page. this is a wart for
1701                  * llite::commit_write() */
1702                 optimal += 16;
1703         }
1704         if (lop->lop_num_pending >= optimal)
1705                 RETURN(1);
1706
1707         RETURN(0);
1708 }
1709
1710 static void on_list(struct list_head *item, struct list_head *list,
1711                     int should_be_on)
1712 {
1713         if (list_empty(item) && should_be_on)
1714                 list_add_tail(item, list);
1715         else if (!list_empty(item) && !should_be_on)
1716                 list_del_init(item);
1717 }
1718
1719 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1720  * can find pages to build into rpcs quickly */
1721 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1722 {
1723         on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1724                 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1725                 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1726
1727         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1728                 loi->loi_write_lop.lop_num_pending);
1729
1730         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1731                 loi->loi_read_lop.lop_num_pending);
1732 }
1733
1734 #define LOI_DEBUG(LOI, STR, args...)                                     \
1735         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
1736                !list_empty(&(LOI)->loi_cli_item),                        \
1737                (LOI)->loi_write_lop.lop_num_pending,                     \
1738                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
1739                (LOI)->loi_read_lop.lop_num_pending,                      \
1740                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
1741                args)                                                     \
1742
1743 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1744 {
1745         ENTRY;
1746         /* first return all objects which we already know to have
1747          * pages ready to be stuffed into rpcs */
1748         if (!list_empty(&cli->cl_loi_ready_list))
1749                 RETURN(list_entry(cli->cl_loi_ready_list.next,
1750                                   struct lov_oinfo, loi_cli_item));
1751
1752         /* then if we have cache waiters, return all objects with queued
1753          * writes.  This is especially important when many small files
1754          * have filled up the cache and not been fired into rpcs because
1755          * they don't pass the nr_pending/object threshhold */
1756         if (!list_empty(&cli->cl_cache_waiters) &&
1757             !list_empty(&cli->cl_loi_write_list))
1758                 RETURN(list_entry(cli->cl_loi_write_list.next,
1759                                   struct lov_oinfo, loi_write_item));
1760
1761         /* then return all queued objects when we have an invalid import
1762          * so that they get flushed */
1763         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1764                 if (!list_empty(&cli->cl_loi_write_list))
1765                         RETURN(list_entry(cli->cl_loi_write_list.next,
1766                                           struct lov_oinfo, loi_write_item));
1767                 if (!list_empty(&cli->cl_loi_read_list))
1768                         RETURN(list_entry(cli->cl_loi_read_list.next,
1769                                           struct lov_oinfo, loi_read_item));
1770         }
1771         RETURN(NULL);
1772 }
1773
1774 /* called with the loi list lock held */
1775 static void osc_check_rpcs(struct client_obd *cli)
1776 {
1777         struct lov_oinfo *loi;
1778         int rc = 0, race_counter = 0;
1779         ENTRY;
1780
1781         while ((loi = osc_next_loi(cli)) != NULL) {
1782                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
1783                 LASSERT(loi->loi_ost_idx != LL_POISON);
1784
1785                 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
1786                         break;
1787
1788                 /* attempt some read/write balancing by alternating between
1789                  * reads and writes in an object.  The makes_rpc checks here
1790                  * would be redundant if we were getting read/write work items
1791                  * instead of objects.  we don't want send_oap_rpc to drain a
1792                  * partial read pending queue when we're given this object to
1793                  * do io on writes while there are cache waiters */
1794                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1795                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1796                                               &loi->loi_write_lop);
1797                         if (rc < 0)
1798                                 break;
1799                         if (rc > 0)
1800                                 race_counter = 0;
1801                         else
1802                                 race_counter++;
1803                 }
1804                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1805                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1806                                               &loi->loi_read_lop);
1807                         if (rc < 0)
1808                                 break;
1809                         if (rc > 0)
1810                                 race_counter = 0;
1811                         else
1812                                 race_counter++;
1813                 }
1814
1815                 /* attempt some inter-object balancing by issueing rpcs
1816                  * for each object in turn */
1817                 if (!list_empty(&loi->loi_cli_item))
1818                         list_del_init(&loi->loi_cli_item);
1819                 if (!list_empty(&loi->loi_write_item))
1820                         list_del_init(&loi->loi_write_item);
1821                 if (!list_empty(&loi->loi_read_item))
1822                         list_del_init(&loi->loi_read_item);
1823
1824                 loi_list_maint(cli, loi);
1825
1826                 /* send_oap_rpc fails with 0 when make_ready tells it to
1827                  * back off.  llite's make_ready does this when it tries
1828                  * to lock a page queued for write that is already locked.
1829                  * we want to try sending rpcs from many objects, but we
1830                  * don't want to spin failing with 0.  */
1831                 if (race_counter == 10)
1832                         break;
1833         }
1834         EXIT;
1835 }
1836
1837 /* we're trying to queue a page in the osc so we're subject to the
1838  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1839  * If the osc's queued pages are already at that limit, then we want to sleep
1840  * until there is space in the osc's queue for us.  We also may be waiting for
1841  * write credits from the OST if there are RPCs in flight that may return some
1842  * before we fall back to sync writes.
1843  *
1844  * We need this know our allocation was granted in the presence of signals */
1845 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1846 {
1847         int rc;
1848         ENTRY;
1849         spin_lock(&cli->cl_loi_list_lock);
1850         rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
1851         spin_unlock(&cli->cl_loi_list_lock);
1852         RETURN(rc);
1853 };
1854
1855 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1856  * grant or cache space. */
1857 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1858                            struct osc_async_page *oap)
1859 {
1860         struct osc_cache_waiter ocw;
1861         struct l_wait_info lwi = { 0 };
1862         struct timeval start, stop;
1863
1864         CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1865                cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1866                cli->cl_avail_grant);
1867
1868         if (cli->cl_dirty_max < PAGE_SIZE)
1869                 return(-EDQUOT);
1870
1871         /* Hopefully normal case - cache space and write credits available */
1872         if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1873             cli->cl_avail_grant >= PAGE_SIZE) {
1874                 /* account for ourselves */
1875                 osc_consume_write_grant(cli, oap);
1876                 return(0);
1877         }
1878
1879         /* Make sure that there are write rpcs in flight to wait for.  This
1880          * is a little silly as this object may not have any pending but
1881          * other objects sure might. */
1882         if (cli->cl_w_in_flight) {
1883                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1884                 init_waitqueue_head(&ocw.ocw_waitq);
1885                 ocw.ocw_oap = oap;
1886                 ocw.ocw_rc = 0;
1887
1888                 loi_list_maint(cli, loi);
1889                 osc_check_rpcs(cli);
1890                 spin_unlock(&cli->cl_loi_list_lock);
1891
1892                 CDEBUG(0, "sleeping for cache space\n");
1893                 do_gettimeofday(&start);
1894                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1895                 do_gettimeofday(&stop);
1896                 spin_lock(&cli->cl_loi_list_lock);
1897                 lprocfs_stime_record(&cli->cl_enter_stime, &stop, &start);
1898                 if (!list_empty(&ocw.ocw_entry)) {
1899                         list_del(&ocw.ocw_entry);
1900                         RETURN(-EINTR);
1901                 }
1902                 RETURN(ocw.ocw_rc);
1903         }
1904
1905         RETURN(-EDQUOT);
1906 }
1907
1908 /* the companion to enter_cache, called when an oap is no longer part of the
1909  * dirty accounting.. so writeback completes or truncate happens before writing
1910  * starts.  must be called with the loi lock held. */
1911 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1912                            int sent)
1913 {
1914         ENTRY;
1915
1916         if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1917                 EXIT;
1918                 return;
1919         }
1920
1921         oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1922         cli->cl_dirty -= PAGE_SIZE;
1923         if (!sent) {
1924                 cli->cl_lost_grant += PAGE_SIZE;
1925                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1926                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1927         }
1928
1929         EXIT;
1930 }
1931
1932 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1933                         struct lov_oinfo *loi, struct page *page,
1934                         obd_off offset, struct obd_async_page_ops *ops,
1935                         void *data, void **res)
1936 {
1937         struct osc_async_page *oap;
1938         ENTRY;
1939
1940         OBD_ALLOC(oap, sizeof(*oap));
1941         if (oap == NULL)
1942                 return -ENOMEM;
1943
1944         oap->oap_magic = OAP_MAGIC;
1945         oap->oap_cli = &exp->exp_obd->u.cli;
1946         oap->oap_loi = loi;
1947
1948         oap->oap_caller_ops = ops;
1949         oap->oap_caller_data = data;
1950
1951         oap->oap_page = page;
1952         oap->oap_obj_off = offset;
1953
1954         INIT_LIST_HEAD(&oap->oap_pending_item);
1955         INIT_LIST_HEAD(&oap->oap_urgent_item);
1956         INIT_LIST_HEAD(&oap->oap_rpc_item);
1957
1958         oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1959
1960         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1961         *res = oap;
1962         RETURN(0);
1963 }
1964
1965 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1966                               struct lov_oinfo *loi, void *cookie,
1967                               int cmd, obd_off off, int count,
1968                               obd_flags brw_flags, enum async_flags async_flags)
1969 {
1970         struct client_obd *cli = &exp->exp_obd->u.cli;
1971         struct osc_async_page *oap;
1972         struct loi_oap_pages *lop;
1973         int rc;
1974         ENTRY;
1975
1976         oap = OAP_FROM_COOKIE(cookie);
1977
1978         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1979                 RETURN(-EIO);
1980
1981         if (!list_empty(&oap->oap_pending_item) ||
1982             !list_empty(&oap->oap_urgent_item) ||
1983             !list_empty(&oap->oap_rpc_item))
1984                 RETURN(-EBUSY);
1985
1986         if (loi == NULL)
1987                 loi = &lsm->lsm_oinfo[0];
1988
1989         spin_lock(&cli->cl_loi_list_lock);
1990
1991         oap->oap_cmd = cmd;
1992         oap->oap_async_flags = async_flags;
1993         oap->oap_page_off = off;
1994         oap->oap_count = count;
1995         oap->oap_brw_flags = brw_flags;
1996
1997         if (cmd == OBD_BRW_WRITE) {
1998                 rc = osc_enter_cache(cli, loi, oap);
1999                 if (rc) {
2000                         spin_unlock(&cli->cl_loi_list_lock);
2001                         RETURN(rc);
2002                 }
2003                 lop = &loi->loi_write_lop;
2004         } else {
2005                 lop = &loi->loi_read_lop;
2006         }
2007
2008         if (oap->oap_async_flags & ASYNC_URGENT)
2009                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2010         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2011         lop_update_pending(cli, lop, cmd, 1);
2012
2013         loi_list_maint(cli, loi);
2014
2015         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2016                   cmd);
2017
2018         osc_check_rpcs(cli);
2019         spin_unlock(&cli->cl_loi_list_lock);
2020
2021         RETURN(0);
2022 }
2023
2024 /* aka (~was & now & flag), but this is more clear :) */
2025 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2026
2027 static int osc_set_async_flags(struct obd_export *exp,
2028                                struct lov_stripe_md *lsm,
2029                                struct lov_oinfo *loi, void *cookie,
2030                                obd_flags async_flags)
2031 {
2032         struct client_obd *cli = &exp->exp_obd->u.cli;
2033         struct loi_oap_pages *lop;
2034         struct osc_async_page *oap;
2035         int rc = 0;
2036         ENTRY;
2037
2038         oap = OAP_FROM_COOKIE(cookie);
2039
2040         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2041                 RETURN(-EIO);
2042
2043         if (loi == NULL)
2044                 loi = &lsm->lsm_oinfo[0];
2045
2046         if (oap->oap_cmd == OBD_BRW_WRITE) {
2047                 lop = &loi->loi_write_lop;
2048         } else {
2049                 lop = &loi->loi_read_lop;
2050         }
2051
2052         spin_lock(&cli->cl_loi_list_lock);
2053
2054         if (list_empty(&oap->oap_pending_item))
2055                 GOTO(out, rc = -EINVAL);
2056
2057         if ((oap->oap_async_flags & async_flags) == async_flags)
2058                 GOTO(out, rc = 0);
2059
2060         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2061                 oap->oap_async_flags |= ASYNC_READY;
2062
2063         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2064                 if (list_empty(&oap->oap_rpc_item)) {
2065                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2066                         loi_list_maint(cli, loi);
2067                 }
2068         }
2069
2070         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2071                         oap->oap_async_flags);
2072 out:
2073         osc_check_rpcs(cli);
2074         spin_unlock(&cli->cl_loi_list_lock);
2075         RETURN(rc);
2076 }
2077
2078 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2079                              struct lov_oinfo *loi,
2080                              struct obd_io_group *oig, void *cookie,
2081                              int cmd, obd_off off, int count,
2082                              obd_flags brw_flags,
2083                              obd_flags async_flags)
2084 {
2085         struct client_obd *cli = &exp->exp_obd->u.cli;
2086         struct osc_async_page *oap;
2087         struct loi_oap_pages *lop;
2088         ENTRY;
2089
2090         oap = OAP_FROM_COOKIE(cookie);
2091
2092         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2093                 RETURN(-EIO);
2094
2095         if (!list_empty(&oap->oap_pending_item) ||
2096             !list_empty(&oap->oap_urgent_item) ||
2097             !list_empty(&oap->oap_rpc_item))
2098                 RETURN(-EBUSY);
2099
2100         if (loi == NULL)
2101                 loi = &lsm->lsm_oinfo[0];
2102
2103         spin_lock(&cli->cl_loi_list_lock);
2104
2105         oap->oap_cmd = cmd;
2106         oap->oap_page_off = off;
2107         oap->oap_count = count;
2108         oap->oap_brw_flags = brw_flags;
2109         oap->oap_async_flags = async_flags;
2110
2111         if (cmd == OBD_BRW_WRITE)
2112                 lop = &loi->loi_write_lop;
2113         else
2114                 lop = &loi->loi_read_lop;
2115
2116         list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2117         if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2118                 oap->oap_oig = oig;
2119                 oig_add_one(oig, &oap->oap_occ);
2120         }
2121
2122         LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
2123
2124         spin_unlock(&cli->cl_loi_list_lock);
2125
2126         RETURN(0);
2127 }
2128
2129 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2130                                  struct loi_oap_pages *lop, int cmd)
2131 {
2132         struct list_head *pos, *tmp;
2133         struct osc_async_page *oap;
2134
2135         list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2136                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2137                 list_del(&oap->oap_pending_item);
2138                 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2139                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2140                 lop_update_pending(cli, lop, cmd, 1);
2141         }
2142         loi_list_maint(cli, loi);
2143 }
2144
2145 static int osc_trigger_group_io(struct obd_export *exp,
2146                                 struct lov_stripe_md *lsm,
2147                                 struct lov_oinfo *loi,
2148                                 struct obd_io_group *oig)
2149 {
2150         struct client_obd *cli = &exp->exp_obd->u.cli;
2151         ENTRY;
2152
2153         if (loi == NULL)
2154                 loi = &lsm->lsm_oinfo[0];
2155
2156         spin_lock(&cli->cl_loi_list_lock);
2157
2158         osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2159         osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2160
2161         osc_check_rpcs(cli);
2162         spin_unlock(&cli->cl_loi_list_lock);
2163
2164         RETURN(0);
2165 }
2166
2167 static int osc_teardown_async_page(struct obd_export *exp,
2168                                    struct lov_stripe_md *lsm,
2169                                    struct lov_oinfo *loi, void *cookie)
2170 {
2171         struct client_obd *cli = &exp->exp_obd->u.cli;
2172         struct loi_oap_pages *lop;
2173         struct osc_async_page *oap;
2174         int rc = 0;
2175         ENTRY;
2176
2177         oap = OAP_FROM_COOKIE(cookie);
2178
2179         if (loi == NULL)
2180                 loi = &lsm->lsm_oinfo[0];
2181
2182         if (oap->oap_cmd == OBD_BRW_WRITE) {
2183                 lop = &loi->loi_write_lop;
2184         } else {
2185                 lop = &loi->loi_read_lop;
2186         }
2187
2188         spin_lock(&cli->cl_loi_list_lock);
2189
2190         if (!list_empty(&oap->oap_rpc_item))
2191                 GOTO(out, rc = -EBUSY);
2192
2193         osc_exit_cache(cli, oap, 0);
2194         osc_wake_cache_waiters(cli);
2195
2196         if (!list_empty(&oap->oap_urgent_item)) {
2197                 list_del_init(&oap->oap_urgent_item);
2198                 oap->oap_async_flags &= ~ASYNC_URGENT;
2199         }
2200         if (!list_empty(&oap->oap_pending_item)) {
2201                 list_del_init(&oap->oap_pending_item);
2202                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2203         }
2204         loi_list_maint(cli, loi);
2205
2206         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2207 out:
2208         spin_unlock(&cli->cl_loi_list_lock);
2209         if (rc == 0)
2210                 OBD_FREE(oap, sizeof(*oap));
2211         RETURN(rc);
2212 }
2213
2214 #ifdef __KERNEL__
2215 /* Note: caller will lock/unlock, and set uptodate on the pages */
2216 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2217 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2218                            struct lov_stripe_md *lsm, obd_count page_count,
2219                            struct brw_page *pga)
2220 {
2221         struct ptlrpc_request *request = NULL;
2222         struct ost_body *body;
2223         struct niobuf_remote *nioptr;
2224         struct obd_ioobj *iooptr;
2225         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2226         int swab;
2227         ENTRY;
2228
2229         /* XXX does not handle 'new' brw protocol */
2230
2231         size[1] = sizeof(struct obd_ioobj);
2232         size[2] = page_count * sizeof(*nioptr);
2233
2234         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2235                                   OST_SAN_READ, 3, size, NULL);
2236         if (!request)
2237                 RETURN(-ENOMEM);
2238
2239         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2240         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2241         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2242                                 sizeof(*nioptr) * page_count);
2243
2244         memcpy(&body->oa, oa, sizeof(body->oa));
2245
2246         obdo_to_ioobj(oa, iooptr);
2247         iooptr->ioo_bufcnt = page_count;
2248
2249         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2250                 LASSERT(PageLocked(pga[mapped].pg));
2251                 LASSERT(mapped == 0 ||
2252                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2253
2254                 nioptr->offset = pga[mapped].disk_offset;
2255                 nioptr->len    = pga[mapped].count;
2256                 nioptr->flags  = pga[mapped].flag;
2257         }
2258
2259         size[1] = page_count * sizeof(*nioptr);
2260         request->rq_replen = lustre_msg_size(2, size);
2261
2262         rc = ptlrpc_queue_wait(request);
2263         if (rc)
2264                 GOTO(out_req, rc);
2265
2266         body = lustre_swab_repbuf(request, 0, sizeof(*body),
2267                                   lustre_swab_ost_body);
2268         if (body == NULL) {
2269                 CERROR("Can't unpack body\n");
2270                 GOTO(out_req, rc = -EPROTO);
2271         }
2272
2273         memcpy(oa, &body->oa, sizeof(*oa));
2274
2275         swab = lustre_msg_swabbed(request->rq_repmsg);
2276         LASSERT_REPSWAB(request, 1);
2277         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2278         if (!nioptr) {
2279                 /* nioptr missing or short */
2280                 GOTO(out_req, rc = -EPROTO);
2281         }
2282
2283         /* actual read */
2284         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2285                 struct page *page = pga[mapped].pg;
2286                 struct buffer_head *bh;
2287                 kdev_t dev;
2288
2289                 if (swab)
2290                         lustre_swab_niobuf_remote (nioptr);
2291
2292                 /* got san device associated */
2293                 LASSERT(exp->exp_obd != NULL);
2294                 dev = exp->exp_obd->u.cli.cl_sandev;
2295
2296                 /* hole */
2297                 if (!nioptr->offset) {
2298                         CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2299                                         page->mapping->host->i_ino,
2300                                         page->index);
2301                         memset(page_address(page), 0, PAGE_SIZE);
2302                         continue;
2303                 }
2304
2305                 if (!page->buffers) {
2306                         create_empty_buffers(page, dev, PAGE_SIZE);
2307                         bh = page->buffers;
2308
2309                         clear_bit(BH_New, &bh->b_state);
2310                         set_bit(BH_Mapped, &bh->b_state);
2311                         bh->b_blocknr = (unsigned long)nioptr->offset;
2312
2313                         clear_bit(BH_Uptodate, &bh->b_state);
2314
2315                         ll_rw_block(READ, 1, &bh);
2316                 } else {
2317                         bh = page->buffers;
2318
2319                         /* if buffer already existed, it must be the
2320                          * one we mapped before, check it */
2321                         LASSERT(!test_bit(BH_New, &bh->b_state));
2322                         LASSERT(test_bit(BH_Mapped, &bh->b_state));
2323                         LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2324
2325                         /* wait it's io completion */
2326                         if (test_bit(BH_Lock, &bh->b_state))
2327                                 wait_on_buffer(bh);
2328
2329                         if (!test_bit(BH_Uptodate, &bh->b_state))
2330                                 ll_rw_block(READ, 1, &bh);
2331                 }
2332
2333
2334                 /* must do syncronous write here */
2335                 wait_on_buffer(bh);
2336                 if (!buffer_uptodate(bh)) {
2337                         /* I/O error */
2338                         rc = -EIO;
2339                         goto out_req;
2340                 }
2341         }
2342
2343 out_req:
2344         ptlrpc_req_finished(request);
2345         RETURN(rc);
2346 }
2347
2348 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2349                             struct lov_stripe_md *lsm, obd_count page_count,
2350                             struct brw_page *pga)
2351 {
2352         struct ptlrpc_request *request = NULL;
2353         struct ost_body *body;
2354         struct niobuf_remote *nioptr;
2355         struct obd_ioobj *iooptr;
2356         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2357         int swab;
2358         ENTRY;
2359
2360         size[1] = sizeof(struct obd_ioobj);
2361         size[2] = page_count * sizeof(*nioptr);
2362
2363         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2364                                   OST_SAN_WRITE, 3, size, NULL);
2365         if (!request)
2366                 RETURN(-ENOMEM);
2367
2368         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2369         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2370         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2371                                 sizeof (*nioptr) * page_count);
2372
2373         memcpy(&body->oa, oa, sizeof(body->oa));
2374
2375         obdo_to_ioobj(oa, iooptr);
2376         iooptr->ioo_bufcnt = page_count;
2377
2378         /* pack request */
2379         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2380                 LASSERT(PageLocked(pga[mapped].pg));
2381                 LASSERT(mapped == 0 ||
2382                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2383
2384                 nioptr->offset = pga[mapped].disk_offset;
2385                 nioptr->len    = pga[mapped].count;
2386                 nioptr->flags  = pga[mapped].flag;
2387         }
2388
2389         size[1] = page_count * sizeof(*nioptr);
2390         request->rq_replen = lustre_msg_size(2, size);
2391
2392         rc = ptlrpc_queue_wait(request);
2393         if (rc)
2394                 GOTO(out_req, rc);
2395
2396         swab = lustre_msg_swabbed (request->rq_repmsg);
2397         LASSERT_REPSWAB (request, 1);
2398         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2399         if (!nioptr) {
2400                 CERROR("absent/short niobuf array\n");
2401                 GOTO(out_req, rc = -EPROTO);
2402         }
2403
2404         /* actual write */
2405         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2406                 struct page *page = pga[mapped].pg;
2407                 struct buffer_head *bh;
2408                 kdev_t dev;
2409
2410                 if (swab)
2411                         lustre_swab_niobuf_remote (nioptr);
2412
2413                 /* got san device associated */
2414                 LASSERT(exp->exp_obd != NULL);
2415                 dev = exp->exp_obd->u.cli.cl_sandev;
2416
2417                 if (!page->buffers) {
2418                         create_empty_buffers(page, dev, PAGE_SIZE);
2419                 } else {
2420                         /* checking */
2421                         LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2422                         LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2423                         LASSERT(page->buffers->b_blocknr ==
2424                                 (unsigned long)nioptr->offset);
2425                 }
2426                 bh = page->buffers;
2427
2428                 LASSERT(bh);
2429
2430                 /* if buffer locked, wait it's io completion */
2431                 if (test_bit(BH_Lock, &bh->b_state))
2432                         wait_on_buffer(bh);
2433
2434                 clear_bit(BH_New, &bh->b_state);
2435                 set_bit(BH_Mapped, &bh->b_state);
2436
2437                 /* override the block nr */
2438                 bh->b_blocknr = (unsigned long)nioptr->offset;
2439
2440                 /* we are about to write it, so set it
2441                  * uptodate/dirty
2442                  * page lock should garentee no race condition here */
2443                 set_bit(BH_Uptodate, &bh->b_state);
2444                 set_bit(BH_Dirty, &bh->b_state);
2445
2446                 ll_rw_block(WRITE, 1, &bh);
2447
2448                 /* must do syncronous write here */
2449                 wait_on_buffer(bh);
2450                 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2451                         /* I/O error */
2452                         rc = -EIO;
2453                         goto out_req;
2454                 }
2455         }
2456
2457 out_req:
2458         ptlrpc_req_finished(request);
2459         RETURN(rc);
2460 }
2461
2462 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2463                       struct lov_stripe_md *lsm, obd_count page_count,
2464                       struct brw_page *pga, struct obd_trans_info *oti)
2465 {
2466         ENTRY;
2467
2468         while (page_count) {
2469                 obd_count pages_per_brw;
2470                 int rc;
2471
2472                 if (page_count > PTLRPC_MAX_BRW_PAGES)
2473                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2474                 else
2475                         pages_per_brw = page_count;
2476
2477                 if (cmd & OBD_BRW_WRITE)
2478                         rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2479                 else
2480                         rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2481
2482                 if (rc != 0)
2483                         RETURN(rc);
2484
2485                 page_count -= pages_per_brw;
2486                 pga += pages_per_brw;
2487         }
2488         RETURN(0);
2489 }
2490 #endif
2491 #endif
2492
2493 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2494 {
2495         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2496
2497         if (lock == NULL) {
2498                 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2499                 return;
2500         }
2501
2502         lock_res_and_lock(lock);
2503 #ifdef __KERNEL__
2504         if (lock->l_ast_data && lock->l_ast_data != data) {
2505                 struct inode *new_inode = data;
2506                 struct inode *old_inode = lock->l_ast_data;
2507                 if (!(old_inode->i_state & I_FREEING))
2508                         LDLM_ERROR(lock, "inconsistent l_ast_data found");
2509                 LASSERTF(old_inode->i_state & I_FREEING,
2510                          "Found existing inode %p/%lu/%u state %lu in lock: "
2511                          "setting data to %p/%lu/%u\n", old_inode,
2512                          old_inode->i_ino, old_inode->i_generation,
2513                          old_inode->i_state,
2514                          new_inode, new_inode->i_ino, new_inode->i_generation);
2515         }
2516 #endif
2517         lock->l_ast_data = data;
2518         unlock_res_and_lock(lock);
2519         LDLM_LOCK_PUT(lock);
2520 }
2521
2522 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2523                              ldlm_iterator_t replace, void *data)
2524 {
2525         struct ldlm_res_id res_id = { .name = {0} };
2526         struct obd_device *obd = class_exp2obd(exp);
2527
2528         res_id.name[0] = lsm->lsm_object_id;
2529         res_id.name[2] = lsm->lsm_object_gr;
2530         ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2531         return 0;
2532 }
2533
2534 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2535                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2536                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2537                        void *data, __u32 lvb_len, void *lvb_swabber,
2538                        struct lustre_handle *lockh)
2539 {
2540         struct obd_device *obd = exp->exp_obd;
2541         struct ldlm_res_id res_id = { .name = {0} };
2542         struct ost_lvb lvb;
2543         struct ldlm_reply *rep;
2544         struct ptlrpc_request *req = NULL;
2545         int rc;
2546         ENTRY;
2547
2548         res_id.name[0] = lsm->lsm_object_id;
2549         res_id.name[2] = lsm->lsm_object_gr;
2550
2551         /* Filesystem lock extents are extended to page boundaries so that
2552          * dealing with the page cache is a little smoother.  */
2553         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2554         policy->l_extent.end |= ~PAGE_MASK;
2555
2556         if (lsm->lsm_oinfo->loi_kms_valid == 0)
2557                 goto no_match;
2558
2559         /* Next, search for already existing extent locks that will cover us */
2560         rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2561                              lockh);
2562         if (rc == 1) {
2563                 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2564                         /* return immediately if no credential held */
2565                         ldlm_lock_decref(lockh, mode);
2566                         RETURN(-EACCES);
2567                 }
2568
2569                 osc_set_data_with_check(lockh, data);
2570                 if (*flags & LDLM_FL_HAS_INTENT) {
2571                         /* I would like to be able to ASSERT here that rss <=
2572                          * kms, but I can't, for reasons which are explained in
2573                          * lov_enqueue() */
2574                 }
2575                 /* We already have a lock, and it's referenced */
2576                 RETURN(ELDLM_OK);
2577         }
2578
2579         /* If we're trying to read, we also search for an existing PW lock.  The
2580          * VFS and page cache already protect us locally, so lots of readers/
2581          * writers can share a single PW lock.
2582          *
2583          * There are problems with conversion deadlocks, so instead of
2584          * converting a read lock to a write lock, we'll just enqueue a new
2585          * one.
2586          *
2587          * At some point we should cancel the read lock instead of making them
2588          * send us a blocking callback, but there are problems with canceling
2589          * locks out from other users right now, too. */
2590
2591         if (mode == LCK_PR) {
2592                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2593                                      policy, LCK_PW, lockh);
2594                 if (rc == 1) {
2595                         if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2596                                 /* return immediately if no credential held */
2597                                 ldlm_lock_decref(lockh, LCK_PW);
2598                                 RETURN(-EACCES);
2599                         }
2600
2601                         /* FIXME: This is not incredibly elegant, but it might
2602                          * be more elegant than adding another parameter to
2603                          * lock_match.  I want a second opinion. */
2604                         ldlm_lock_addref(lockh, LCK_PR);
2605                         ldlm_lock_decref(lockh, LCK_PW);
2606                         osc_set_data_with_check(lockh, data);
2607                         RETURN(ELDLM_OK);
2608                 }
2609         }
2610         if (mode == LCK_PW) {
2611                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2612                                      policy, LCK_PR, lockh);
2613                 if (rc == 1) {
2614                         rc = ldlm_cli_convert(lockh, mode, flags);
2615                         if (!rc) {
2616                                 /* Update readers/writers accounting */
2617                                 ldlm_lock_addref(lockh, LCK_PW);
2618                                 ldlm_lock_decref(lockh, LCK_PR);
2619                                 osc_set_data_with_check(lockh, data);
2620                                 RETURN(ELDLM_OK);
2621                         }
2622                         /* If the conversion failed, we need to drop refcount
2623                            on matched lock before we get new one */
2624                         /* XXX Won't it save us some efforts if we cancel PR
2625                            lock here? We are going to take PW lock anyway and it
2626                            will invalidate PR lock */
2627                         ldlm_lock_decref(lockh, LCK_PR);
2628                         if (rc != EDEADLOCK) {
2629                                 RETURN(rc);
2630                         }
2631                 }
2632         }
2633
2634 no_match:
2635         if (*flags & LDLM_FL_HAS_INTENT) {
2636                 int size[2] = {0, sizeof(struct ldlm_request)};
2637
2638                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
2639                                       LDLM_ENQUEUE, 2, size, NULL);
2640                 if (req == NULL)
2641                         RETURN(-ENOMEM);
2642
2643                 size[0] = sizeof(*rep);
2644                 size[1] = sizeof(lvb);
2645                 req->rq_replen = lustre_msg_size(2, size);
2646         }
2647         rc = ldlm_cli_enqueue(exp, req, obd->obd_namespace, res_id, type,
2648                               policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2649                               &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2650         if (req != NULL) {
2651                 if (rc == ELDLM_LOCK_ABORTED) {
2652                         /* swabbed by ldlm_cli_enqueue() */
2653                         LASSERT_REPSWABBED(req, 0);
2654                         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
2655                         LASSERT(rep != NULL);
2656                         if (rep->lock_policy_res1)
2657                                 rc = rep->lock_policy_res1;
2658                 }
2659                 ptlrpc_req_finished(req);
2660         }
2661
2662         if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2663                 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2664                        lvb.lvb_size, lvb.lvb_blocks);
2665                 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2666                 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2667         }
2668
2669         RETURN(rc);
2670 }
2671
2672 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2673                      __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2674                      int *flags, void *data, struct lustre_handle *lockh)
2675 {
2676         struct ldlm_res_id res_id = { .name = {0} };
2677         struct obd_device *obd = exp->exp_obd;
2678         int rc;
2679         ENTRY;
2680
2681         res_id.name[0] = lsm->lsm_object_id;
2682         res_id.name[2] = lsm->lsm_object_gr;
2683
2684         OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2685
2686         /* Filesystem lock extents are extended to page boundaries so that
2687          * dealing with the page cache is a little smoother */
2688         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2689         policy->l_extent.end |= ~PAGE_MASK;
2690
2691         /* Next, search for already existing extent locks that will cover us */
2692         rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2693                              policy, mode, lockh);
2694         if (rc) {
2695                // if (!(*flags & LDLM_FL_TEST_LOCK))
2696                         osc_set_data_with_check(lockh, data);
2697                 RETURN(rc);
2698         }
2699         /* If we're trying to read, we also search for an existing PW lock.  The
2700          * VFS and page cache already protect us locally, so lots of readers/
2701          * writers can share a single PW lock. */
2702         if (mode == LCK_PR) {
2703                 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2704                                      policy, LCK_PW, lockh);
2705                 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2706                         /* FIXME: This is not incredibly elegant, but it might
2707                          * be more elegant than adding another parameter to
2708                          * lock_match.  I want a second opinion. */
2709                         osc_set_data_with_check(lockh, data);
2710                         ldlm_lock_addref(lockh, LCK_PR);
2711                         ldlm_lock_decref(lockh, LCK_PW);
2712                 }
2713         }
2714         RETURN(rc);
2715 }
2716
2717 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2718                       __u32 mode, struct lustre_handle *lockh)
2719 {
2720         ENTRY;
2721
2722         if (mode == LCK_GROUP)
2723                 ldlm_lock_decref_and_cancel(lockh, mode);
2724         else
2725                 ldlm_lock_decref(lockh, mode);
2726
2727         RETURN(0);
2728 }
2729
2730 static int osc_cancel_unused(struct obd_export *exp,
2731                              struct lov_stripe_md *lsm,
2732                              int flags, void *opaque)
2733 {
2734         struct obd_device *obd = class_exp2obd(exp);
2735         struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2736
2737         if (lsm != NULL) {
2738                 res_id.name[0] = lsm->lsm_object_id;
2739                 res_id.name[2] = lsm->lsm_object_gr;
2740                 resp = &res_id;
2741         }
2742
2743         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2744 }
2745
2746 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2747                       unsigned long max_age)
2748 {
2749         struct obd_statfs *msfs;
2750         struct ptlrpc_request *request;
2751         int rc, size = sizeof(*osfs);
2752         ENTRY;
2753
2754         /* We could possibly pass max_age in the request (as an absolute
2755          * timestamp or a "seconds.usec ago") so the target can avoid doing
2756          * extra calls into the filesystem if that isn't necessary (e.g.
2757          * during mount that would help a bit).  Having relative timestamps
2758          * is not so great if request processing is slow, while absolute
2759          * timestamps are not ideal because they need time synchronization. */
2760         request = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OBD_VERSION,
2761                                   OST_STATFS, 0, NULL, NULL);
2762         if (!request)
2763                 RETURN(-ENOMEM);
2764
2765         request->rq_replen = lustre_msg_size(1, &size);
2766         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2767
2768         rc = ptlrpc_queue_wait(request);
2769         if (rc)
2770                 GOTO(out, rc);
2771
2772         msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2773                                   lustre_swab_obd_statfs);
2774         if (msfs == NULL) {
2775                 CERROR("Can't unpack obd_statfs\n");
2776                 GOTO(out, rc = -EPROTO);
2777         }
2778
2779         memcpy(osfs, msfs, sizeof(*osfs));
2780
2781         EXIT;
2782  out:
2783         ptlrpc_req_finished(request);
2784         return rc;
2785 }
2786
2787 /* Retrieve object striping information.
2788  *
2789  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2790  * the maximum number of OST indices which will fit in the user buffer.
2791  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2792  */
2793 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2794 {
2795         struct lov_user_md lum, *lumk;
2796         int rc, lum_size;
2797         ENTRY;
2798
2799         if (!lsm)
2800                 RETURN(-ENODATA);
2801
2802         rc = copy_from_user(&lum, lump, sizeof(lum));
2803         if (rc)
2804                 RETURN(-EFAULT);
2805
2806         if (lum.lmm_magic != LOV_USER_MAGIC)
2807                 RETURN(-EINVAL);
2808
2809         if (lum.lmm_stripe_count > 0) {
2810                 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2811                 OBD_ALLOC(lumk, lum_size);
2812                 if (!lumk)
2813                         RETURN(-ENOMEM);
2814
2815                 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2816                 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2817         } else {
2818                 lum_size = sizeof(lum);
2819                 lumk = &lum;
2820         }
2821
2822         lumk->lmm_object_id = lsm->lsm_object_id;
2823         lumk->lmm_object_gr = lsm->lsm_object_gr;
2824         lumk->lmm_stripe_count = 1;
2825
2826         if (copy_to_user(lump, lumk, lum_size))
2827                 rc = -EFAULT;
2828
2829         if (lumk != &lum)
2830                 OBD_FREE(lumk, lum_size);
2831
2832         RETURN(rc);
2833 }
2834
2835 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2836                          void *karg, void *uarg)
2837 {
2838         struct obd_device *obd = exp->exp_obd;
2839         struct obd_ioctl_data *data = karg;
2840         int err = 0;
2841         ENTRY;
2842
2843 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2844         MOD_INC_USE_COUNT;
2845 #else
2846         if (!try_module_get(THIS_MODULE)) {
2847                 CERROR("Can't get module. Is it alive?");
2848                 return -EINVAL;
2849         }
2850 #endif
2851         switch (cmd) {
2852         case OBD_IOC_LOV_GET_CONFIG: {
2853                 char *buf;
2854                 struct lov_desc *desc;
2855                 struct obd_uuid uuid;
2856
2857                 buf = NULL;
2858                 len = 0;
2859                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2860                         GOTO(out, err = -EINVAL);
2861
2862                 data = (struct obd_ioctl_data *)buf;
2863
2864                 if (sizeof(*desc) > data->ioc_inllen1) {
2865                         OBD_FREE(buf, len);
2866                         GOTO(out, err = -EINVAL);
2867                 }
2868
2869                 if (data->ioc_inllen2 < sizeof(uuid)) {
2870                         OBD_FREE(buf, len);
2871                         GOTO(out, err = -EINVAL);
2872                 }
2873
2874                 if (data->ioc_inllen3 < sizeof(__u32)) {
2875                         OBD_FREE(buf, len);
2876                         GOTO(out, err = -EINVAL);
2877                 }
2878
2879                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2880                 desc->ld_tgt_count = 1;
2881                 desc->ld_active_tgt_count = 1;
2882                 desc->ld_default_stripe_count = 1;
2883                 desc->ld_default_stripe_size = 0;
2884                 desc->ld_default_stripe_offset = 0;
2885                 desc->ld_pattern = 0;
2886                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2887                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2888                 *((__u32 *)data->ioc_inlbuf3) = 1;
2889
2890                 err = copy_to_user((void *)uarg, buf, len);
2891                 if (err)
2892                         err = -EFAULT;
2893                 obd_ioctl_freedata(buf, len);
2894                 GOTO(out, err);
2895         }
2896         case LL_IOC_LOV_SETSTRIPE:
2897                 err = obd_alloc_memmd(exp, karg);
2898                 if (err > 0)
2899                         err = 0;
2900                 GOTO(out, err);
2901         case LL_IOC_LOV_GETSTRIPE:
2902                 err = osc_getstripe(karg, uarg);
2903                 GOTO(out, err);
2904         case OBD_IOC_CLIENT_RECOVER:
2905                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2906                                             data->ioc_inlbuf1);
2907                 if (err > 0)
2908                         err = 0;
2909                 GOTO(out, err);
2910         case IOC_OSC_SET_ACTIVE:
2911                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2912                                                data->ioc_offset);
2913                 GOTO(out, err);
2914         case IOC_OSC_CTL_RECOVERY:
2915                 err = ptlrpc_import_control_recovery(obd->u.cli.cl_import,
2916                                                      data->ioc_offset);
2917                 GOTO(out, err);
2918         default:
2919                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2920                 GOTO(out, err = -ENOTTY);
2921         }
2922 out:
2923 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2924         MOD_DEC_USE_COUNT;
2925 #else
2926         module_put(THIS_MODULE);
2927 #endif
2928         return err;
2929 }
2930
2931 static int osc_get_info(struct obd_export *exp, __u32 keylen,
2932                         void *key, __u32 *vallen, void *val)
2933 {
2934         ENTRY;
2935         if (!vallen || !val)
2936                 RETURN(-EFAULT);
2937
2938         if (keylen > strlen("lock_to_stripe") &&
2939             strcmp(key, "lock_to_stripe") == 0) {
2940                 __u32 *stripe = val;
2941                 *vallen = sizeof(*stripe);
2942                 *stripe = 0;
2943                 RETURN(0);
2944         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2945                 struct ptlrpc_request *req;
2946                 obd_id *reply;
2947                 char *bufs[1] = {key};
2948                 int rc;
2949                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2950                                       OST_GET_INFO, 1, (int *)&keylen, bufs);
2951                 if (req == NULL)
2952                         RETURN(-ENOMEM);
2953
2954                 req->rq_replen = lustre_msg_size(1, (int *)vallen);
2955                 rc = ptlrpc_queue_wait(req);
2956                 if (rc)
2957                         GOTO(out, rc);
2958
2959                 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2960                                            lustre_swab_ost_last_id);
2961                 if (reply == NULL) {
2962                         CERROR("Can't unpack OST last ID\n");
2963                         GOTO(out, rc = -EPROTO);
2964                 }
2965                 *((obd_id *)val) = *reply;
2966         out:
2967                 ptlrpc_req_finished(req);
2968                 RETURN(rc);
2969         } else if (keylen == 10 && strcmp(key, "client_nid") == 0) {
2970                 struct ptlrpc_connection * conn;
2971                 ptl_nid_t * nid = val;
2972                 ptl_process_id_t id;
2973                 int rc;
2974
2975                 *vallen = sizeof(*nid);
2976                 conn = class_exp2cliimp(exp)->imp_connection;
2977
2978                 if (!conn || !conn->c_peer.peer_ni)
2979                         RETURN(-ENOTCONN);
2980
2981                 rc = PtlGetId(conn->c_peer.peer_ni->pni_ni_h, &id);
2982                 if (rc == PTL_OK)
2983                         *nid = id.nid;
2984
2985                 RETURN(0);
2986         }
2987         RETURN(-EPROTO);
2988 }
2989
2990 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2991                         void *key, obd_count vallen, void *val)
2992 {
2993         struct obd_device  *obd = exp->exp_obd;
2994         struct obd_import *imp = class_exp2cliimp(exp);
2995         struct llog_ctxt *ctxt;
2996         int rc = 0;
2997         ENTRY;
2998
2999         if (keylen == strlen("unlinked") &&
3000             memcmp(key, "unlinked", keylen) == 0) {
3001                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3002                 spin_lock(&oscc->oscc_lock);
3003                 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3004                 spin_unlock(&oscc->oscc_lock);
3005                 RETURN(0);
3006         }
3007
3008         if (keylen == strlen("unrecovery") &&
3009             memcmp(key, "unrecovery", keylen) == 0) {
3010                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3011                 spin_lock(&oscc->oscc_lock);
3012                 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3013                 spin_unlock(&oscc->oscc_lock);
3014                 RETURN(0);
3015         }
3016
3017         if (keylen == strlen("initial_recov") &&
3018             memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
3019                 struct obd_import *imp = class_exp2cliimp(exp);
3020                 if (vallen != sizeof(int))
3021                         RETURN(-EINVAL);
3022                 imp->imp_initial_recov = *(int *)val;
3023                 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
3024                        exp->exp_obd->obd_name,
3025                        imp->imp_initial_recov);
3026                 RETURN(0);
3027         }
3028
3029         if (keylen == strlen("async") &&
3030             memcmp(key, "async", keylen) == 0) {
3031                 struct client_obd *cl = &obd->u.cli;
3032                 if (vallen != sizeof(int))
3033                         RETURN(-EINVAL);
3034                 cl->cl_async = *(int *)val;
3035                 CDEBUG(D_HA, "%s: set async = %d\n",
3036                        obd->obd_name, cl->cl_async);
3037                 RETURN(0);
3038         }
3039
3040         if (keylen == 5 && strcmp(key, "audit") == 0) {
3041                 struct ptlrpc_request *req;
3042                 char *bufs[2] = {key, val};
3043                 int size[2] = {keylen, vallen};
3044
3045                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3046                                       OST_SET_INFO, 2, size, bufs);
3047                 if (req == NULL)
3048                         RETURN(-ENOMEM);
3049
3050                 req->rq_replen = lustre_msg_size(0, size);
3051                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
3052                                    lustre_swab_audit_attr);
3053                 rc = ptlrpc_queue_wait(req);
3054
3055                 ptlrpc_req_finished(req);
3056                 RETURN(rc);
3057         }
3058
3059         if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
3060                 struct ptlrpc_request *req;
3061                 char *bufs[2] = {key, val};
3062                 int size[2] = {keylen, vallen};
3063
3064                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3065                                       OST_SET_INFO, 2, size, bufs);
3066                 if (req == NULL)
3067                         RETURN(-ENOMEM);
3068
3069                 req->rq_replen = lustre_msg_size(0, size);
3070                 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
3071                                    lustre_swab_obdo);
3072                 rc = ptlrpc_queue_wait(req);
3073
3074                 ptlrpc_req_finished(req);
3075                 RETURN(rc);
3076         }
3077
3078         if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
3079                 struct ptlrpc_request *req;
3080                 char *bufs[2] = {key, val};
3081                 int size[2] = {keylen, vallen};
3082
3083                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3084                                       OST_SET_INFO, 2, size, bufs);
3085                 if (req == NULL)
3086                         RETURN(-ENOMEM);
3087
3088                 req->rq_replen = lustre_msg_size(0, size);
3089                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
3090                                    lustre_swab_audit_msg);
3091                 rc = ptlrpc_queue_wait(req);
3092
3093                 ptlrpc_req_finished(req);
3094                 RETURN(rc);
3095         }
3096
3097         if (keylen == strlen("sec") && memcmp(key, "sec", keylen) == 0) {
3098                 struct client_obd *cli = &exp->exp_obd->u.cli;
3099
3100                 cli->cl_sec_flavor = ptlrpcs_name2flavor(val);
3101                 if (cli->cl_sec_flavor == PTLRPCS_FLVR_INVALID) {
3102                         CERROR("unrecognized security flavor %s\n", (char*) val);
3103                         RETURN(-EINVAL);
3104                 }
3105
3106                 RETURN(0);
3107         }
3108
3109         if (keylen == strlen("sec_flags") &&
3110             memcmp(key, "sec_flags", keylen) == 0) {
3111                 struct client_obd *cli = &exp->exp_obd->u.cli;
3112
3113                 cli->cl_sec_flags = *((unsigned long *) val);
3114                 RETURN(0);
3115         }
3116
3117         if (keylen == strlen("flush_cred") &&
3118             memcmp(key, "flush_cred", keylen) == 0) {
3119                 struct client_obd *cli = &exp->exp_obd->u.cli;
3120
3121                 if (cli->cl_import)
3122                         ptlrpcs_import_flush_current_creds(cli->cl_import);
3123                 RETURN(0);
3124         }
3125         if (keylen == strlen("crypto_cb") &&
3126             memcmp(key, "crypto_cb", keylen) == 0) {
3127                 LASSERT(vallen == sizeof(crypt_cb_t));
3128                 osc_crypt_cb = (crypt_cb_t)val;
3129                 RETURN(0);
3130         }
3131
3132         if (keylen == 8 && memcmp(key, "capa_key", 8) == 0) {
3133                 struct ptlrpc_request *req;
3134                 char *bufs[2] = {key, val};
3135                 unsigned long irqflags;
3136                 int rc, size[2] = {keylen, vallen};
3137
3138                 LASSERT(vallen == sizeof(struct lustre_capa_key));
3139
3140                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3141                                       OST_SET_INFO, 2, size, bufs);
3142                 if (req == NULL)
3143                         RETURN(-ENOMEM);
3144
3145                 spin_lock_irqsave (&req->rq_lock, irqflags);
3146                 req->rq_replay = 1;
3147                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
3148
3149                 req->rq_replen = lustre_msg_size(0, NULL);
3150                 rc = ptlrpc_queue_wait(req);
3151                 ptlrpc_req_finished(req);
3152                 RETURN(rc);
3153         }
3154
3155         if (keylen == strlen("setext") &&
3156             memcmp(key, "setext", keylen) == 0) {
3157                 struct client_obd *cli = &exp->exp_obd->u.cli;
3158                 struct osc_creator *oscc = &cli->cl_oscc;
3159                 struct fid_extent *ext = val;
3160
3161                 oscc->oscc_next_id = (obd_id)ext->fe_start;
3162                 RETURN(0);
3163         }
3164
3165         if (keylen < strlen("mds_conn") ||
3166             memcmp(key, "mds_conn", keylen) != 0)
3167                 RETURN(-EINVAL);
3168
3169         ctxt = llog_get_context(&exp->exp_obd->obd_llogs,
3170                                 LLOG_UNLINK_ORIG_CTXT);
3171         if (ctxt) {
3172                 if (rc == 0)
3173                         rc = llog_initiator_connect(ctxt);
3174                 else
3175                         CERROR("cannot establish the connect for "
3176                                "ctxt %p: %d\n", ctxt, rc);
3177         }
3178
3179         imp->imp_server_timeout = 1;
3180         CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
3181         imp->imp_pingable = 1;
3182
3183         RETURN(rc);
3184 }
3185
3186
3187 static struct llog_operations osc_size_repl_logops = {
3188         lop_cancel: llog_obd_repl_cancel
3189 };
3190
3191 static struct llog_operations osc_unlink_orig_logops;
3192
3193 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
3194                          struct obd_device *tgt, int count,
3195                          struct llog_catid *catid)
3196 {
3197         int rc;
3198         ENTRY;
3199
3200         osc_unlink_orig_logops = llog_lvfs_ops;
3201         osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
3202         osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
3203         osc_unlink_orig_logops.lop_add = llog_catalog_add;
3204         osc_unlink_orig_logops.lop_connect = llog_origin_connect;
3205
3206         rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
3207                             &catid->lci_logid, &osc_unlink_orig_logops);
3208         if (rc)
3209                 RETURN(rc);
3210
3211         rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
3212                             &osc_size_repl_logops);
3213         RETURN(rc);
3214 }
3215
3216 static int osc_llog_finish(struct obd_device *obd,
3217                            struct obd_llogs *llogs, int count)
3218 {
3219         int rc;
3220         ENTRY;
3221
3222         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
3223         if (rc)
3224                 RETURN(rc);
3225
3226         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
3227         RETURN(rc);
3228 }
3229
3230 static int osc_connect(struct lustre_handle *exph,
3231                        struct obd_device *obd, struct obd_uuid *cluuid,
3232                        struct obd_connect_data *data,
3233                        unsigned long connect_flags)
3234 {
3235         int rc;
3236         ENTRY;
3237         rc = client_connect_import(exph, obd, cluuid, data, connect_flags);
3238         RETURN(rc);
3239 }
3240
3241 static int osc_disconnect(struct obd_export *exp, unsigned long flags)
3242 {
3243         struct obd_device *obd = class_exp2obd(exp);
3244         struct llog_ctxt *ctxt;
3245         int rc;
3246         ENTRY;
3247
3248         ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
3249         if (obd->u.cli.cl_conn_count == 1)
3250                 /* flush any remaining cancel messages out to the target */
3251                 llog_sync(ctxt, exp);
3252
3253         rc = client_disconnect_export(exp, flags);
3254         RETURN(rc);
3255 }
3256
3257 static int osc_import_event(struct obd_device *obd,
3258                             struct obd_import *imp,
3259                             enum obd_import_event event)
3260 {
3261         struct client_obd *cli;
3262         int rc = 0;
3263         ENTRY;
3264
3265         LASSERT(imp->imp_obd == obd);
3266
3267         switch (event) {
3268         case IMP_EVENT_DISCON: {
3269                 /* Only do this on the MDS OSC's */
3270                 if (imp->imp_server_timeout) {
3271                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3272
3273                         spin_lock(&oscc->oscc_lock);
3274                         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3275                         spin_unlock(&oscc->oscc_lock);
3276                 }
3277                 break;
3278         }
3279         case IMP_EVENT_INACTIVE: {
3280                 if (obd->obd_observer)
3281                         rc = obd_notify(obd->obd_observer, obd, 0, 0);
3282                 break;
3283         }
3284         case IMP_EVENT_INVALIDATE: {
3285                 struct ldlm_namespace *ns = obd->obd_namespace;
3286
3287                 /* Reset grants */
3288                 cli = &obd->u.cli;
3289                 spin_lock(&cli->cl_loi_list_lock);
3290                 cli->cl_avail_grant = 0;
3291                 cli->cl_lost_grant = 0;
3292                 /* all pages go to failing rpcs due to the invalid import */
3293                 osc_check_rpcs(cli);
3294                 spin_unlock(&cli->cl_loi_list_lock);
3295
3296                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3297
3298                 break;
3299         }
3300         case IMP_EVENT_ACTIVE: {
3301                 /* Only do this on the MDS OSC's */
3302                 if (imp->imp_server_timeout) {
3303                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3304
3305                         spin_lock(&oscc->oscc_lock);
3306                         oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3307                         spin_unlock(&oscc->oscc_lock);
3308                 }
3309
3310                 if (obd->obd_observer)
3311                         rc = obd_notify(obd->obd_observer, obd, 1, 0);
3312                 break;
3313         }
3314         default:
3315                 CERROR("Unknown import event %d\n", event);
3316                 LBUG();
3317         }
3318         RETURN(rc);
3319 }
3320
3321 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
3322 {
3323         struct lprocfs_static_vars lvars;
3324         int rc;
3325         ENTRY;
3326
3327         lprocfs_init_vars(osc,&lvars);
3328         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3329         if (rc < 0)
3330                 RETURN(rc);
3331
3332         rc = lproc_osc_attach_seqstat(dev);
3333         if (rc < 0) {
3334                 lprocfs_obd_detach(dev);
3335                 RETURN(rc);
3336         }
3337
3338         ptlrpc_lprocfs_register_obd(dev);
3339         RETURN(0);
3340 }
3341
3342 static int osc_detach(struct obd_device *dev)
3343 {
3344         ptlrpc_lprocfs_unregister_obd(dev);
3345         return lprocfs_obd_detach(dev);
3346 }
3347
3348 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3349 {
3350         int rc;
3351         ENTRY;
3352         rc = ptlrpcd_addref();
3353         if (rc)
3354                 RETURN(rc);
3355
3356         rc = client_obd_setup(obd, len, buf);
3357         if (rc)
3358                 ptlrpcd_decref();
3359         else
3360                 oscc_init(obd);
3361
3362         RETURN(rc);
3363 }
3364
3365 static int osc_cleanup(struct obd_device *obd, int flags)
3366 {
3367         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3368         int rc;
3369
3370         rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3371                                     LDLM_FL_CONFIG_CHANGE, NULL);
3372         if (rc)
3373                 RETURN(rc);
3374
3375         spin_lock(&oscc->oscc_lock);
3376         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3377         oscc->oscc_flags |= OSCC_FLAG_EXITING;
3378         spin_unlock(&oscc->oscc_lock);
3379
3380         rc = client_obd_cleanup(obd, flags);
3381         ptlrpcd_decref();
3382         RETURN(rc);
3383 }
3384
3385
3386 struct obd_ops osc_obd_ops = {
3387         .o_owner                = THIS_MODULE,
3388         .o_attach               = osc_attach,
3389         .o_detach               = osc_detach,
3390         .o_setup                = osc_setup,
3391         .o_cleanup              = osc_cleanup,
3392         .o_add_conn             = client_import_add_conn,
3393         .o_del_conn             = client_import_del_conn,
3394         .o_connect              = osc_connect,
3395         .o_disconnect           = osc_disconnect,
3396         .o_statfs               = osc_statfs,
3397         .o_packmd               = osc_packmd,
3398         .o_unpackmd             = osc_unpackmd,
3399         .o_create               = osc_create,
3400         .o_destroy              = osc_destroy,
3401         .o_getattr              = osc_getattr,
3402         .o_getattr_async        = osc_getattr_async,
3403         .o_setattr              = osc_setattr,
3404         .o_brw                  = osc_brw,
3405         .o_brw_async            = osc_brw_async,
3406         .o_prep_async_page      = osc_prep_async_page,
3407         .o_queue_async_io       = osc_queue_async_io,
3408         .o_set_async_flags      = osc_set_async_flags,
3409         .o_queue_group_io       = osc_queue_group_io,
3410         .o_trigger_group_io     = osc_trigger_group_io,
3411         .o_teardown_async_page  = osc_teardown_async_page,
3412         .o_punch                = osc_punch,
3413         .o_sync                 = osc_sync,
3414         .o_enqueue              = osc_enqueue,
3415         .o_match                = osc_match,
3416         .o_change_cbdata        = osc_change_cbdata,
3417         .o_cancel               = osc_cancel,
3418         .o_cancel_unused        = osc_cancel_unused,
3419         .o_iocontrol            = osc_iocontrol,
3420         .o_get_info             = osc_get_info,
3421         .o_set_info             = osc_set_info,
3422         .o_import_event         = osc_import_event,
3423         .o_llog_init            = osc_llog_init,
3424         .o_llog_finish          = osc_llog_finish,
3425 };
3426
3427 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3428 struct obd_ops sanosc_obd_ops = {
3429         .o_owner                = THIS_MODULE,
3430         .o_attach               = osc_attach,
3431         .o_detach               = osc_detach,
3432         .o_cleanup              = client_obd_cleanup,
3433         .o_add_conn             = client_import_add_conn,
3434         .o_del_conn             = client_import_del_conn,
3435         .o_connect              = osc_connect,
3436         .o_disconnect           = client_disconnect_export,
3437         .o_statfs               = osc_statfs,
3438         .o_packmd               = osc_packmd,
3439         .o_unpackmd             = osc_unpackmd,
3440         .o_create               = osc_real_create,
3441         .o_destroy              = osc_destroy,
3442         .o_getattr              = osc_getattr,
3443         .o_getattr_async        = osc_getattr_async,
3444         .o_setattr              = osc_setattr,
3445         .o_setup                = client_sanobd_setup,
3446         .o_brw                  = sanosc_brw,
3447         .o_punch                = osc_punch,
3448         .o_sync                 = osc_sync,
3449         .o_enqueue              = osc_enqueue,
3450         .o_match                = osc_match,
3451         .o_change_cbdata        = osc_change_cbdata,
3452         .o_cancel               = osc_cancel,
3453         .o_cancel_unused        = osc_cancel_unused,
3454         .o_iocontrol            = osc_iocontrol,
3455         .o_import_event         = osc_import_event,
3456         .o_llog_init            = osc_llog_init,
3457         .o_llog_finish          = osc_llog_finish,
3458 };
3459 #endif
3460
3461 int __init osc_init(void)
3462 {
3463         struct lprocfs_static_vars lvars;
3464 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3465         struct lprocfs_static_vars sanlvars;
3466 #endif
3467         int rc;
3468         ENTRY;
3469
3470         lprocfs_init_vars(osc, &lvars);
3471 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3472         lprocfs_init_vars(osc, &sanlvars);
3473 #endif
3474
3475         rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3476                                  OBD_OSC_DEVICENAME);
3477         if (rc)
3478                 RETURN(rc);
3479
3480 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3481         rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3482                                  OBD_SANOSC_DEVICENAME);
3483         if (rc)
3484                 class_unregister_type(OBD_OSC_DEVICENAME);
3485 #endif
3486
3487         RETURN(rc);
3488 }
3489
3490 #ifdef __KERNEL__
3491 static void /*__exit*/ osc_exit(void)
3492 {
3493 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3494         class_unregister_type(OBD_SANOSC_DEVICENAME);
3495 #endif
3496         class_unregister_type(OBD_OSC_DEVICENAME);
3497 }
3498
3499 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3500 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3501 MODULE_LICENSE("GPL");
3502
3503 module_init(osc_init);
3504 module_exit(osc_exit);
3505 #endif