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                        struct lustre_capa *capa)
 258 {
 259         struct ptlrpc_request *request;
 260         struct ost_body *body;
 261         int rc, size = sizeof(*body);
 262         ENTRY;
 263
 264         LASSERT(!(oa->o_valid & OBD_MD_FLGROUP) || oa->o_gr > 0);
 265
 266         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 267                                   OST_SETATTR, 1, &size, NULL);
 268         if (!request)
 269                 RETURN(-ENOMEM);
 270
 271         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
 272         memcpy(&body->oa, oa, sizeof(*oa));
 273
 274         request->rq_replen = lustre_msg_size(1, &size);
 275
 276         if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
 277                 ptlrpcd_add_req(request);
 278                 rc = 0;
 279         } else {
 280                 rc = ptlrpc_queue_wait(request);
 281                 if (rc)
 282                         GOTO(out, rc);
 283
 284                 body = lustre_swab_repbuf(request, 0, sizeof(*body),
 285                                           lustre_swab_ost_body);
 286                 if (body == NULL)
 287                         GOTO(out, rc = -EPROTO);
 288
 289                 memcpy(oa, &body->oa, sizeof(*oa));
 290         }
 291         EXIT;
 292 out:
 293         ptlrpc_req_finished(request);
 294         RETURN(0);
 295 }
 296
 297 int osc_real_create(struct obd_export *exp, struct obdo *oa,
 298                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
 299 {
 300         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 301         struct ptlrpc_request *request;
 302         struct ost_body *body;
 303         struct lov_stripe_md *lsm;
 304         int rc, size = sizeof(*body);
 305         ENTRY;
 306
 307         LASSERT(oa);
 308         LASSERT(ea);
 309
 310         lsm = *ea;
 311         if (!lsm) {
 312                 rc = obd_alloc_memmd(exp, &lsm);
 313                 if (rc < 0)
 314                         RETURN(rc);
 315         }
 316
 317         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 318                                   OST_CREATE, 1, &size, NULL);
 319         if (!request)
 320                 GOTO(out, rc = -ENOMEM);
 321
 322         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 323         memcpy(&body->oa, oa, sizeof(body->oa));
 324
 325         request->rq_replen = lustre_msg_size(1, &size);
 326         if (oa->o_valid & OBD_MD_FLINLINE) {
 327                 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
 328                         oa->o_flags == OBD_FL_DELORPHAN);
 329                 DEBUG_REQ(D_HA, request,
 330                           "delorphan from OST integration");
 331                 /* Don't resend the delorphan request */
 332                 request->rq_no_resend = request->rq_no_delay = 1;
 333         }
 334
 335         rc = ptlrpc_queue_wait(request);
 336         if (rc)
 337                 GOTO(out_req, rc);
 338
 339         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 340                                   lustre_swab_ost_body);
 341         if (body == NULL) {
 342                 CERROR ("can't unpack ost_body\n");
 343                 GOTO (out_req, rc = -EPROTO);
 344         }
 345
 346         if ((oa->o_valid & OBD_MD_FLFLAGS) && oa->o_flags == OBD_FL_DELORPHAN) {
 347                 struct obd_import *imp = class_exp2cliimp(exp);
 348                 /* MDS declares last known object, OSS responses
 349                  * with next possible object -bzzz */
 350                 spin_lock(&oscc->oscc_lock);
 351                 oscc->oscc_next_id = body->oa.o_id;
 352                 spin_unlock(&oscc->oscc_lock);
 353                 CDEBUG(D_HA, "%s: set nextid "LPD64" after recovery\n",
 354                        imp->imp_target_uuid.uuid, oa->o_id);
 355         }
 356         memcpy(oa, &body->oa, sizeof(*oa));
 357
 358         /* This should really be sent by the OST */
 359         oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
 360         oa->o_valid |= OBD_MD_FLBLKSZ;
 361
 362         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
 363          * have valid lsm_oinfo data structs, so don't go touching that.
 364          * This needs to be fixed in a big way.
 365          */
 366         lsm->lsm_object_id = oa->o_id;
 367         lsm->lsm_object_gr = oa->o_gr;
 368         *ea = lsm;
 369
 370         if (oti != NULL) {
 371                 oti->oti_transno = request->rq_repmsg->transno;
 372
 373                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
 374                         if (!oti->oti_logcookies)
 375                                 oti_alloc_cookies(oti, 1);
 376                         memcpy(oti->oti_logcookies, obdo_logcookie(oa),
 377                                sizeof(oti->oti_onecookie));
 378                 }
 379         }
 380
 381         CDEBUG(D_HA, "transno: "LPD64"\n", request->rq_repmsg->transno);
 382         EXIT;
 383 out_req:
 384         ptlrpc_req_finished(request);
 385 out:
 386         if (rc && !*ea)
 387                 obd_free_memmd(exp, &lsm);
 388         return rc;
 389 }
 390
 391 static int osc_punch(struct obd_export *exp, struct obdo *oa,
 392                      struct lov_stripe_md *md, obd_size start,
 393                      obd_size end, struct obd_trans_info *oti,
 394                      struct lustre_capa *capa)
 395 {
 396         struct ptlrpc_request *request;
 397         struct ost_body *body;
 398         struct lustre_capa *req_capa;
 399         int bufcnt = 0;
 400         int rc, size[2] = { sizeof(*body), sizeof(*capa) };
 401         ENTRY;
 402
 403         if (!oa) {
 404                 CERROR("oa NULL\n");
 405                 RETURN(-EINVAL);
 406         }
 407
 408         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 409                                   OST_PUNCH, capa ? 2 : 1, size, NULL);
 410         if (!request)
 411                 RETURN(-ENOMEM);
 412
 413         body = lustre_msg_buf(request->rq_reqmsg, bufcnt++, sizeof (*body));
 414
 415         memcpy(&body->oa, oa, sizeof(*oa));
 416
 417         /* overload the size and blocks fields in the oa with start/end */
 418         body->oa.o_size = start;
 419         body->oa.o_blocks = end;
 420         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 421
 422         if (capa) {
 423                 req_capa = lustre_msg_buf(request->rq_reqmsg, bufcnt++,
 424                                           sizeof(*capa));
 425                 capa_dup2(req_capa, capa);
 426                 body->oa.o_valid |= OBD_MD_CAPA;
 427         }
 428
 429         request->rq_replen = lustre_msg_size(1, size);
 430
 431         rc = ptlrpc_queue_wait(request);
 432         if (rc)
 433                 GOTO(out, rc);
 434
 435         body = lustre_swab_repbuf (request, 0, sizeof (*body),
 436                                    lustre_swab_ost_body);
 437         if (body == NULL) {
 438                 CERROR ("can't unpack ost_body\n");
 439                 GOTO (out, rc = -EPROTO);
 440         }
 441
 442         memcpy(oa, &body->oa, sizeof(*oa));
 443
 444         EXIT;
 445  out:
 446         ptlrpc_req_finished(request);
 447         return rc;
 448 }
 449
 450 static int osc_sync(struct obd_export *exp, struct obdo *oa,
 451                     struct lov_stripe_md *md, obd_size start,
 452                     obd_size end)
 453 {
 454         struct ptlrpc_request *request;
 455         struct ost_body *body;
 456         int rc, size = sizeof(*body);
 457         ENTRY;
 458
 459         if (!oa) {
 460                 CERROR("oa NULL\n");
 461                 RETURN(-EINVAL);
 462         }
 463
 464         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 465                                   OST_SYNC, 1, &size, NULL);
 466         if (!request)
 467                 RETURN(-ENOMEM);
 468
 469         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 470         memcpy(&body->oa, oa, sizeof(*oa));
 471
 472         /* overload the size and blocks fields in the oa with start/end */
 473         body->oa.o_size = start;
 474         body->oa.o_blocks = end;
 475         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 476
 477         request->rq_replen = lustre_msg_size(1, &size);
 478
 479         rc = ptlrpc_queue_wait(request);
 480         if (rc)
 481                 GOTO(out, rc);
 482
 483         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 484                                   lustre_swab_ost_body);
 485         if (body == NULL) {
 486                 CERROR ("can't unpack ost_body\n");
 487                 GOTO (out, rc = -EPROTO);
 488         }
 489
 490         memcpy(oa, &body->oa, sizeof(*oa));
 491
 492         EXIT;
 493  out:
 494         ptlrpc_req_finished(request);
 495         return rc;
 496 }
 497
 498 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
 499                        struct lov_stripe_md *ea, struct obd_trans_info *oti)
 500 {
 501         struct ptlrpc_request *request;
 502         struct ost_body *body;
 503         int rc, size = sizeof(*body);
 504         ENTRY;
 505
 506         if (!oa) {
 507                 CERROR("oa NULL\n");
 508                 RETURN(-EINVAL);
 509         }
 510
 511         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
 512                                   OST_DESTROY, 1, &size, NULL);
 513         if (!request)
 514                 RETURN(-ENOMEM);
 515         request->rq_request_portal = OST_DESTROY_PORTAL;
 516
 517         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 518
 519         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
 520                 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
 521                        sizeof(*oti->oti_logcookies));
 522                 oti->oti_logcookies++;
 523         }
 524
 525         memcpy(&body->oa, oa, sizeof(*oa));
 526         request->rq_replen = lustre_msg_size(1, &size);
 527
 528         if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
 529                 ptlrpcd_add_req(request);
 530                 rc = 0;
 531         } else {
 532                 rc = ptlrpc_queue_wait(request);
 533
 534                 if (rc == -ENOENT)
 535                         rc = 0;
 536
 537                 if (rc) {
 538                         ptlrpc_req_finished(request);
 539                         RETURN(rc);
 540                 }
 541
 542                 body = lustre_swab_repbuf(request, 0, sizeof(*body),
 543                                           lustre_swab_ost_body);
 544                 if (body == NULL) {
 545                         CERROR ("Can't unpack body\n");
 546                         ptlrpc_req_finished(request);
 547                         RETURN(-EPROTO);
 548                 }
 549
 550                 memcpy(oa, &body->oa, sizeof(*oa));
 551                 ptlrpc_req_finished(request);
 552         }
 553         RETURN(rc);
 554 }
 555
 556 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 557                                 long writing_bytes)
 558 {
 559         obd_valid bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
 560
 561         LASSERT(!(oa->o_valid & bits));
 562
 563         oa->o_valid |= bits;
 564         spin_lock(&cli->cl_loi_list_lock);
 565         oa->o_dirty = cli->cl_dirty;
 566         oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
 567         oa->o_grant = cli->cl_avail_grant;
 568         oa->o_dropped = cli->cl_lost_grant;
 569         cli->cl_lost_grant = 0;
 570         spin_unlock(&cli->cl_loi_list_lock);
 571         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
 572                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 573 }
 574
 575 /* caller must hold loi_list_lock */
 576 static void osc_consume_write_grant(struct client_obd *cli,
 577                                     struct osc_async_page *oap)
 578 {
 579         cli->cl_dirty += PAGE_SIZE;
 580         cli->cl_avail_grant -= PAGE_SIZE;
 581         oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
 582         CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
 583         LASSERT(cli->cl_avail_grant >= 0);
 584 }
 585
 586 static unsigned long rpcs_in_flight(struct client_obd *cli)
 587 {
 588         return cli->cl_r_in_flight + cli->cl_w_in_flight;
 589 }
 590
 591 /* caller must hold loi_list_lock */
 592 void osc_wake_cache_waiters(struct client_obd *cli)
 593 {
 594         struct list_head *l, *tmp;
 595         struct osc_cache_waiter *ocw;
 596
 597         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 598                 /* if we can't dirty more, we must wait until some is written */
 599                 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
 600                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
 601                                cli->cl_dirty, cli->cl_dirty_max);
 602                         return;
 603                 }
 604
 605                 /* if still dirty cache but no grant wait for pending RPCs that
 606                  * may yet return us some grant before doing sync writes */
 607                 if (cli->cl_w_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
 608                         CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
 609                                cli->cl_w_in_flight);
 610                 }
 611                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
 612                 list_del_init(&ocw->ocw_entry);
 613                 if (cli->cl_avail_grant < PAGE_SIZE) {
 614                         /* no more RPCs in flight to return grant, do sync IO */
 615                         ocw->ocw_rc = -EDQUOT;
 616                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
 617                 } else {
 618                         osc_consume_write_grant(cli, ocw->ocw_oap);
 619                 }
 620
 621                 wake_up(&ocw->ocw_waitq);
 622         }
 623
 624         EXIT;
 625 }
 626
 627 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 628 {
 629         spin_lock(&cli->cl_loi_list_lock);
 630         CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
 631         cli->cl_avail_grant += body->oa.o_grant;
 632         /* waiters are woken in brw_interpret_oap */
 633         spin_unlock(&cli->cl_loi_list_lock);
 634 }
 635
 636 /* We assume that the reason this OSC got a short read is because it read
 637  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
 638  * via the LOV, and it _knows_ it's reading inside the file, it's just that
 639  * this stripe never got written at or beyond this stripe offset yet. */
 640 static void handle_short_read(int nob_read, obd_count page_count,
 641                               struct brw_page *pga)
 642 {
 643         char *ptr;
 644
 645         /* skip bytes read OK */
 646         while (nob_read > 0) {
 647                 LASSERT (page_count > 0);
 648
 649                 if (pga->count > nob_read) {
 650                         /* EOF inside this page */
 651                         ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
 652                         memset(ptr + nob_read, 0, pga->count - nob_read);
 653                         kunmap(pga->pg);
 654                         page_count--;
 655                         pga++;
 656                         break;
 657                 }
 658
 659                 nob_read -= pga->count;
 660                 page_count--;
 661                 pga++;
 662         }
 663
 664         /* zero remaining pages */
 665         while (page_count-- > 0) {
 666                 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
 667                 memset(ptr, 0, pga->count);
 668                 kunmap(pga->pg);
 669                 pga++;
 670         }
 671 }
 672
 673 static int check_write_rcs(struct ptlrpc_request *request,
 674                            int requested_nob, int niocount,
 675                            obd_count page_count, struct brw_page *pga)
 676 {
 677         int *remote_rcs, i;
 678
 679         /* return error if any niobuf was in error */
 680         remote_rcs = lustre_swab_repbuf(request, 1,
 681                                         sizeof(*remote_rcs) * niocount, NULL);
 682         if (remote_rcs == NULL) {
 683                 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
 684                 return(-EPROTO);
 685         }
 686         if (lustre_msg_swabbed(request->rq_repmsg))
 687                 for (i = 0; i < niocount; i++)
 688                         __swab32s((__u32 *)&remote_rcs[i]);
 689
 690         for (i = 0; i < niocount; i++) {
 691                 if (remote_rcs[i] < 0)
 692                         return(remote_rcs[i]);
 693
 694                 if (remote_rcs[i] != 0) {
 695                         CERROR("rc[%d] invalid (%d) req %p\n",
 696                                 i, remote_rcs[i], request);
 697                         return(-EPROTO);
 698                 }
 699         }
 700
 701         if (request->rq_bulk->bd_nob_transferred != requested_nob) {
 702                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
 703                        requested_nob, request->rq_bulk->bd_nob_transferred);
 704                 return(-EPROTO);
 705         }
 706
 707         return (0);
 708 }
 709
 710 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
 711 {
 712         if (p1->flag != p2->flag) {
 713                 unsigned mask = ~OBD_BRW_FROM_GRANT;
 714
 715                 /* warn if we try to combine flags that we don't know to be
 716                  * safe to combine */
 717                 if ((p1->flag & mask) != (p2->flag & mask))
 718                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
 719                                "same brw?\n", p1->flag, p2->flag);
 720                 return 0;
 721         }
 722
 723         return (p1->disk_offset + p1->count == p2->disk_offset);
 724 }
 725
 726 #if CHECKSUM_BULK
 727 static obd_count cksum_pages(int nob, obd_count page_count,
 728                              struct brw_page *pga)
 729 {
 730         obd_count cksum = 0;
 731         char *ptr;
 732
 733         while (nob > 0) {
 734                 LASSERT (page_count > 0);
 735
 736                 ptr = kmap(pga->pg);
 737                 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
 738                              pga->count > nob ? nob : pga->count);
 739                 kunmap(pga->pg);
 740
 741                 nob -= pga->count;
 742                 page_count--;
 743                 pga++;
 744         }
 745
 746         return (cksum);
 747 }
 748 #endif
 749
 750 #define osc_encrypt_page(page, off, count)  \
 751         osc_crypt_page(page, off, count, ENCRYPT_DATA)
 752 #define osc_decrypt_page(page, off, count)  \
 753         osc_crypt_page(page, off, count, DECRYPT_DATA)
 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_decrypt_pages(struct brw_page *pga, int page_count)
 772 {
 773         int i =0;
 774         ENTRY;
 775
 776         for (i = 0; i < page_count; i++) {
 777                 struct brw_page *pg = &pga[i];
 778                 osc_decrypt_page(pg->pg, pg->page_offset, pg->count);
 779         }
 780         RETURN(0);
 781 }
 782
 783 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
 784                                 struct lov_stripe_md *lsm, obd_count page_count,
 785                                 struct brw_page *pga, int *requested_nobp,
 786                                 int *niocountp, struct ptlrpc_request **reqp)
 787 {
 788         struct ptlrpc_request   *req;
 789         struct ptlrpc_bulk_desc *desc;
 790         struct client_obd       *cli = &imp->imp_obd->u.cli;
 791         struct ost_body         *body;
 792         struct lustre_id        *raw_id = obdo_id(oa);
 793         struct obd_capa         *ocapa = NULL;
 794         struct lustre_capa      *capa = NULL;
 795         struct obd_ioobj        *ioobj;
 796         struct niobuf_remote    *niobuf;
 797         int                      niocount;
 798         int                      size[4];
 799         int                      i, bufcnt = 0;
 800         int                      requested_nob;
 801         int                      opc;
 802         int                      capa_op;
 803         int                      rc;
 804
 805         opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
 806
 807         for (niocount = i = 1; i < page_count; i++)
 808                 if (!can_merge_pages(&pga[i - 1], &pga[i]))
 809                         niocount++;
 810
 811         /* TODO: this could be optimized: thie capability can be
 812          * found from ll_inode_info->lli_capas. */
 813         /* partial write might cause read, both CAPA_READ and CAPA_WRITE
 814          * capability could be used here */
 815         capa_op = (opc == OST_WRITE) ? CAPA_WRITE : CAPA_READ | CAPA_WRITE;
 816         ocapa = capa_get(oa->o_fsuid, capa_op, raw_id->li_fid.lf_group,
 817                          id_ino(raw_id), id_gen(raw_id), CLIENT_CAPA);
 818
 819         size[bufcnt++] = sizeof(*body);
 820         size[bufcnt++] = sizeof(*ioobj);
 821         if (ocapa)
 822                 size[bufcnt++] = sizeof(*capa);
 823         size[bufcnt++] = niocount * sizeof(*niobuf);
 824
 825         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, opc, bufcnt, size, NULL);
 826         if (req == NULL)
 827                 return (-ENOMEM);
 828
 829         if (opc == OST_WRITE)
 830                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 831                                              BULK_GET_SOURCE, OST_BULK_PORTAL);
 832         else
 833                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 834                                              BULK_PUT_SINK, OST_BULK_PORTAL);
 835         if (desc == NULL)
 836                 GOTO(out, rc = -ENOMEM);
 837         /* NB request now owns desc and will free it when it gets freed */
 838
 839         bufcnt = 0;
 840         body = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*body));
 841         memcpy(&body->oa, oa, sizeof(*oa));
 842         ioobj = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*ioobj));
 843         if (ocapa) {
 844                 capa = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*capa));
 845                 capa_dup(capa, ocapa);
 846                 body->oa.o_valid |= OBD_MD_CAPA;
 847         }
 848         niobuf = lustre_msg_buf(req->rq_reqmsg, bufcnt++,
 849                                 niocount * sizeof(*niobuf));
 850
 851         obdo_to_ioobj(oa, ioobj);
 852         ioobj->ioo_bufcnt = niocount;
 853
 854         LASSERT (page_count > 0);
 855
 856         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
 857                 struct brw_page *pg = &pga[i];
 858                 struct brw_page *pg_prev = pg - 1;
 859
 860                 LASSERT(pg->count > 0);
 861                 LASSERTF((pg->page_offset & ~PAGE_MASK)+ pg->count <= PAGE_SIZE,
 862                          "i: %d pg: %p pg_off: "LPU64", count: %u\n", i, pg,
 863                          pg->page_offset, pg->count);
 864                 LASSERTF(i == 0 || pg->disk_offset > pg_prev->disk_offset,
 865                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
 866                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
 867                          i, page_count,
 868                          pg->pg, pg->pg->private, pg->pg->index, pg->disk_offset,
 869                          pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
 870                          pg_prev->disk_offset);
 871
 872                 if (opc == OST_WRITE) {
 873                         rc = osc_encrypt_page(pg->pg, pg->page_offset, pg->count);
 874                         if (rc)
 875                                 GOTO(out, rc);
 876                 }
 877
 878                 ptlrpc_prep_bulk_page(desc, pg->pg,
 879                                       pg->page_offset & ~PAGE_MASK, pg->count);
 880                 requested_nob += pg->count;
 881
 882                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
 883                         niobuf--;
 884                         niobuf->len += pg->count;
 885                 } else {
 886                         niobuf->offset = pg->disk_offset;
 887                         niobuf->len    = pg->count;
 888                         niobuf->flags  = pg->flag;
 889                 }
 890         }
 891
 892         LASSERT((void *)(niobuf - niocount) ==
 893                 lustre_msg_buf(req->rq_reqmsg, bufcnt - 1,
 894                                niocount * sizeof(*niobuf)));
 895         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
 896
 897         /* size[0] still sizeof (*body) */
 898         if (opc == OST_WRITE) {
 899 #if CHECKSUM_BULK
 900                 body->oa.o_valid |= OBD_MD_FLCKSUM;
 901                 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
 902 #endif
 903                 /* 1 RC per niobuf */
 904                 size[1] = sizeof(__u32) * niocount;
 905                 req->rq_replen = lustre_msg_size(2, size);
 906         } else {
 907                 /* 1 RC for the whole I/O */
 908                 req->rq_replen = lustre_msg_size(1, size);
 909         }
 910
 911         *niocountp = niocount;
 912         *requested_nobp = requested_nob;
 913         *reqp = req;
 914         return (0);
 915
 916  out:
 917         ptlrpc_req_finished (req);
 918         return (rc);
 919 }
 920
 921 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
 922                                 int requested_nob, int niocount,
 923                                 obd_count page_count, struct brw_page *pga,
 924                                 int rc)
 925 {
 926         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 927         struct ost_body *body;
 928         ENTRY;
 929
 930         if (rc < 0)
 931                 RETURN(rc);
 932
 933         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 934         if (body == NULL) {
 935                 CERROR ("Can't unpack body\n");
 936                 RETURN(-EPROTO);
 937         }
 938
 939         osc_update_grant(cli, body);
 940         memcpy(oa, &body->oa, sizeof(*oa));
 941
 942         if (req->rq_reqmsg->opc == OST_WRITE) {
 943                 if (rc > 0) {
 944                         CERROR ("Unexpected +ve rc %d\n", rc);
 945                         RETURN(-EPROTO);
 946                 }
 947                 LASSERT (req->rq_bulk->bd_nob == requested_nob);
 948                 osc_decrypt_pages(pga, page_count);
 949                 RETURN(check_write_rcs(req, requested_nob, niocount,
 950                                        page_count, pga));
 951         }
 952
 953         if (rc > requested_nob) {
 954                 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
 955                 RETURN(-EPROTO);
 956         }
 957
 958         if (rc != req->rq_bulk->bd_nob_transferred) {
 959                 CERROR ("Unexpected rc %d (%d transferred)\n",
 960                         rc, req->rq_bulk->bd_nob_transferred);
 961                 return (-EPROTO);
 962         }
 963
 964         if (rc < requested_nob)
 965                 handle_short_read(rc, page_count, pga);
 966
 967 #if CHECKSUM_BULK
 968         if (oa->o_valid & OBD_MD_FLCKSUM) {
 969                 const struct ptlrpc_peer *peer =
 970                         &req->rq_import->imp_connection->c_peer;
 971                 static int cksum_counter;
 972                 obd_count server_cksum = oa->o_cksum;
 973                 obd_count cksum = cksum_pages(rc, page_count, pga);
 974                 char str[PTL_NALFMT_SIZE];
 975
 976                 ptlrpc_peernid2str(peer, str);
 977
 978                 cksum_counter++;
 979                 if (server_cksum != cksum) {
 980                         CERROR("Bad checksum: server %x, client %x, server NID "
 981                                LPX64" (%s)\n", server_cksum, cksum,
 982                                peer->peer_id.nid, str);
 983                         cksum_counter = 0;
 984                         oa->o_cksum = cksum;
 985                 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
 986                         CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
 987                               cksum_counter, peer->peer_id.nid, str, cksum);
 988                 }
 989         } else {
 990                 static int cksum_missed;
 991
 992                 cksum_missed++;
 993                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
 994                         CERROR("Request checksum %u from "LPX64", no reply\n",
 995                                cksum_missed,
 996                                req->rq_import->imp_connection->c_peer.peer_id.nid);
 997         }
 998 #endif
 999         osc_decrypt_pages(pga, page_count);
1000         RETURN(0);
1001 }
1002
1003 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
1004                             struct lov_stripe_md *lsm,
1005                             obd_count page_count, struct brw_page *pga)
1006 {
1007         int                    requested_nob;
1008         int                    niocount;
1009         struct ptlrpc_request *request;
1010         int                    rc;
1011         ENTRY;
1012
1013 restart_bulk:
1014         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1015                                   page_count, pga, &requested_nob, &niocount,
1016                                   &request);
1017         if (rc != 0)
1018                 return (rc);
1019
1020         rc = ptlrpc_queue_wait(request);
1021
1022         if (rc == -ETIMEDOUT && request->rq_resend) {
1023                 DEBUG_REQ(D_HA, request,  "BULK TIMEOUT");
1024                 ptlrpc_req_finished(request);
1025                 goto restart_bulk;
1026         }
1027
1028         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
1029                                   page_count, pga, rc);
1030
1031         ptlrpc_req_finished(request);
1032         RETURN (rc);
1033 }
1034
1035 static int brw_interpret(struct ptlrpc_request *request,
1036                          struct osc_brw_async_args *aa, int rc)
1037 {
1038         struct obdo *oa      = aa->aa_oa;
1039         int requested_nob    = aa->aa_requested_nob;
1040         int niocount         = aa->aa_nio_count;
1041         obd_count page_count = aa->aa_page_count;
1042         struct brw_page *pga = aa->aa_pga;
1043         ENTRY;
1044
1045         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
1046                                   page_count, pga, rc);
1047         RETURN (rc);
1048 }
1049
1050 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1051                           struct lov_stripe_md *lsm, obd_count page_count,
1052                           struct brw_page *pga, struct ptlrpc_request_set *set)
1053 {
1054         struct ptlrpc_request     *request;
1055         int                        requested_nob;
1056         int                        nio_count;
1057         struct osc_brw_async_args *aa;
1058         int                        rc;
1059         ENTRY;
1060
1061         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1062                                   page_count, pga, &requested_nob, &nio_count,
1063                                   &request);
1064         if (rc == 0) {
1065                 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1066                 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1067                 aa->aa_oa = oa;
1068                 aa->aa_requested_nob = requested_nob;
1069                 aa->aa_nio_count = nio_count;
1070                 aa->aa_page_count = page_count;
1071                 aa->aa_pga = pga;
1072
1073                 request->rq_interpret_reply = brw_interpret;
1074                 ptlrpc_set_add_req(set, request);
1075         }
1076         RETURN (rc);
1077 }
1078
1079 #ifndef min_t
1080 #define min_t(type,x,y) \
1081         ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
1082 #endif
1083
1084 /*
1085  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1086  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1087  * fine for our small page arrays and doesn't require allocation.  its an
1088  * insertion sort that swaps elements that are strides apart, shrinking the
1089  * stride down until its '1' and the array is sorted.
1090  */
1091 static void sort_brw_pages(struct brw_page *array, int num)
1092 {
1093         int stride, i, j;
1094         struct brw_page tmp;
1095
1096         if (num == 1)
1097                 return;
1098         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1099                 ;
1100
1101         do {
1102                 stride /= 3;
1103                 for (i = stride ; i < num ; i++) {
1104                         tmp = array[i];
1105                         j = i;
1106                         while (j >= stride && array[j - stride].disk_offset >
1107                                 tmp.disk_offset) {
1108                                 array[j] = array[j - stride];
1109                                 j -= stride;
1110                         }
1111                         array[j] = tmp;
1112                 }
1113         } while (stride > 1);
1114 }
1115
1116 /* make sure we the regions we're passing to elan don't violate its '4
1117  * fragments' constraint.  portal headers are a fragment, all full
1118  * PAGE_SIZE long pages count as 1 fragment, and each partial page
1119  * counts as a fragment.  I think.  see bug 934. */
1120 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1121 {
1122         int frags_left = 3;
1123         int saw_whole_frag = 0;
1124         int i;
1125
1126         for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1127                 if (pg->count == PAGE_SIZE) {
1128                         if (!saw_whole_frag) {
1129                                 saw_whole_frag = 1;
1130                                 frags_left--;
1131                         }
1132                 } else {
1133                         frags_left--;
1134                 }
1135         }
1136         return i;
1137 }
1138
1139 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1140                    struct lov_stripe_md *lsm, obd_count page_count,
1141                    struct brw_page *pga, struct obd_trans_info *oti)
1142 {
1143         ENTRY;
1144
1145         if (cmd == OBD_BRW_CHECK) {
1146                 /* The caller just wants to know if there's a chance that this
1147                  * I/O can succeed */
1148                 struct obd_import *imp = class_exp2cliimp(exp);
1149
1150                 if (imp == NULL || imp->imp_invalid)
1151                         RETURN(-EIO);
1152                 RETURN(0);
1153         }
1154
1155         while (page_count) {
1156                 obd_count pages_per_brw;
1157                 int rc;
1158
1159                 if (page_count > PTLRPC_MAX_BRW_PAGES)
1160                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1161                 else
1162                         pages_per_brw = page_count;
1163
1164                 sort_brw_pages(pga, pages_per_brw);
1165                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1166
1167                 rc = osc_brw_internal(cmd, exp, oa, lsm, pages_per_brw, pga);
1168
1169                 if (rc != 0)
1170                         RETURN(rc);
1171
1172                 page_count -= pages_per_brw;
1173                 pga += pages_per_brw;
1174         }
1175         RETURN(0);
1176 }
1177
1178 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1179                          struct lov_stripe_md *lsm, obd_count page_count,
1180                          struct brw_page *pga, struct ptlrpc_request_set *set,
1181                          struct obd_trans_info *oti)
1182 {
1183         ENTRY;
1184
1185         if (cmd == OBD_BRW_CHECK) {
1186                 /* The caller just wants to know if there's a chance that this
1187                  * I/O can succeed */
1188                 struct obd_import *imp = class_exp2cliimp(exp);
1189
1190                 if (imp == NULL || imp->imp_invalid)
1191                         RETURN(-EIO);
1192                 RETURN(0);
1193         }
1194
1195         while (page_count) {
1196                 obd_count pages_per_brw;
1197                 int rc;
1198
1199                 if (page_count > PTLRPC_MAX_BRW_PAGES)
1200                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1201                 else
1202                         pages_per_brw = page_count;
1203
1204                 sort_brw_pages(pga, pages_per_brw);
1205                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1206
1207                 rc = async_internal(cmd, exp, oa, lsm, pages_per_brw, pga, set);
1208
1209                 if (rc != 0)
1210                         RETURN(rc);
1211
1212                 page_count -= pages_per_brw;
1213                 pga += pages_per_brw;
1214         }
1215         RETURN(0);
1216 }
1217
1218 static void osc_check_rpcs(struct client_obd *cli);
1219 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1220                            int sent);
1221 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1222 static void lop_update_pending(struct client_obd *cli,
1223                                struct loi_oap_pages *lop, int cmd, int delta);
1224
1225 /* this is called when a sync waiter receives an interruption.  Its job is to
1226  * get the caller woken as soon as possible.  If its page hasn't been put in an
1227  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
1228  * desiring interruption which will forcefully complete the rpc once the rpc
1229  * has timed out */
1230 static void osc_occ_interrupted(struct oig_callback_context *occ)
1231 {
1232         struct osc_async_page *oap;
1233         struct loi_oap_pages *lop;
1234         struct lov_oinfo *loi;
1235         ENTRY;
1236
1237         /* XXX member_of() */
1238         oap = list_entry(occ, struct osc_async_page, oap_occ);
1239
1240         spin_lock(&oap->oap_cli->cl_loi_list_lock);
1241
1242         oap->oap_interrupted = 1;
1243
1244         /* ok, it's been put in an rpc. */
1245         if (oap->oap_request != NULL) {
1246                 ptlrpc_mark_interrupted(oap->oap_request);
1247                 ptlrpcd_wake(oap->oap_request);
1248                 GOTO(unlock, 0);
1249         }
1250
1251         /* we don't get interruption callbacks until osc_trigger_sync_io()
1252          * has been called and put the sync oaps in the pending/urgent lists.*/
1253         if (!list_empty(&oap->oap_pending_item)) {
1254                 list_del_init(&oap->oap_pending_item);
1255                 if (oap->oap_async_flags & ASYNC_URGENT)
1256                         list_del_init(&oap->oap_urgent_item);
1257
1258                 loi = oap->oap_loi;
1259                 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1260                         &loi->loi_write_lop : &loi->loi_read_lop;
1261                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1262                 loi_list_maint(oap->oap_cli, oap->oap_loi);
1263
1264                 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1265                 oap->oap_oig = NULL;
1266         }
1267
1268 unlock:
1269         spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1270 }
1271
1272 /* this must be called holding the loi list lock to give coverage to exit_cache,
1273  * async_flag maintenance, and oap_request */
1274 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1275                               struct osc_async_page *oap, int sent, int rc)
1276 {
1277         osc_exit_cache(cli, oap, sent);
1278         oap->oap_async_flags = 0;
1279         oap->oap_interrupted = 0;
1280
1281         if (oap->oap_request != NULL) {
1282                 ptlrpc_req_finished(oap->oap_request);
1283                 oap->oap_request = NULL;
1284         }
1285
1286         if (rc == 0 && oa != NULL)
1287                 oap->oap_loi->loi_blocks = oa->o_blocks;
1288
1289         if (oap->oap_oig) {
1290                 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1291                 oap->oap_oig = NULL;
1292                 EXIT;
1293                 return;
1294         }
1295
1296         oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1297                                            oa, rc);
1298 }
1299
1300 static int brw_interpret_oap(struct ptlrpc_request *request,
1301                              struct osc_brw_async_args *aa, int rc)
1302 {
1303         struct osc_async_page *oap;
1304         struct client_obd *cli;
1305         struct list_head *pos, *n;
1306         struct timeval now;
1307         ENTRY;
1308
1309         do_gettimeofday(&now);
1310         rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1311                                   aa->aa_nio_count, aa->aa_page_count,
1312                                   aa->aa_pga, rc);
1313
1314         CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1315
1316         cli = aa->aa_cli;
1317
1318         /* in failout recovery we ignore writeback failure and want
1319          * to just tell llite to unlock the page and continue */
1320         if (request->rq_reqmsg->opc == OST_WRITE &&
1321             (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1322                 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1323                        cli->cl_import,
1324                        cli->cl_import ? cli->cl_import->imp_invalid : -1);
1325                 rc = 0;
1326         }
1327
1328         spin_lock(&cli->cl_loi_list_lock);
1329
1330         if (request->rq_reqmsg->opc == OST_WRITE)
1331                 lprocfs_stime_record(&cli->cl_write_stime, &now,
1332                                      &request->rq_rpcd_start);
1333         else
1334                 lprocfs_stime_record(&cli->cl_read_stime, &now,
1335                                      &request->rq_rpcd_start);
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         LASSERT(lop->lop_pending.next != LP_POISON);
1480
1481         /* first we find the pages we're allowed to work with */
1482         list_for_each_safe(pos, tmp, &lop->lop_pending) {
1483                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1484                 ops = oap->oap_caller_ops;
1485
1486                 LASSERT(oap->oap_magic == OAP_MAGIC);
1487
1488                 /* in llite being 'ready' equates to the page being locked
1489                  * until completion unlocks it.  commit_write submits a page
1490                  * as not ready because its unlock will happen unconditionally
1491                  * as the call returns.  if we race with commit_write giving
1492                  * us that page we dont' want to create a hole in the page
1493                  * stream, so we stop and leave the rpc to be fired by
1494                  * another dirtier or kupdated interval (the not ready page
1495                  * will still be on the dirty list).  we could call in
1496                  * at the end of ll_file_write to process the queue again. */
1497                 if (!(oap->oap_async_flags & ASYNC_READY)) {
1498                         int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1499                         if (rc < 0)
1500                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
1501                                                 "instead of ready\n", oap,
1502                                                 oap->oap_page, rc);
1503                         switch (rc) {
1504                         case -EAGAIN:
1505                                 /* llite is telling us that the page is still
1506                                  * in commit_write and that we should try
1507                                  * and put it in an rpc again later.  we
1508                                  * break out of the loop so we don't create
1509                                  * a hole in the sequence of pages in the rpc
1510                                  * stream.*/
1511                                 pos = NULL;
1512                                 break;
1513                         case -EINTR:
1514                                 /* the io isn't needed.. tell the checks
1515                                  * below to complete the rpc with EINTR */
1516                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1517                                 oap->oap_count = -EINTR;
1518                                 break;
1519                         case 0:
1520                                 oap->oap_async_flags |= ASYNC_READY;
1521                                 break;
1522                         default:
1523                                 LASSERTF(0, "oap %p page %p returned %d "
1524                                             "from make_ready\n", oap,
1525                                             oap->oap_page, rc);
1526                                 break;
1527                         }
1528                 }
1529                 if (pos == NULL)
1530                         break;
1531                 /*
1532                  * Page submitted for IO has to be locked. Either by
1533                  * ->ap_make_ready() or by higher layers.
1534                  *
1535                  * XXX nikita: this assertion should be adjusted when lustre
1536                  * starts using PG_writeback for pages being written out.
1537                  */
1538                 LASSERT(PageLocked(oap->oap_page));
1539
1540                 /* take the page out of our book-keeping */
1541                 list_del_init(&oap->oap_pending_item);
1542                 lop_update_pending(cli, lop, cmd, -1);
1543                 list_del_init(&oap->oap_urgent_item);
1544
1545                 /* ask the caller for the size of the io as the rpc leaves. */
1546                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1547                         oap->oap_count =
1548                                 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1549                 if (oap->oap_count <= 0) {
1550                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1551                                oap->oap_count);
1552                         osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1553                         continue;
1554                 }
1555
1556                 /* now put the page back in our accounting */
1557                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1558                 if (++page_count >= cli->cl_max_pages_per_rpc)
1559                         break;
1560         }
1561
1562         osc_wake_cache_waiters(cli);
1563
1564         if (page_count == 0)
1565                 RETURN(0);
1566
1567         loi_list_maint(cli, loi);
1568         spin_unlock(&cli->cl_loi_list_lock);
1569
1570         request = osc_build_req(cli, &rpc_list, page_count, cmd);
1571         if (IS_ERR(request)) {
1572                 /* this should happen rarely and is pretty bad, it makes the
1573                  * pending list not follow the dirty order */
1574                 spin_lock(&cli->cl_loi_list_lock);
1575                 list_for_each_safe(pos, tmp, &rpc_list) {
1576                         oap = list_entry(pos, struct osc_async_page,
1577                                          oap_rpc_item);
1578                         list_del_init(&oap->oap_rpc_item);
1579
1580                         /* queued sync pages can be torn down while the pages
1581                          * were between the pending list and the rpc */
1582                         if (oap->oap_interrupted) {
1583                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1584                                 osc_ap_completion(cli, NULL, oap, 0,
1585                                                   oap->oap_count);
1586                                 continue;
1587                         }
1588
1589                         /* put the page back in the loi/lop lists */
1590                         list_add_tail(&oap->oap_pending_item,
1591                                       &lop->lop_pending);
1592                         lop_update_pending(cli, lop, cmd, 1);
1593                         if (oap->oap_async_flags & ASYNC_URGENT)
1594                                 list_add(&oap->oap_urgent_item,
1595                                          &lop->lop_urgent);
1596                 }
1597                 loi_list_maint(cli, loi);
1598                 RETURN(PTR_ERR(request));
1599         }
1600
1601         LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1602         aa = (struct osc_brw_async_args *)&request->rq_async_args;
1603         INIT_LIST_HEAD(&aa->aa_oaps);
1604         list_splice(&rpc_list, &aa->aa_oaps);
1605         INIT_LIST_HEAD(&rpc_list);
1606
1607 #ifdef __KERNEL__
1608         if (cmd == OBD_BRW_READ) {
1609                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1610                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1611         } else {
1612                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1613                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1614                                  cli->cl_w_in_flight);
1615         }
1616 #endif
1617
1618         spin_lock(&cli->cl_loi_list_lock);
1619
1620         /* collect write gaps and sum of them */
1621         if (cmd == OBD_BRW_WRITE && cli->cl_w_in_flight == 0) {
1622                 struct timeval now;
1623                 long diff;
1624
1625                 do_gettimeofday(&now);
1626
1627                 if (cli->cl_last_write_time.tv_sec) {
1628                         diff = timeval_sub(&now, &cli->cl_last_write_time);
1629                         if (diff < CLI_ODD_WRITE_GAP) {
1630                                 cli->cl_write_gap_sum += diff;
1631                                 cli->cl_write_gaps++;
1632                         }
1633                 }
1634         }
1635
1636         if (cmd == OBD_BRW_READ) {
1637                 cli->cl_r_in_flight++;
1638                 cli->cl_read_num++;
1639         } else {
1640                 cli->cl_w_in_flight++;
1641                 cli->cl_write_num++;
1642         }
1643
1644         /* queued sync pages can be torn down while the pages
1645          * were between the pending list and the rpc */
1646         list_for_each(pos, &aa->aa_oaps) {
1647                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1648                 if (oap->oap_interrupted) {
1649                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1650                                oap, request);
1651                         ptlrpc_mark_interrupted(request);
1652                         break;
1653                 }
1654         }
1655
1656         CDEBUG(D_INODE, "req %p: %d pages, aa %p.  now %dr/%dw in flight\n",
1657                         request, page_count, aa, cli->cl_r_in_flight,
1658                         cli->cl_w_in_flight);
1659
1660         oap->oap_request = ptlrpc_request_addref(request);
1661         request->rq_interpret_reply = brw_interpret_oap;
1662
1663         ptlrpcd_add_req(request);
1664         RETURN(1);
1665 }
1666
1667 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1668                          int cmd)
1669 {
1670         int optimal;
1671         ENTRY;
1672
1673         if (lop->lop_num_pending == 0)
1674                 RETURN(0);
1675
1676         /* if we have an invalid import we want to drain the queued pages
1677          * by forcing them through rpcs that immediately fail and complete
1678          * the pages.  recovery relies on this to empty the queued pages
1679          * before canceling the locks and evicting down the llite pages */
1680         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1681                 RETURN(1);
1682
1683         /* stream rpcs in queue order as long as as there is an urgent page
1684          * queued.  this is our cheap solution for good batching in the case
1685          * where writepage marks some random page in the middle of the file as
1686          * urgent because of, say, memory pressure */
1687         if (!list_empty(&lop->lop_urgent))
1688                 RETURN(1);
1689
1690         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1691         optimal = cli->cl_max_pages_per_rpc;
1692         if (cmd == OBD_BRW_WRITE) {
1693                 /* trigger a write rpc stream as long as there are dirtiers
1694                  * waiting for space.  as they're waiting, they're not going to
1695                  * create more pages to coallesce with what's waiting.. */
1696                 if (!list_empty(&cli->cl_cache_waiters))
1697                         RETURN(1);
1698
1699                 /* *2 to avoid triggering rpcs that would want to include pages
1700                  * that are being queued but which can't be made ready until
1701                  * the queuer finishes with the page. this is a wart for
1702                  * llite::commit_write() */
1703                 optimal += 16;
1704         }
1705         if (lop->lop_num_pending >= optimal)
1706                 RETURN(1);
1707
1708         RETURN(0);
1709 }
1710
1711 static void on_list(struct list_head *item, struct list_head *list,
1712                     int should_be_on)
1713 {
1714         if (list_empty(item) && should_be_on)
1715                 list_add_tail(item, list);
1716         else if (!list_empty(item) && !should_be_on)
1717                 list_del_init(item);
1718 }
1719
1720 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1721  * can find pages to build into rpcs quickly */
1722 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1723 {
1724         on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1725                 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1726                 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1727
1728         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1729                 loi->loi_write_lop.lop_num_pending);
1730
1731         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1732                 loi->loi_read_lop.lop_num_pending);
1733 }
1734
1735 #define LOI_DEBUG(LOI, STR, args...)                                     \
1736         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
1737                !list_empty(&(LOI)->loi_cli_item),                        \
1738                (LOI)->loi_write_lop.lop_num_pending,                     \
1739                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
1740                (LOI)->loi_read_lop.lop_num_pending,                      \
1741                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
1742                args)                                                     \
1743
1744 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1745 {
1746         ENTRY;
1747         /* first return all objects which we already know to have
1748          * pages ready to be stuffed into rpcs */
1749         if (!list_empty(&cli->cl_loi_ready_list))
1750                 RETURN(list_entry(cli->cl_loi_ready_list.next,
1751                                   struct lov_oinfo, loi_cli_item));
1752
1753         /* then if we have cache waiters, return all objects with queued
1754          * writes.  This is especially important when many small files
1755          * have filled up the cache and not been fired into rpcs because
1756          * they don't pass the nr_pending/object threshhold */
1757         if (!list_empty(&cli->cl_cache_waiters) &&
1758             !list_empty(&cli->cl_loi_write_list))
1759                 RETURN(list_entry(cli->cl_loi_write_list.next,
1760                                   struct lov_oinfo, loi_write_item));
1761
1762         /* then return all queued objects when we have an invalid import
1763          * so that they get flushed */
1764         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1765                 if (!list_empty(&cli->cl_loi_write_list))
1766                         RETURN(list_entry(cli->cl_loi_write_list.next,
1767                                           struct lov_oinfo, loi_write_item));
1768                 if (!list_empty(&cli->cl_loi_read_list))
1769                         RETURN(list_entry(cli->cl_loi_read_list.next,
1770                                           struct lov_oinfo, loi_read_item));
1771         }
1772         RETURN(NULL);
1773 }
1774
1775 /* called with the loi list lock held */
1776 static void osc_check_rpcs(struct client_obd *cli)
1777 {
1778         struct lov_oinfo *loi;
1779         int rc = 0, race_counter = 0;
1780         ENTRY;
1781
1782         while ((loi = osc_next_loi(cli)) != NULL) {
1783                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
1784                 LASSERT(loi->loi_ost_idx != LL_POISON);
1785
1786                 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
1787                         break;
1788
1789                 /* attempt some read/write balancing by alternating between
1790                  * reads and writes in an object.  The makes_rpc checks here
1791                  * would be redundant if we were getting read/write work items
1792                  * instead of objects.  we don't want send_oap_rpc to drain a
1793                  * partial read pending queue when we're given this object to
1794                  * do io on writes while there are cache waiters */
1795                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1796                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1797                                               &loi->loi_write_lop);
1798                         if (rc < 0)
1799                                 break;
1800                         if (rc > 0)
1801                                 race_counter = 0;
1802                         else
1803                                 race_counter++;
1804                 }
1805                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1806                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1807                                               &loi->loi_read_lop);
1808                         if (rc < 0)
1809                                 break;
1810                         if (rc > 0)
1811                                 race_counter = 0;
1812                         else
1813                                 race_counter++;
1814                 }
1815
1816                 /* attempt some inter-object balancing by issueing rpcs
1817                  * for each object in turn */
1818                 if (!list_empty(&loi->loi_cli_item))
1819                         list_del_init(&loi->loi_cli_item);
1820                 if (!list_empty(&loi->loi_write_item))
1821                         list_del_init(&loi->loi_write_item);
1822                 if (!list_empty(&loi->loi_read_item))
1823                         list_del_init(&loi->loi_read_item);
1824
1825                 loi_list_maint(cli, loi);
1826
1827                 /* send_oap_rpc fails with 0 when make_ready tells it to
1828                  * back off.  llite's make_ready does this when it tries
1829                  * to lock a page queued for write that is already locked.
1830                  * we want to try sending rpcs from many objects, but we
1831                  * don't want to spin failing with 0.  */
1832                 if (race_counter == 10)
1833                         break;
1834         }
1835         EXIT;
1836 }
1837
1838 /* we're trying to queue a page in the osc so we're subject to the
1839  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1840  * If the osc's queued pages are already at that limit, then we want to sleep
1841  * until there is space in the osc's queue for us.  We also may be waiting for
1842  * write credits from the OST if there are RPCs in flight that may return some
1843  * before we fall back to sync writes.
1844  *
1845  * We need this know our allocation was granted in the presence of signals */
1846 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1847 {
1848         int rc;
1849         ENTRY;
1850         spin_lock(&cli->cl_loi_list_lock);
1851         rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
1852         spin_unlock(&cli->cl_loi_list_lock);
1853         RETURN(rc);
1854 };
1855
1856 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1857  * grant or cache space. */
1858 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1859                            struct osc_async_page *oap)
1860 {
1861         struct osc_cache_waiter ocw;
1862         struct l_wait_info lwi = { 0 };
1863         struct timeval start, stop;
1864         ENTRY;
1865
1866         CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1867                cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1868                cli->cl_avail_grant);
1869
1870         if (cli->cl_dirty_max < PAGE_SIZE)
1871                 GOTO(out, -EDQUOT);
1872
1873         if (~0ul - cli->cl_dirty_sum <= cli->cl_dirty) {
1874                 cli->cl_dirty_av = (cli->cl_dirty_av +
1875                                     (cli->cl_dirty_sum / cli->cl_dirty_num)) / 2;
1876                 cli->cl_dirty_num = 0;
1877                 cli->cl_dirty_sum = 0;
1878         } else {
1879                 if (cli->cl_dirty_num)
1880                         cli->cl_dirty_av = (cli->cl_dirty_sum / cli->cl_dirty_num);
1881         }
1882
1883         cli->cl_dirty_num++;
1884         cli->cl_dirty_sum += cli->cl_dirty;
1885
1886         if (cli->cl_dirty > cli->cl_dirty_dmax)
1887                 cli->cl_dirty_dmax = cli->cl_dirty;
1888         if (cli->cl_dirty < cli->cl_dirty_dmin || !cli->cl_dirty_dmin)
1889                 cli->cl_dirty_dmin = cli->cl_dirty;
1890
1891         /* Hopefully normal case - cache space and write credits available */
1892         if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1893             cli->cl_avail_grant >= PAGE_SIZE) {
1894                 /* account for ourselves */
1895                 osc_consume_write_grant(cli, oap);
1896                 RETURN(0);
1897         }
1898
1899         /* Make sure that there are write rpcs in flight to wait for.  This
1900          * is a little silly as this object may not have any pending but
1901          * other objects sure might. */
1902         if (cli->cl_w_in_flight) {
1903                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1904                 init_waitqueue_head(&ocw.ocw_waitq);
1905                 ocw.ocw_oap = oap;
1906                 ocw.ocw_rc = 0;
1907
1908                 loi_list_maint(cli, loi);
1909                 osc_check_rpcs(cli);
1910                 spin_unlock(&cli->cl_loi_list_lock);
1911
1912                 CDEBUG(0, "sleeping for cache space\n");
1913                 do_gettimeofday(&start);
1914                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1915                 do_gettimeofday(&stop);
1916
1917                 cli->cl_cache_wait_num++;
1918                 cli->cl_cache_wait_sum += timeval_sub(&stop, &start);
1919
1920                 spin_lock(&cli->cl_loi_list_lock);
1921                 lprocfs_stime_record(&cli->cl_enter_stime, &stop, &start);
1922                 if (!list_empty(&ocw.ocw_entry)) {
1923                         list_del(&ocw.ocw_entry);
1924                         RETURN(-EINTR);
1925                 }
1926                 RETURN(ocw.ocw_rc);
1927         }
1928
1929         EXIT;
1930 out:
1931         cli->cl_sync_rpcs++;
1932         return -EDQUOT;
1933 }
1934
1935 /* the companion to enter_cache, called when an oap is no longer part of the
1936  * dirty accounting.. so writeback completes or truncate happens before writing
1937  * starts.  must be called with the loi lock held. */
1938 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1939                            int sent)
1940 {
1941         ENTRY;
1942
1943         if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1944                 EXIT;
1945                 return;
1946         }
1947
1948         if (~0ul - cli->cl_dirty_sum <= cli->cl_dirty) {
1949                 cli->cl_dirty_av = (cli->cl_dirty_av +
1950                                     (cli->cl_dirty_sum / cli->cl_dirty_num)) / 2;
1951                 cli->cl_dirty_num = 0;
1952                 cli->cl_dirty_sum = 0;
1953         } else {
1954                 if (cli->cl_dirty_num)
1955                         cli->cl_dirty_av = (cli->cl_dirty_sum / cli->cl_dirty_num);
1956         }
1957
1958         cli->cl_dirty_num++;
1959         cli->cl_dirty_sum += cli->cl_dirty;
1960
1961         if (cli->cl_dirty > cli->cl_dirty_dmax)
1962                 cli->cl_dirty_dmax = cli->cl_dirty;
1963         if (cli->cl_dirty < cli->cl_dirty_dmin || !cli->cl_dirty_dmin)
1964                 cli->cl_dirty_dmin = cli->cl_dirty;
1965
1966         oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1967         cli->cl_dirty -= PAGE_SIZE;
1968         if (!sent) {
1969                 cli->cl_lost_grant += PAGE_SIZE;
1970                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1971                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1972         }
1973
1974         EXIT;
1975 }
1976
1977 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1978                         struct lov_oinfo *loi, struct page *page,
1979                         obd_off offset, struct obd_async_page_ops *ops,
1980                         void *data, void **res)
1981 {
1982         struct osc_async_page *oap;
1983         ENTRY;
1984
1985         OBD_ALLOC(oap, sizeof(*oap));
1986         if (oap == NULL)
1987                 return -ENOMEM;
1988
1989         oap->oap_magic = OAP_MAGIC;
1990         oap->oap_cli = &exp->exp_obd->u.cli;
1991         oap->oap_loi = loi;
1992
1993         oap->oap_caller_ops = ops;
1994         oap->oap_caller_data = data;
1995
1996         oap->oap_page = page;
1997         oap->oap_obj_off = offset;
1998
1999         INIT_LIST_HEAD(&oap->oap_pending_item);
2000         INIT_LIST_HEAD(&oap->oap_urgent_item);
2001         INIT_LIST_HEAD(&oap->oap_rpc_item);
2002
2003         oap->oap_occ.occ_interrupted = osc_occ_interrupted;
2004
2005         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2006         *res = oap;
2007         RETURN(0);
2008 }
2009
2010 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2011                               struct lov_oinfo *loi, void *cookie,
2012                               int cmd, obd_off off, int count,
2013                               obd_flags brw_flags, enum async_flags async_flags)
2014 {
2015         struct client_obd *cli = &exp->exp_obd->u.cli;
2016         struct osc_async_page *oap;
2017         struct loi_oap_pages *lop;
2018         int rc;
2019         ENTRY;
2020
2021         oap = OAP_FROM_COOKIE(cookie);
2022
2023         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2024                 RETURN(-EIO);
2025
2026         if (!list_empty(&oap->oap_pending_item) ||
2027             !list_empty(&oap->oap_urgent_item) ||
2028             !list_empty(&oap->oap_rpc_item))
2029                 RETURN(-EBUSY);
2030
2031         if (loi == NULL)
2032                 loi = &lsm->lsm_oinfo[0];
2033
2034         spin_lock(&cli->cl_loi_list_lock);
2035
2036         oap->oap_cmd = cmd;
2037         oap->oap_async_flags = async_flags;
2038         oap->oap_page_off = off;
2039         oap->oap_count = count;
2040         oap->oap_brw_flags = brw_flags;
2041
2042         if (cmd == OBD_BRW_WRITE) {
2043                 rc = osc_enter_cache(cli, loi, oap);
2044                 if (rc) {
2045                         spin_unlock(&cli->cl_loi_list_lock);
2046                         RETURN(rc);
2047                 }
2048                 lop = &loi->loi_write_lop;
2049         } else {
2050                 lop = &loi->loi_read_lop;
2051         }
2052
2053         if (oap->oap_async_flags & ASYNC_URGENT)
2054                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2055         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2056         lop_update_pending(cli, lop, cmd, 1);
2057
2058         loi_list_maint(cli, loi);
2059
2060         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2061                   cmd);
2062
2063         osc_check_rpcs(cli);
2064         spin_unlock(&cli->cl_loi_list_lock);
2065
2066         RETURN(0);
2067 }
2068
2069 /* aka (~was & now & flag), but this is more clear :) */
2070 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2071
2072 static int osc_set_async_flags(struct obd_export *exp,
2073                                struct lov_stripe_md *lsm,
2074                                struct lov_oinfo *loi, void *cookie,
2075                                obd_flags async_flags)
2076 {
2077         struct client_obd *cli = &exp->exp_obd->u.cli;
2078         struct loi_oap_pages *lop;
2079         struct osc_async_page *oap;
2080         int rc = 0;
2081         ENTRY;
2082
2083         oap = OAP_FROM_COOKIE(cookie);
2084
2085         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2086                 RETURN(-EIO);
2087
2088         if (loi == NULL)
2089                 loi = &lsm->lsm_oinfo[0];
2090
2091         if (oap->oap_cmd == OBD_BRW_WRITE) {
2092                 lop = &loi->loi_write_lop;
2093         } else {
2094                 lop = &loi->loi_read_lop;
2095         }
2096
2097         spin_lock(&cli->cl_loi_list_lock);
2098
2099         if (list_empty(&oap->oap_pending_item))
2100                 GOTO(out, rc = -EINVAL);
2101
2102         if ((oap->oap_async_flags & async_flags) == async_flags)
2103                 GOTO(out, rc = 0);
2104
2105         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2106                 oap->oap_async_flags |= ASYNC_READY;
2107
2108         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2109                 if (list_empty(&oap->oap_rpc_item)) {
2110                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2111                         loi_list_maint(cli, loi);
2112                 }
2113         }
2114
2115         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2116                         oap->oap_async_flags);
2117 out:
2118         osc_check_rpcs(cli);
2119         spin_unlock(&cli->cl_loi_list_lock);
2120         RETURN(rc);
2121 }
2122
2123 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2124                              struct lov_oinfo *loi,
2125                              struct obd_io_group *oig, void *cookie,
2126                              int cmd, obd_off off, int count,
2127                              obd_flags brw_flags,
2128                              obd_flags async_flags)
2129 {
2130         struct client_obd *cli = &exp->exp_obd->u.cli;
2131         struct osc_async_page *oap;
2132         struct loi_oap_pages *lop;
2133         ENTRY;
2134
2135         oap = OAP_FROM_COOKIE(cookie);
2136
2137         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2138                 RETURN(-EIO);
2139
2140         if (!list_empty(&oap->oap_pending_item) ||
2141             !list_empty(&oap->oap_urgent_item) ||
2142             !list_empty(&oap->oap_rpc_item))
2143                 RETURN(-EBUSY);
2144
2145         if (loi == NULL)
2146                 loi = &lsm->lsm_oinfo[0];
2147
2148         spin_lock(&cli->cl_loi_list_lock);
2149
2150         oap->oap_cmd = cmd;
2151         oap->oap_page_off = off;
2152         oap->oap_count = count;
2153         oap->oap_brw_flags = brw_flags;
2154         oap->oap_async_flags = async_flags;
2155
2156         if (cmd == OBD_BRW_WRITE)
2157                 lop = &loi->loi_write_lop;
2158         else
2159                 lop = &loi->loi_read_lop;
2160
2161         list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2162         if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2163                 oap->oap_oig = oig;
2164                 oig_add_one(oig, &oap->oap_occ);
2165         }
2166
2167         LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
2168
2169         spin_unlock(&cli->cl_loi_list_lock);
2170
2171         RETURN(0);
2172 }
2173
2174 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2175                                  struct loi_oap_pages *lop, int cmd)
2176 {
2177         struct list_head *pos, *tmp;
2178         struct osc_async_page *oap;
2179
2180         list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2181                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2182                 list_del(&oap->oap_pending_item);
2183                 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2184                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2185                 lop_update_pending(cli, lop, cmd, 1);
2186         }
2187         loi_list_maint(cli, loi);
2188 }
2189
2190 static int osc_trigger_group_io(struct obd_export *exp,
2191                                 struct lov_stripe_md *lsm,
2192                                 struct lov_oinfo *loi,
2193                                 struct obd_io_group *oig)
2194 {
2195         struct client_obd *cli = &exp->exp_obd->u.cli;
2196         ENTRY;
2197
2198         if (loi == NULL)
2199                 loi = &lsm->lsm_oinfo[0];
2200
2201         spin_lock(&cli->cl_loi_list_lock);
2202
2203         osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2204         osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2205
2206         osc_check_rpcs(cli);
2207         spin_unlock(&cli->cl_loi_list_lock);
2208
2209         RETURN(0);
2210 }
2211
2212 static int osc_teardown_async_page(struct obd_export *exp,
2213                                    struct lov_stripe_md *lsm,
2214                                    struct lov_oinfo *loi, void *cookie)
2215 {
2216         struct client_obd *cli = &exp->exp_obd->u.cli;
2217         struct loi_oap_pages *lop;
2218         struct osc_async_page *oap;
2219         int rc = 0;
2220         ENTRY;
2221
2222         oap = OAP_FROM_COOKIE(cookie);
2223
2224         if (loi == NULL)
2225                 loi = &lsm->lsm_oinfo[0];
2226
2227         if (oap->oap_cmd == OBD_BRW_WRITE) {
2228                 lop = &loi->loi_write_lop;
2229         } else {
2230                 lop = &loi->loi_read_lop;
2231         }
2232
2233         spin_lock(&cli->cl_loi_list_lock);
2234
2235         if (!list_empty(&oap->oap_rpc_item))
2236                 GOTO(out, rc = -EBUSY);
2237
2238         osc_exit_cache(cli, oap, 0);
2239         osc_wake_cache_waiters(cli);
2240
2241         if (!list_empty(&oap->oap_urgent_item)) {
2242                 list_del_init(&oap->oap_urgent_item);
2243                 oap->oap_async_flags &= ~ASYNC_URGENT;
2244         }
2245         if (!list_empty(&oap->oap_pending_item)) {
2246                 list_del_init(&oap->oap_pending_item);
2247                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2248         }
2249         loi_list_maint(cli, loi);
2250
2251         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2252 out:
2253         spin_unlock(&cli->cl_loi_list_lock);
2254         if (rc == 0)
2255                 OBD_FREE(oap, sizeof(*oap));
2256         RETURN(rc);
2257 }
2258
2259 #ifdef __KERNEL__
2260 /* Note: caller will lock/unlock, and set uptodate on the pages */
2261 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2262 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2263                            struct lov_stripe_md *lsm, obd_count page_count,
2264                            struct brw_page *pga)
2265 {
2266         struct ptlrpc_request *request = NULL;
2267         struct ost_body *body;
2268         struct niobuf_remote *nioptr;
2269         struct obd_ioobj *iooptr;
2270         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2271         int swab;
2272         ENTRY;
2273
2274         /* XXX does not handle 'new' brw protocol */
2275
2276         size[1] = sizeof(struct obd_ioobj);
2277         size[2] = page_count * sizeof(*nioptr);
2278
2279         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2280                                   OST_SAN_READ, 3, size, NULL);
2281         if (!request)
2282                 RETURN(-ENOMEM);
2283
2284         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2285         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2286         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2287                                 sizeof(*nioptr) * page_count);
2288
2289         memcpy(&body->oa, oa, sizeof(body->oa));
2290
2291         obdo_to_ioobj(oa, iooptr);
2292         iooptr->ioo_bufcnt = page_count;
2293
2294         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2295                 LASSERT(PageLocked(pga[mapped].pg));
2296                 LASSERT(mapped == 0 ||
2297                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2298
2299                 nioptr->offset = pga[mapped].disk_offset;
2300                 nioptr->len    = pga[mapped].count;
2301                 nioptr->flags  = pga[mapped].flag;
2302         }
2303
2304         size[1] = page_count * sizeof(*nioptr);
2305         request->rq_replen = lustre_msg_size(2, size);
2306
2307         rc = ptlrpc_queue_wait(request);
2308         if (rc)
2309                 GOTO(out_req, rc);
2310
2311         body = lustre_swab_repbuf(request, 0, sizeof(*body),
2312                                   lustre_swab_ost_body);
2313         if (body == NULL) {
2314                 CERROR("Can't unpack body\n");
2315                 GOTO(out_req, rc = -EPROTO);
2316         }
2317
2318         memcpy(oa, &body->oa, sizeof(*oa));
2319
2320         swab = lustre_msg_swabbed(request->rq_repmsg);
2321         LASSERT_REPSWAB(request, 1);
2322         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2323         if (!nioptr) {
2324                 /* nioptr missing or short */
2325                 GOTO(out_req, rc = -EPROTO);
2326         }
2327
2328         /* actual read */
2329         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2330                 struct page *page = pga[mapped].pg;
2331                 struct buffer_head *bh;
2332                 kdev_t dev;
2333
2334                 if (swab)
2335                         lustre_swab_niobuf_remote (nioptr);
2336
2337                 /* got san device associated */
2338                 LASSERT(exp->exp_obd != NULL);
2339                 dev = exp->exp_obd->u.cli.cl_sandev;
2340
2341                 /* hole */
2342                 if (!nioptr->offset) {
2343                         CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2344                                         page->mapping->host->i_ino,
2345                                         page->index);
2346                         memset(page_address(page), 0, PAGE_SIZE);
2347                         continue;
2348                 }
2349
2350                 if (!page->buffers) {
2351                         create_empty_buffers(page, dev, PAGE_SIZE);
2352                         bh = page->buffers;
2353
2354                         clear_bit(BH_New, &bh->b_state);
2355                         set_bit(BH_Mapped, &bh->b_state);
2356                         bh->b_blocknr = (unsigned long)nioptr->offset;
2357
2358                         clear_bit(BH_Uptodate, &bh->b_state);
2359
2360                         ll_rw_block(READ, 1, &bh);
2361                 } else {
2362                         bh = page->buffers;
2363
2364                         /* if buffer already existed, it must be the
2365                          * one we mapped before, check it */
2366                         LASSERT(!test_bit(BH_New, &bh->b_state));
2367                         LASSERT(test_bit(BH_Mapped, &bh->b_state));
2368                         LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2369
2370                         /* wait it's io completion */
2371                         if (test_bit(BH_Lock, &bh->b_state))
2372                                 wait_on_buffer(bh);
2373
2374                         if (!test_bit(BH_Uptodate, &bh->b_state))
2375                                 ll_rw_block(READ, 1, &bh);
2376                 }
2377
2378
2379                 /* must do syncronous write here */
2380                 wait_on_buffer(bh);
2381                 if (!buffer_uptodate(bh)) {
2382                         /* I/O error */
2383                         rc = -EIO;
2384                         goto out_req;
2385                 }
2386         }
2387
2388 out_req:
2389         ptlrpc_req_finished(request);
2390         RETURN(rc);
2391 }
2392
2393 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2394                             struct lov_stripe_md *lsm, obd_count page_count,
2395                             struct brw_page *pga)
2396 {
2397         struct ptlrpc_request *request = NULL;
2398         struct ost_body *body;
2399         struct niobuf_remote *nioptr;
2400         struct obd_ioobj *iooptr;
2401         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2402         int swab;
2403         ENTRY;
2404
2405         size[1] = sizeof(struct obd_ioobj);
2406         size[2] = page_count * sizeof(*nioptr);
2407
2408         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2409                                   OST_SAN_WRITE, 3, size, NULL);
2410         if (!request)
2411                 RETURN(-ENOMEM);
2412
2413         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2414         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2415         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2416                                 sizeof (*nioptr) * page_count);
2417
2418         memcpy(&body->oa, oa, sizeof(body->oa));
2419
2420         obdo_to_ioobj(oa, iooptr);
2421         iooptr->ioo_bufcnt = page_count;
2422
2423         /* pack request */
2424         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2425                 LASSERT(PageLocked(pga[mapped].pg));
2426                 LASSERT(mapped == 0 ||
2427                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2428
2429                 nioptr->offset = pga[mapped].disk_offset;
2430                 nioptr->len    = pga[mapped].count;
2431                 nioptr->flags  = pga[mapped].flag;
2432         }
2433
2434         size[1] = page_count * sizeof(*nioptr);
2435         request->rq_replen = lustre_msg_size(2, size);
2436
2437         rc = ptlrpc_queue_wait(request);
2438         if (rc)
2439                 GOTO(out_req, rc);
2440
2441         swab = lustre_msg_swabbed (request->rq_repmsg);
2442         LASSERT_REPSWAB (request, 1);
2443         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2444         if (!nioptr) {
2445                 CERROR("absent/short niobuf array\n");
2446                 GOTO(out_req, rc = -EPROTO);
2447         }
2448
2449         /* actual write */
2450         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2451                 struct page *page = pga[mapped].pg;
2452                 struct buffer_head *bh;
2453                 kdev_t dev;
2454
2455                 if (swab)
2456                         lustre_swab_niobuf_remote (nioptr);
2457
2458                 /* got san device associated */
2459                 LASSERT(exp->exp_obd != NULL);
2460                 dev = exp->exp_obd->u.cli.cl_sandev;
2461
2462                 if (!page->buffers) {
2463                         create_empty_buffers(page, dev, PAGE_SIZE);
2464                 } else {
2465                         /* checking */
2466                         LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2467                         LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2468                         LASSERT(page->buffers->b_blocknr ==
2469                                 (unsigned long)nioptr->offset);
2470                 }
2471                 bh = page->buffers;
2472
2473                 LASSERT(bh);
2474
2475                 /* if buffer locked, wait it's io completion */
2476                 if (test_bit(BH_Lock, &bh->b_state))
2477                         wait_on_buffer(bh);
2478
2479                 clear_bit(BH_New, &bh->b_state);
2480                 set_bit(BH_Mapped, &bh->b_state);
2481
2482                 /* override the block nr */
2483                 bh->b_blocknr = (unsigned long)nioptr->offset;
2484
2485                 /* we are about to write it, so set it
2486                  * uptodate/dirty
2487                  * page lock should garentee no race condition here */
2488                 set_bit(BH_Uptodate, &bh->b_state);
2489                 set_bit(BH_Dirty, &bh->b_state);
2490
2491                 ll_rw_block(WRITE, 1, &bh);
2492
2493                 /* must do syncronous write here */
2494                 wait_on_buffer(bh);
2495                 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2496                         /* I/O error */
2497                         rc = -EIO;
2498                         goto out_req;
2499                 }
2500         }
2501
2502 out_req:
2503         ptlrpc_req_finished(request);
2504         RETURN(rc);
2505 }
2506
2507 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2508                       struct lov_stripe_md *lsm, obd_count page_count,
2509                       struct brw_page *pga, struct obd_trans_info *oti)
2510 {
2511         ENTRY;
2512
2513         while (page_count) {
2514                 obd_count pages_per_brw;
2515                 int rc;
2516
2517                 if (page_count > PTLRPC_MAX_BRW_PAGES)
2518                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2519                 else
2520                         pages_per_brw = page_count;
2521
2522                 if (cmd & OBD_BRW_WRITE)
2523                         rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2524                 else
2525                         rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2526
2527                 if (rc != 0)
2528                         RETURN(rc);
2529
2530                 page_count -= pages_per_brw;
2531                 pga += pages_per_brw;
2532         }
2533         RETURN(0);
2534 }
2535 #endif
2536 #endif
2537
2538 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2539 {
2540         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2541
2542         if (lock == NULL) {
2543                 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2544                 return;
2545         }
2546
2547         lock_res_and_lock(lock);
2548 #ifdef __KERNEL__
2549         if (lock->l_ast_data && lock->l_ast_data != data) {
2550                 struct inode *new_inode = data;
2551                 struct inode *old_inode = lock->l_ast_data;
2552                 if (!(old_inode->i_state & I_FREEING))
2553                         LDLM_ERROR(lock, "inconsistent l_ast_data found");
2554                 LASSERTF(old_inode->i_state & I_FREEING,
2555                          "Found existing inode %p/%lu/%u state %lu in lock: "
2556                          "setting data to %p/%lu/%u\n", old_inode,
2557                          old_inode->i_ino, old_inode->i_generation,
2558                          old_inode->i_state,
2559                          new_inode, new_inode->i_ino, new_inode->i_generation);
2560         }
2561 #endif
2562         lock->l_ast_data = data;
2563         unlock_res_and_lock(lock);
2564         LDLM_LOCK_PUT(lock);
2565 }
2566
2567 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2568                              ldlm_iterator_t replace, void *data)
2569 {
2570         struct ldlm_res_id res_id = { .name = {0} };
2571         struct obd_device *obd = class_exp2obd(exp);
2572
2573         res_id.name[0] = lsm->lsm_object_id;
2574         res_id.name[2] = lsm->lsm_object_gr;
2575         ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2576         return 0;
2577 }
2578
2579 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2580                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2581                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2582                        void *data, __u32 lvb_len, void *lvb_swabber,
2583                        struct lustre_handle *lockh)
2584 {
2585         struct obd_device *obd = exp->exp_obd;
2586         struct ldlm_res_id res_id = { .name = {0} };
2587         struct ost_lvb lvb;
2588         struct ldlm_reply *rep;
2589         struct ptlrpc_request *req = NULL;
2590         int rc;
2591         ENTRY;
2592
2593         res_id.name[0] = lsm->lsm_object_id;
2594         res_id.name[2] = lsm->lsm_object_gr;
2595
2596         /* Filesystem lock extents are extended to page boundaries so that
2597          * dealing with the page cache is a little smoother.  */
2598         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2599         policy->l_extent.end |= ~PAGE_MASK;
2600
2601         if (lsm->lsm_oinfo->loi_kms_valid == 0)
2602                 goto no_match;
2603
2604         /* Next, search for already existing extent locks that will cover us */
2605         rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2606                              lockh);
2607         if (rc == 1) {
2608                 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2609                         /* return immediately if no credential held */
2610                         ldlm_lock_decref(lockh, mode);
2611                         RETURN(-EACCES);
2612                 }
2613
2614                 osc_set_data_with_check(lockh, data);
2615                 if (*flags & LDLM_FL_HAS_INTENT) {
2616                         /* I would like to be able to ASSERT here that rss <=
2617                          * kms, but I can't, for reasons which are explained in
2618                          * lov_enqueue() */
2619                 }
2620                 /* We already have a lock, and it's referenced */
2621                 RETURN(ELDLM_OK);
2622         }
2623
2624         /* If we're trying to read, we also search for an existing PW lock.  The
2625          * VFS and page cache already protect us locally, so lots of readers/
2626          * writers can share a single PW lock.
2627          *
2628          * There are problems with conversion deadlocks, so instead of
2629          * converting a read lock to a write lock, we'll just enqueue a new
2630          * one.
2631          *
2632          * At some point we should cancel the read lock instead of making them
2633          * send us a blocking callback, but there are problems with canceling
2634          * locks out from other users right now, too. */
2635
2636         if (mode == LCK_PR) {
2637                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2638                                      policy, LCK_PW, lockh);
2639                 if (rc == 1) {
2640                         if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2641                                 /* return immediately if no credential held */
2642                                 ldlm_lock_decref(lockh, LCK_PW);
2643                                 RETURN(-EACCES);
2644                         }
2645
2646                         /* FIXME: This is not incredibly elegant, but it might
2647                          * be more elegant than adding another parameter to
2648                          * lock_match.  I want a second opinion. */
2649                         ldlm_lock_addref(lockh, LCK_PR);
2650                         ldlm_lock_decref(lockh, LCK_PW);
2651                         osc_set_data_with_check(lockh, data);
2652                         RETURN(ELDLM_OK);
2653                 }
2654         }
2655         if (mode == LCK_PW) {
2656                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2657                                      policy, LCK_PR, lockh);
2658                 if (rc == 1) {
2659                         rc = ldlm_cli_convert(lockh, mode, flags);
2660                         if (!rc) {
2661                                 /* Update readers/writers accounting */
2662                                 ldlm_lock_addref(lockh, LCK_PW);
2663                                 ldlm_lock_decref(lockh, LCK_PR);
2664                                 osc_set_data_with_check(lockh, data);
2665                                 RETURN(ELDLM_OK);
2666                         }
2667                         /* If the conversion failed, we need to drop refcount
2668                            on matched lock before we get new one */
2669                         /* XXX Won't it save us some efforts if we cancel PR
2670                            lock here? We are going to take PW lock anyway and it
2671                            will invalidate PR lock */
2672                         ldlm_lock_decref(lockh, LCK_PR);
2673                         if (rc != EDEADLOCK) {
2674                                 RETURN(rc);
2675                         }
2676                 }
2677         }
2678
2679 no_match:
2680         if (*flags & LDLM_FL_HAS_INTENT) {
2681                 int size[2] = {0, sizeof(struct ldlm_request)};
2682
2683                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
2684                                       LDLM_ENQUEUE, 2, size, NULL);
2685                 if (req == NULL)
2686                         RETURN(-ENOMEM);
2687
2688                 size[0] = sizeof(*rep);
2689                 size[1] = sizeof(lvb);
2690                 req->rq_replen = lustre_msg_size(2, size);
2691         }
2692         rc = ldlm_cli_enqueue(exp, req, obd->obd_namespace, res_id, type,
2693                               policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2694                               &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2695         if (req != NULL) {
2696                 if (rc == ELDLM_LOCK_ABORTED) {
2697                         /* swabbed by ldlm_cli_enqueue() */
2698                         LASSERT_REPSWABBED(req, 0);
2699                         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
2700                         LASSERT(rep != NULL);
2701                         if (rep->lock_policy_res1)
2702                                 rc = rep->lock_policy_res1;
2703                 }
2704                 ptlrpc_req_finished(req);
2705         }
2706
2707         if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2708                 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2709                        lvb.lvb_size, lvb.lvb_blocks);
2710                 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2711                 lsm->lsm_oinfo->loi_mtime = lvb.lvb_mtime;
2712                 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2713         }
2714
2715         RETURN(rc);
2716 }
2717
2718 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2719                      __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2720                      int *flags, void *data, struct lustre_handle *lockh)
2721 {
2722         struct ldlm_res_id res_id = { .name = {0} };
2723         struct obd_device *obd = exp->exp_obd;
2724         int rc;
2725         ENTRY;
2726
2727         res_id.name[0] = lsm->lsm_object_id;
2728         res_id.name[2] = lsm->lsm_object_gr;
2729
2730         OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2731
2732         /* Filesystem lock extents are extended to page boundaries so that
2733          * dealing with the page cache is a little smoother */
2734         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2735         policy->l_extent.end |= ~PAGE_MASK;
2736
2737         /* Next, search for already existing extent locks that will cover us */
2738         rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2739                              policy, mode, lockh);
2740         if (rc) {
2741                // if (!(*flags & LDLM_FL_TEST_LOCK))
2742                         osc_set_data_with_check(lockh, data);
2743                 RETURN(rc);
2744         }
2745         /* If we're trying to read, we also search for an existing PW lock.  The
2746          * VFS and page cache already protect us locally, so lots of readers/
2747          * writers can share a single PW lock. */
2748         if (mode == LCK_PR) {
2749                 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2750                                      policy, LCK_PW, lockh);
2751                 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2752                         /* FIXME: This is not incredibly elegant, but it might
2753                          * be more elegant than adding another parameter to
2754                          * lock_match.  I want a second opinion. */
2755                         osc_set_data_with_check(lockh, data);
2756                         ldlm_lock_addref(lockh, LCK_PR);
2757                         ldlm_lock_decref(lockh, LCK_PW);
2758                 }
2759         }
2760         RETURN(rc);
2761 }
2762
2763 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2764                       __u32 mode, struct lustre_handle *lockh)
2765 {
2766         ENTRY;
2767
2768         ldlm_lock_decref(lockh, mode);
2769         RETURN(0);
2770 }
2771
2772 static int osc_cancel_unused(struct obd_export *exp,
2773                              struct lov_stripe_md *lsm,
2774                              int flags, void *opaque)
2775 {
2776         struct obd_device *obd = class_exp2obd(exp);
2777         struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2778
2779         if (lsm != NULL) {
2780                 res_id.name[0] = lsm->lsm_object_id;
2781                 res_id.name[2] = lsm->lsm_object_gr;
2782                 resp = &res_id;
2783         }
2784
2785         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2786 }
2787
2788 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2789                       unsigned long max_age)
2790 {
2791         struct obd_statfs *msfs;
2792         struct ptlrpc_request *request;
2793         int rc, size = sizeof(*osfs);
2794         ENTRY;
2795
2796         /* We could possibly pass max_age in the request (as an absolute
2797          * timestamp or a "seconds.usec ago") so the target can avoid doing
2798          * extra calls into the filesystem if that isn't necessary (e.g.
2799          * during mount that would help a bit).  Having relative timestamps
2800          * is not so great if request processing is slow, while absolute
2801          * timestamps are not ideal because they need time synchronization. */
2802         request = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OBD_VERSION,
2803                                   OST_STATFS, 0, NULL, NULL);
2804         if (!request)
2805                 RETURN(-ENOMEM);
2806
2807         request->rq_replen = lustre_msg_size(1, &size);
2808         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2809
2810         rc = ptlrpc_queue_wait(request);
2811         if (rc)
2812                 GOTO(out, rc);
2813
2814         msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2815                                   lustre_swab_obd_statfs);
2816         if (msfs == NULL) {
2817                 CERROR("Can't unpack obd_statfs\n");
2818                 GOTO(out, rc = -EPROTO);
2819         }
2820
2821         memcpy(osfs, msfs, sizeof(*osfs));
2822
2823         EXIT;
2824  out:
2825         ptlrpc_req_finished(request);
2826         return rc;
2827 }
2828
2829 /* Retrieve object striping information.
2830  *
2831  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2832  * the maximum number of OST indices which will fit in the user buffer.
2833  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2834  */
2835 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2836 {
2837         struct lov_user_md lum, *lumk;
2838         int rc, lum_size;
2839         ENTRY;
2840
2841         if (!lsm)
2842                 RETURN(-ENODATA);
2843
2844         rc = copy_from_user(&lum, lump, sizeof(lum));
2845         if (rc)
2846                 RETURN(-EFAULT);
2847
2848         if (lum.lmm_magic != LOV_USER_MAGIC)
2849                 RETURN(-EINVAL);
2850
2851         if (lum.lmm_stripe_count > 0) {
2852                 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2853                 OBD_ALLOC(lumk, lum_size);
2854                 if (!lumk)
2855                         RETURN(-ENOMEM);
2856
2857                 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2858                 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2859         } else {
2860                 lum_size = sizeof(lum);
2861                 lumk = &lum;
2862         }
2863
2864         lumk->lmm_object_id = lsm->lsm_object_id;
2865         lumk->lmm_object_gr = lsm->lsm_object_gr;
2866         lumk->lmm_stripe_count = 1;
2867
2868         if (copy_to_user(lump, lumk, lum_size))
2869                 rc = -EFAULT;
2870
2871         if (lumk != &lum)
2872                 OBD_FREE(lumk, lum_size);
2873
2874         RETURN(rc);
2875 }
2876
2877 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2878                          void *karg, void *uarg)
2879 {
2880         struct obd_device *obd = exp->exp_obd;
2881         struct obd_ioctl_data *data = karg;
2882         int err = 0;
2883         ENTRY;
2884
2885 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2886         MOD_INC_USE_COUNT;
2887 #else
2888         if (!try_module_get(THIS_MODULE)) {
2889                 CERROR("Can't get module. Is it alive?");
2890                 return -EINVAL;
2891         }
2892 #endif
2893         switch (cmd) {
2894         case OBD_IOC_LOV_GET_CONFIG: {
2895                 char *buf;
2896                 struct lov_desc *desc;
2897                 struct obd_uuid uuid;
2898
2899                 buf = NULL;
2900                 len = 0;
2901                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2902                         GOTO(out, err = -EINVAL);
2903
2904                 data = (struct obd_ioctl_data *)buf;
2905
2906                 if (sizeof(*desc) > data->ioc_inllen1) {
2907                         OBD_FREE(buf, len);
2908                         GOTO(out, err = -EINVAL);
2909                 }
2910
2911                 if (data->ioc_inllen2 < sizeof(uuid)) {
2912                         OBD_FREE(buf, len);
2913                         GOTO(out, err = -EINVAL);
2914                 }
2915
2916                 if (data->ioc_inllen3 < sizeof(__u32)) {
2917                         OBD_FREE(buf, len);
2918                         GOTO(out, err = -EINVAL);
2919                 }
2920
2921                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2922                 desc->ld_tgt_count = 1;
2923                 desc->ld_active_tgt_count = 1;
2924                 desc->ld_default_stripe_count = 1;
2925                 desc->ld_default_stripe_size = 0;
2926                 desc->ld_default_stripe_offset = 0;
2927                 desc->ld_pattern = 0;
2928                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2929                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2930                 *((__u32 *)data->ioc_inlbuf3) = 1;
2931
2932                 err = copy_to_user((void *)uarg, buf, len);
2933                 if (err)
2934                         err = -EFAULT;
2935                 obd_ioctl_freedata(buf, len);
2936                 GOTO(out, err);
2937         }
2938         case LL_IOC_LOV_SETSTRIPE:
2939                 err = obd_alloc_memmd(exp, karg);
2940                 if (err > 0)
2941                         err = 0;
2942                 GOTO(out, err);
2943         case LL_IOC_LOV_GETSTRIPE:
2944                 err = osc_getstripe(karg, uarg);
2945                 GOTO(out, err);
2946         case OBD_IOC_CLIENT_RECOVER:
2947                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2948                                             data->ioc_inlbuf1);
2949                 if (err > 0)
2950                         err = 0;
2951                 GOTO(out, err);
2952         case IOC_OSC_SET_ACTIVE:
2953                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2954                                                data->ioc_offset);
2955                 GOTO(out, err);
2956         case IOC_OSC_CTL_RECOVERY:
2957                 err = ptlrpc_import_control_recovery(obd->u.cli.cl_import,
2958                                                      data->ioc_offset);
2959                 GOTO(out, err);
2960         default:
2961                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2962                 GOTO(out, err = -ENOTTY);
2963         }
2964 out:
2965 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2966         MOD_DEC_USE_COUNT;
2967 #else
2968         module_put(THIS_MODULE);
2969 #endif
2970         return err;
2971 }
2972
2973 static int osc_get_info(struct obd_export *exp, __u32 keylen,
2974                         void *key, __u32 *vallen, void *val)
2975 {
2976         ENTRY;
2977         if (!vallen || !val)
2978                 RETURN(-EFAULT);
2979
2980         if (keylen > strlen("lock_to_stripe") &&
2981             strcmp(key, "lock_to_stripe") == 0) {
2982                 __u32 *stripe = val;
2983                 *vallen = sizeof(*stripe);
2984                 *stripe = 0;
2985                 RETURN(0);
2986         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2987                 struct ptlrpc_request *req;
2988                 obd_id *reply;
2989                 char *bufs[1] = {key};
2990                 int rc;
2991                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2992                                       OST_GET_INFO, 1, (int *)&keylen, bufs);
2993                 if (req == NULL)
2994                         RETURN(-ENOMEM);
2995
2996                 req->rq_replen = lustre_msg_size(1, (int *)vallen);
2997                 rc = ptlrpc_queue_wait(req);
2998                 if (rc)
2999                         GOTO(out, rc);
3000
3001                 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
3002                                            lustre_swab_ost_last_id);
3003                 if (reply == NULL) {
3004                         CERROR("Can't unpack OST last ID\n");
3005                         GOTO(out, rc = -EPROTO);
3006                 }
3007                 *((obd_id *)val) = *reply;
3008         out:
3009                 ptlrpc_req_finished(req);
3010                 RETURN(rc);
3011         } else if (keylen == 10 && strcmp(key, "client_nid") == 0) {
3012                 struct ptlrpc_connection * conn;
3013                 ptl_nid_t * nid = val;
3014                 ptl_process_id_t id;
3015                 int rc;
3016
3017                 *vallen = sizeof(*nid);
3018                 conn = class_exp2cliimp(exp)->imp_connection;
3019
3020                 if (!conn || !conn->c_peer.peer_ni)
3021                         RETURN(-ENOTCONN);
3022
3023                 rc = PtlGetId(conn->c_peer.peer_ni->pni_ni_h, &id);
3024                 if (rc == PTL_OK)
3025                         *nid = id.nid;
3026
3027                 RETURN(0);
3028         }
3029         RETURN(-EPROTO);
3030 }
3031
3032 static int osc_set_info(struct obd_export *exp, obd_count keylen,
3033                         void *key, obd_count vallen, void *val)
3034 {
3035         struct obd_device  *obd = exp->exp_obd;
3036         struct obd_import *imp = class_exp2cliimp(exp);
3037         struct llog_ctxt *ctxt;
3038         int rc = 0;
3039         ENTRY;
3040
3041         if (keylen == strlen("unlinked") &&
3042             memcmp(key, "unlinked", keylen) == 0) {
3043                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3044                 spin_lock(&oscc->oscc_lock);
3045                 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3046                 spin_unlock(&oscc->oscc_lock);
3047                 RETURN(0);
3048         }
3049
3050         if (keylen == strlen("unrecovery") &&
3051             memcmp(key, "unrecovery", keylen) == 0) {
3052                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3053                 spin_lock(&oscc->oscc_lock);
3054                 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3055                 spin_unlock(&oscc->oscc_lock);
3056                 RETURN(0);
3057         }
3058
3059         if (keylen == strlen("initial_recov") &&
3060             memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
3061                 struct obd_import *imp = class_exp2cliimp(exp);
3062                 if (vallen != sizeof(int))
3063                         RETURN(-EINVAL);
3064                 imp->imp_initial_recov = *(int *)val;
3065                 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
3066                        exp->exp_obd->obd_name,
3067                        imp->imp_initial_recov);
3068                 RETURN(0);
3069         }
3070
3071         if (keylen == strlen("async") &&
3072             memcmp(key, "async", keylen) == 0) {
3073                 struct client_obd *cl = &obd->u.cli;
3074                 if (vallen != sizeof(int))
3075                         RETURN(-EINVAL);
3076                 cl->cl_async = *(int *)val;
3077                 CDEBUG(D_HA, "%s: set async = %d\n",
3078                        obd->obd_name, cl->cl_async);
3079                 RETURN(0);
3080         }
3081
3082         if (keylen == 5 && strcmp(key, "audit") == 0) {
3083                 struct ptlrpc_request *req;
3084                 char *bufs[2] = {key, val};
3085                 int size[2] = {keylen, vallen};
3086
3087                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3088                                       OST_SET_INFO, 2, size, bufs);
3089                 if (req == NULL)
3090                         RETURN(-ENOMEM);
3091
3092                 req->rq_replen = lustre_msg_size(0, size);
3093                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
3094                                    lustre_swab_audit_attr);
3095                 rc = ptlrpc_queue_wait(req);
3096
3097                 ptlrpc_req_finished(req);
3098                 RETURN(rc);
3099         }
3100
3101         if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
3102                 struct ptlrpc_request *req;
3103                 char *bufs[2] = {key, val};
3104                 int size[2] = {keylen, vallen};
3105
3106                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3107                                       OST_SET_INFO, 2, size, bufs);
3108                 if (req == NULL)
3109                         RETURN(-ENOMEM);
3110
3111                 req->rq_replen = lustre_msg_size(0, size);
3112                 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
3113                                    lustre_swab_obdo);
3114                 rc = ptlrpc_queue_wait(req);
3115
3116                 ptlrpc_req_finished(req);
3117                 RETURN(rc);
3118         }
3119
3120         if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
3121                 struct ptlrpc_request *req;
3122                 char *bufs[2] = {key, val};
3123                 int size[2] = {keylen, vallen};
3124
3125                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3126                                       OST_SET_INFO, 2, size, bufs);
3127                 if (req == NULL)
3128                         RETURN(-ENOMEM);
3129
3130                 req->rq_replen = lustre_msg_size(0, size);
3131                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
3132                                    lustre_swab_audit_msg);
3133                 rc = ptlrpc_queue_wait(req);
3134
3135                 ptlrpc_req_finished(req);
3136                 RETURN(rc);
3137         }
3138
3139         if (keylen == strlen("sec") && memcmp(key, "sec", keylen) == 0) {
3140                 struct client_obd *cli = &exp->exp_obd->u.cli;
3141
3142                 cli->cl_sec_flavor = ptlrpcs_name2flavor(val);
3143                 if (cli->cl_sec_flavor == PTLRPCS_FLVR_INVALID) {
3144                         CERROR("unrecognized security flavor %s\n", (char*) val);
3145                         RETURN(-EINVAL);
3146                 }
3147
3148                 RETURN(0);
3149         }
3150
3151         if (keylen == strlen("sec_flags") &&
3152             memcmp(key, "sec_flags", keylen) == 0) {
3153                 struct client_obd *cli = &exp->exp_obd->u.cli;
3154
3155                 cli->cl_sec_flags = *((unsigned long *) val);
3156                 RETURN(0);
3157         }
3158
3159         if (keylen == strlen("flush_cred") &&
3160             memcmp(key, "flush_cred", keylen) == 0) {
3161                 struct client_obd *cli = &exp->exp_obd->u.cli;
3162
3163                 if (cli->cl_import)
3164                         ptlrpcs_import_flush_current_creds(cli->cl_import);
3165                 RETURN(0);
3166         }
3167         if (keylen == strlen("crypto_cb") &&
3168             memcmp(key, "crypto_cb", keylen) == 0) {
3169                 LASSERT(vallen == sizeof(crypt_cb_t));
3170                 osc_crypt_cb = (crypt_cb_t)val;
3171                 RETURN(0);
3172         }
3173
3174         if (keylen == 8 && memcmp(key, "capa_key", 8) == 0) {
3175                 struct ptlrpc_request *req;
3176                 char *bufs[2] = {key, val};
3177                 unsigned long irqflags;
3178                 int rc, size[2] = {keylen, vallen};
3179
3180                 LASSERT(vallen == sizeof(struct lustre_capa_key));
3181
3182                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3183                                       OST_SET_INFO, 2, size, bufs);
3184                 if (req == NULL)
3185                         RETURN(-ENOMEM);
3186
3187                 spin_lock_irqsave (&req->rq_lock, irqflags);
3188                 req->rq_replay = 1;
3189                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
3190
3191                 req->rq_replen = lustre_msg_size(0, NULL);
3192                 rc = ptlrpc_queue_wait(req);
3193                 ptlrpc_req_finished(req);
3194                 RETURN(rc);
3195         }
3196
3197         if (keylen == strlen("setext") &&
3198             memcmp(key, "setext", keylen) == 0) {
3199                 struct client_obd *cli = &exp->exp_obd->u.cli;
3200                 struct osc_creator *oscc = &cli->cl_oscc;
3201                 struct fid_extent *ext = val;
3202
3203                 oscc->oscc_next_id = (obd_id)ext->fe_start;
3204                 RETURN(0);
3205         }
3206
3207         if (keylen < strlen("mds_conn") ||
3208             memcmp(key, "mds_conn", keylen) != 0)
3209                 RETURN(-EINVAL);
3210
3211         ctxt = llog_get_context(&exp->exp_obd->obd_llogs,
3212                                 LLOG_UNLINK_ORIG_CTXT);
3213         if (ctxt) {
3214                 if (rc == 0)
3215                         rc = llog_initiator_connect(ctxt);
3216                 else
3217                         CERROR("cannot establish the connect for "
3218                                "ctxt %p: %d\n", ctxt, rc);
3219         }
3220
3221         imp->imp_server_timeout = 1;
3222         CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
3223         imp->imp_pingable = 1;
3224
3225         RETURN(rc);
3226 }
3227
3228
3229 static struct llog_operations osc_size_repl_logops = {
3230         lop_cancel: llog_obd_repl_cancel
3231 };
3232
3233 static struct llog_operations osc_unlink_orig_logops;
3234
3235 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
3236                          struct obd_device *tgt, int count,
3237                          struct llog_catid *catid)
3238 {
3239         int rc;
3240         ENTRY;
3241
3242         osc_unlink_orig_logops = llog_lvfs_ops;
3243         osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
3244         osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
3245         osc_unlink_orig_logops.lop_add = llog_catalog_add;
3246         osc_unlink_orig_logops.lop_connect = llog_origin_connect;
3247
3248         rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
3249                             &catid->lci_logid, &osc_unlink_orig_logops);
3250         if (rc)
3251                 RETURN(rc);
3252
3253         rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
3254                             &osc_size_repl_logops);
3255         RETURN(rc);
3256 }
3257
3258 static int osc_llog_finish(struct obd_device *obd,
3259                            struct obd_llogs *llogs, int count)
3260 {
3261         int rc;
3262         ENTRY;
3263
3264         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
3265         if (rc)
3266                 RETURN(rc);
3267
3268         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
3269         RETURN(rc);
3270 }
3271
3272 static int osc_connect(struct lustre_handle *exph,
3273                        struct obd_device *obd, struct obd_uuid *cluuid,
3274                        struct obd_connect_data *data,
3275                        unsigned long connect_flags)
3276 {
3277         int rc;
3278         ENTRY;
3279         rc = client_connect_import(exph, obd, cluuid, data, connect_flags);
3280         RETURN(rc);
3281 }
3282
3283 static int osc_disconnect(struct obd_export *exp, unsigned long flags)
3284 {
3285         struct obd_device *obd = class_exp2obd(exp);
3286         struct llog_ctxt *ctxt;
3287         int rc;
3288         ENTRY;
3289
3290         ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
3291         if (obd->u.cli.cl_conn_count == 1)
3292                 /* flush any remaining cancel messages out to the target */
3293                 llog_sync(ctxt, exp);
3294
3295         rc = client_disconnect_export(exp, flags);
3296         RETURN(rc);
3297 }
3298
3299 static int osc_import_event(struct obd_device *obd,
3300                             struct obd_import *imp,
3301                             enum obd_import_event event)
3302 {
3303         struct client_obd *cli;
3304         int rc = 0;
3305         ENTRY;
3306
3307         LASSERT(imp->imp_obd == obd);
3308
3309         switch (event) {
3310         case IMP_EVENT_DISCON: {
3311                 /* Only do this on the MDS OSC's */
3312                 if (imp->imp_server_timeout) {
3313                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3314
3315                         spin_lock(&oscc->oscc_lock);
3316                         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3317                         spin_unlock(&oscc->oscc_lock);
3318                 }
3319                 break;
3320         }
3321         case IMP_EVENT_INACTIVE: {
3322                 if (obd->obd_observer)
3323                         rc = obd_notify(obd->obd_observer, obd, 0, 0);
3324                 break;
3325         }
3326         case IMP_EVENT_INVALIDATE: {
3327                 struct ldlm_namespace *ns = obd->obd_namespace;
3328
3329                 /* Reset grants */
3330                 cli = &obd->u.cli;
3331                 spin_lock(&cli->cl_loi_list_lock);
3332                 cli->cl_avail_grant = 0;
3333                 cli->cl_lost_grant = 0;
3334                 /* all pages go to failing rpcs due to the invalid import */
3335                 osc_check_rpcs(cli);
3336                 spin_unlock(&cli->cl_loi_list_lock);
3337
3338                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3339
3340                 break;
3341         }
3342         case IMP_EVENT_ACTIVE: {
3343                 /* Only do this on the MDS OSC's */
3344                 if (imp->imp_server_timeout) {
3345                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3346
3347                         spin_lock(&oscc->oscc_lock);
3348                         oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3349                         spin_unlock(&oscc->oscc_lock);
3350                 }
3351
3352                 if (obd->obd_observer)
3353                         rc = obd_notify(obd->obd_observer, obd, 1, 0);
3354                 break;
3355         }
3356         default:
3357                 CERROR("Unknown import event %d\n", event);
3358                 LBUG();
3359         }
3360         RETURN(rc);
3361 }
3362
3363 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
3364 {
3365         struct lprocfs_static_vars lvars;
3366         int rc;
3367         ENTRY;
3368
3369         lprocfs_init_vars(osc,&lvars);
3370         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3371         if (rc < 0)
3372                 RETURN(rc);
3373
3374         rc = lproc_osc_attach_seqstat(dev);
3375         if (rc < 0) {
3376                 lprocfs_obd_detach(dev);
3377                 RETURN(rc);
3378         }
3379
3380         ptlrpc_lprocfs_register_obd(dev);
3381         RETURN(0);
3382 }
3383
3384 static int osc_detach(struct obd_device *dev)
3385 {
3386         ptlrpc_lprocfs_unregister_obd(dev);
3387         return lprocfs_obd_detach(dev);
3388 }
3389
3390 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3391 {
3392         int rc;
3393         ENTRY;
3394         rc = ptlrpcd_addref();
3395         if (rc)
3396                 RETURN(rc);
3397
3398         rc = client_obd_setup(obd, len, buf);
3399         if (rc)
3400                 ptlrpcd_decref();
3401         else
3402                 oscc_init(obd);
3403
3404         RETURN(rc);
3405 }
3406
3407 static int osc_cleanup(struct obd_device *obd, int flags)
3408 {
3409         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3410         int rc;
3411
3412         rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3413                                     LDLM_FL_CONFIG_CHANGE, NULL);
3414         if (rc)
3415                 RETURN(rc);
3416
3417         spin_lock(&oscc->oscc_lock);
3418         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3419         oscc->oscc_flags |= OSCC_FLAG_EXITING;
3420         spin_unlock(&oscc->oscc_lock);
3421
3422         rc = client_obd_cleanup(obd, flags);
3423         ptlrpcd_decref();
3424         RETURN(rc);
3425 }
3426
3427
3428 struct obd_ops osc_obd_ops = {
3429         .o_owner                = THIS_MODULE,
3430         .o_attach               = osc_attach,
3431         .o_detach               = osc_detach,
3432         .o_setup                = osc_setup,
3433         .o_cleanup              = osc_cleanup,
3434         .o_add_conn             = client_import_add_conn,
3435         .o_del_conn             = client_import_del_conn,
3436         .o_connect              = osc_connect,
3437         .o_disconnect           = osc_disconnect,
3438         .o_statfs               = osc_statfs,
3439         .o_packmd               = osc_packmd,
3440         .o_unpackmd             = osc_unpackmd,
3441         .o_create               = osc_create,
3442         .o_destroy              = osc_destroy,
3443         .o_getattr              = osc_getattr,
3444         .o_getattr_async        = osc_getattr_async,
3445         .o_setattr              = osc_setattr,
3446         .o_brw                  = osc_brw,
3447         .o_brw_async            = osc_brw_async,
3448         .o_prep_async_page      = osc_prep_async_page,
3449         .o_queue_async_io       = osc_queue_async_io,
3450         .o_set_async_flags      = osc_set_async_flags,
3451         .o_queue_group_io       = osc_queue_group_io,
3452         .o_trigger_group_io     = osc_trigger_group_io,
3453         .o_teardown_async_page  = osc_teardown_async_page,
3454         .o_punch                = osc_punch,
3455         .o_sync                 = osc_sync,
3456         .o_enqueue              = osc_enqueue,
3457         .o_match                = osc_match,
3458         .o_change_cbdata        = osc_change_cbdata,
3459         .o_cancel               = osc_cancel,
3460         .o_cancel_unused        = osc_cancel_unused,
3461         .o_iocontrol            = osc_iocontrol,
3462         .o_get_info             = osc_get_info,
3463         .o_set_info             = osc_set_info,
3464         .o_import_event         = osc_import_event,
3465         .o_llog_init            = osc_llog_init,
3466         .o_llog_finish          = osc_llog_finish,
3467 };
3468
3469 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3470 struct obd_ops sanosc_obd_ops = {
3471         .o_owner                = THIS_MODULE,
3472         .o_attach               = osc_attach,
3473         .o_detach               = osc_detach,
3474         .o_cleanup              = client_obd_cleanup,
3475         .o_add_conn             = client_import_add_conn,
3476         .o_del_conn             = client_import_del_conn,
3477         .o_connect              = osc_connect,
3478         .o_disconnect           = client_disconnect_export,
3479         .o_statfs               = osc_statfs,
3480         .o_packmd               = osc_packmd,
3481         .o_unpackmd             = osc_unpackmd,
3482         .o_create               = osc_real_create,
3483         .o_destroy              = osc_destroy,
3484         .o_getattr              = osc_getattr,
3485         .o_getattr_async        = osc_getattr_async,
3486         .o_setattr              = osc_setattr,
3487         .o_setup                = client_sanobd_setup,
3488         .o_brw                  = sanosc_brw,
3489         .o_punch                = osc_punch,
3490         .o_sync                 = osc_sync,
3491         .o_enqueue              = osc_enqueue,
3492         .o_match                = osc_match,
3493         .o_change_cbdata        = osc_change_cbdata,
3494         .o_cancel               = osc_cancel,
3495         .o_cancel_unused        = osc_cancel_unused,
3496         .o_iocontrol            = osc_iocontrol,
3497         .o_import_event         = osc_import_event,
3498         .o_llog_init            = osc_llog_init,
3499         .o_llog_finish          = osc_llog_finish,
3500 };
3501 #endif
3502
3503 int __init osc_init(void)
3504 {
3505         struct lprocfs_static_vars lvars;
3506 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3507         struct lprocfs_static_vars sanlvars;
3508 #endif
3509         int rc;
3510         ENTRY;
3511
3512         lprocfs_init_vars(osc, &lvars);
3513 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3514         lprocfs_init_vars(osc, &sanlvars);
3515 #endif
3516
3517         rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3518                                  OBD_OSC_DEVICENAME);
3519         if (rc)
3520                 RETURN(rc);
3521
3522 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3523         rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3524                                  OBD_SANOSC_DEVICENAME);
3525         if (rc)
3526                 class_unregister_type(OBD_OSC_DEVICENAME);
3527 #endif
3528
3529         RETURN(rc);
3530 }
3531
3532 #ifdef __KERNEL__
3533 static void /*__exit*/ osc_exit(void)
3534 {
3535 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3536         class_unregister_type(OBD_SANOSC_DEVICENAME);
3537 #endif
3538         class_unregister_type(OBD_OSC_DEVICENAME);
3539 }
3540
3541 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3542 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3543 MODULE_LICENSE("GPL");
3544
3545 module_init(osc_init);
3546 module_exit(osc_exit);
3547 #endif