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
 755 /*Put a global call back var here is Ugly, but put it to client_obd
 756  *also seems not a good idea, WangDi*/
 757 crypt_cb_t  osc_crypt_cb = NULL;
 758
 759 static int osc_crypt_page(struct page *page, obd_off page_off, obd_off count,
 760                           int flags)
 761 {
 762         int rc = 0;
 763         ENTRY;
 764
 765         if (osc_crypt_cb != NULL)
 766                 rc = osc_crypt_cb(page, page_off, count, flags);
 767         if (rc != 0)
 768                 CERROR("crypt page error %d \n", rc);
 769         RETURN(rc);
 770 }
 771
 772 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
 773                                 struct lov_stripe_md *lsm, obd_count page_count,
 774                                 struct brw_page *pga, int *requested_nobp,
 775                                 int *niocountp, struct ptlrpc_request **reqp)
 776 {
 777         struct ptlrpc_request   *req;
 778         struct ptlrpc_bulk_desc *desc;
 779         struct client_obd       *cli = &imp->imp_obd->u.cli;
 780         struct ost_body         *body;
 781         struct lustre_id        *raw_id = obdo_id(oa);
 782         struct obd_capa         *ocapa = NULL;
 783         struct lustre_capa      *capa = NULL;
 784         struct obd_ioobj        *ioobj;
 785         struct niobuf_remote    *niobuf;
 786         int                      niocount;
 787         int                      size[4];
 788         int                      i, bufcnt = 0;
 789         int                      requested_nob;
 790         int                      opc;
 791         int                      capa_op;
 792         int                      rc;
 793
 794         opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
 795
 796         for (niocount = i = 1; i < page_count; i++)
 797                 if (!can_merge_pages(&pga[i - 1], &pga[i]))
 798                         niocount++;
 799
 800         capa_op = (opc == OST_WRITE) ? MAY_WRITE : MAY_READ;
 801 get_capa:
 802         ocapa = capa_get(oa->o_fsuid, capa_op, raw_id->li_fid.lf_group,
 803                          raw_id->li_stc.u.e3s.l3s_ino, CLIENT_CAPA,
 804                          NULL, NULL, NULL);
 805         if (!ocapa) {
 806                 if (opc == OST_READ && capa_op == MAY_READ) {
 807                         /* partial write might cause read, MAY_WRITE capability
 808                          * should be used here */
 809                         capa_op = MAY_WRITE;
 810                         goto get_capa;
 811                 }
 812         }
 813
 814         size[bufcnt++] = sizeof(*body);
 815         size[bufcnt++] = sizeof(*ioobj);
 816         if (ocapa)
 817                 size[bufcnt++] = sizeof(*capa);
 818         size[bufcnt++] = niocount * sizeof(*niobuf);
 819
 820         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, opc, bufcnt, size, NULL);
 821         if (req == NULL)
 822                 return (-ENOMEM);
 823
 824         if (opc == OST_WRITE)
 825                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 826                                              BULK_GET_SOURCE, OST_BULK_PORTAL);
 827         else
 828                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 829                                              BULK_PUT_SINK, OST_BULK_PORTAL);
 830         if (desc == NULL)
 831                 GOTO(out, rc = -ENOMEM);
 832         /* NB request now owns desc and will free it when it gets freed */
 833
 834         bufcnt = 0;
 835         body = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*body));
 836         ioobj = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*ioobj));
 837         if (ocapa)
 838                 capa = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*capa));
 839         niobuf = lustre_msg_buf(req->rq_reqmsg, bufcnt++,
 840                                 niocount * sizeof(*niobuf));
 841
 842         memcpy(&body->oa, oa, sizeof(*oa));
 843
 844         obdo_to_ioobj(oa, ioobj);
 845         ioobj->ioo_bufcnt = niocount;
 846
 847         if (ocapa) {
 848                 capa_dup(capa, ocapa);
 849                 body->oa.o_valid |= OBD_MD_CAPA;
 850                 capa_put(ocapa, CLIENT_CAPA);
 851         }
 852
 853         LASSERT (page_count > 0);
 854
 855         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
 856                 struct brw_page *pg = &pga[i];
 857                 struct brw_page *pg_prev = pg - 1;
 858
 859                 LASSERT(pg->count > 0);
 860                 LASSERTF((pg->page_offset & ~PAGE_MASK)+ pg->count <= PAGE_SIZE,
 861                          "i: %d pg: %p pg_off: "LPU64", count: %u\n", i, pg,
 862                          pg->page_offset, pg->count);
 863                 LASSERTF(i == 0 || pg->disk_offset > pg_prev->disk_offset,
 864                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
 865                          " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
 866                          i, page_count,
 867                          pg->pg, pg->pg->private, pg->pg->index, pg->disk_offset,
 868                          pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
 869                          pg_prev->disk_offset);
 870
 871                 if (opc == OST_WRITE) {
 872                         rc = osc_encrypt_page(pg->pg, pg->page_offset, pg->count);
 873                         if (rc)
 874                                 GOTO(out, rc);
 875                 }
 876
 877                 ptlrpc_prep_bulk_page(desc, pg->pg,
 878                                       pg->page_offset & ~PAGE_MASK, pg->count);
 879                 requested_nob += pg->count;
 880
 881                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
 882                         niobuf--;
 883                         niobuf->len += pg->count;
 884                 } else {
 885                         niobuf->offset = pg->disk_offset;
 886                         niobuf->len    = pg->count;
 887                         niobuf->flags  = pg->flag;
 888                 }
 889         }
 890
 891         LASSERT((void *)(niobuf - niocount) ==
 892                 lustre_msg_buf(req->rq_reqmsg, bufcnt - 1,
 893                                niocount * sizeof(*niobuf)));
 894         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
 895
 896         /* size[0] still sizeof (*body) */
 897         if (opc == OST_WRITE) {
 898 #if CHECKSUM_BULK
 899                 body->oa.o_valid |= OBD_MD_FLCKSUM;
 900                 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
 901 #endif
 902                 /* 1 RC per niobuf */
 903                 size[1] = sizeof(__u32) * niocount;
 904                 req->rq_replen = lustre_msg_size(2, size);
 905         } else {
 906                 /* 1 RC for the whole I/O */
 907                 req->rq_replen = lustre_msg_size(1, size);
 908         }
 909
 910         *niocountp = niocount;
 911         *requested_nobp = requested_nob;
 912         *reqp = req;
 913         return (0);
 914
 915  out:
 916         ptlrpc_req_finished (req);
 917         return (rc);
 918 }
 919
 920 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
 921                                 int requested_nob, int niocount,
 922                                 obd_count page_count, struct brw_page *pga,
 923                                 int rc)
 924 {
 925         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 926         struct ost_body *body;
 927         ENTRY;
 928
 929         if (rc < 0)
 930                 RETURN(rc);
 931
 932         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 933         if (body == NULL) {
 934                 CERROR ("Can't unpack body\n");
 935                 RETURN(-EPROTO);
 936         }
 937
 938         osc_update_grant(cli, body);
 939         memcpy(oa, &body->oa, sizeof(*oa));
 940
 941         if (req->rq_reqmsg->opc == OST_WRITE) {
 942                 if (rc > 0) {
 943                         CERROR ("Unexpected +ve rc %d\n", rc);
 944                         RETURN(-EPROTO);
 945                 }
 946                 LASSERT (req->rq_bulk->bd_nob == requested_nob);
 947                 osc_decrypt_page(pga->pg, pga->page_offset,
 948                                  pga->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_page(pga->pg, pga->page_offset, pga->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         /* in failout recovery we ignore writeback failure and want
1318          * to just tell llite to unlock the page and continue */
1319         if (request->rq_reqmsg->opc == OST_WRITE &&
1320             (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1321                 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1322                        cli->cl_import,
1323                        cli->cl_import ? cli->cl_import->imp_invalid : -1);
1324                 rc = 0;
1325         }
1326
1327         spin_lock(&cli->cl_loi_list_lock);
1328
1329         if (request->rq_reqmsg->opc == OST_WRITE)
1330                 lprocfs_stime_record(&cli->cl_write_stime, &now,
1331                                      &request->rq_rpcd_start);
1332         else
1333                 lprocfs_stime_record(&cli->cl_read_stime, &now,
1334                                      &request->rq_rpcd_start);
1335
1336
1337
1338         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1339          * is called so we know whether to go to sync BRWs or wait for more
1340          * RPCs to complete */
1341         if (request->rq_reqmsg->opc == OST_WRITE)
1342                 cli->cl_w_in_flight--;
1343         else
1344                 cli->cl_r_in_flight--;
1345
1346         /* the caller may re-use the oap after the completion call so
1347          * we need to clean it up a little */
1348         list_for_each_safe(pos, n, &aa->aa_oaps) {
1349                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1350
1351                 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1352                        //oap->oap_page, oap->oap_page->index, oap);
1353
1354                 list_del_init(&oap->oap_rpc_item);
1355                 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1356         }
1357
1358         /* no write RPCs in flight, reset the time */
1359         if (request->rq_reqmsg->opc == OST_WRITE && cli->cl_w_in_flight == 0)
1360                 do_gettimeofday(&cli->cl_last_write_time);
1361
1362         osc_wake_cache_waiters(cli);
1363         osc_check_rpcs(cli);
1364         spin_unlock(&cli->cl_loi_list_lock);
1365
1366         obdo_free(aa->aa_oa);
1367         OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1368
1369         RETURN(0);
1370 }
1371
1372 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1373                                             struct list_head *rpc_list,
1374                                             int page_count, int cmd)
1375 {
1376         struct ptlrpc_request *req;
1377         struct brw_page *pga = NULL;
1378         int requested_nob, nio_count;
1379         struct osc_brw_async_args *aa;
1380         struct obdo *oa = NULL;
1381         struct obd_async_page_ops *ops = NULL;
1382         void *caller_data = NULL;
1383         struct list_head *pos;
1384         int i, rc;
1385
1386         LASSERT(!list_empty(rpc_list));
1387
1388         OBD_ALLOC(pga, sizeof(*pga) * page_count);
1389         if (pga == NULL)
1390                 RETURN(ERR_PTR(-ENOMEM));
1391
1392         oa = obdo_alloc();
1393         if (oa == NULL)
1394                 GOTO(out, req = ERR_PTR(-ENOMEM));
1395
1396         i = 0;
1397         list_for_each(pos, rpc_list) {
1398                 struct osc_async_page *oap;
1399
1400                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1401                 if (ops == NULL) {
1402                         ops = oap->oap_caller_ops;
1403                         caller_data = oap->oap_caller_data;
1404                 }
1405                 pga[i].disk_offset = oap->oap_obj_off + oap->oap_page_off;
1406                 pga[i].page_offset = pga[i].disk_offset;
1407                 pga[i].pg = oap->oap_page;
1408                 pga[i].count = oap->oap_count;
1409                 pga[i].flag = oap->oap_brw_flags;
1410                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1411                        pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1412                 i++;
1413         }
1414
1415         /* always get the data for the obdo for the rpc */
1416         LASSERT(ops != NULL);
1417         ops->ap_fill_obdo(caller_data, cmd, oa);
1418
1419         sort_brw_pages(pga, page_count);
1420         rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1421                                   pga, &requested_nob, &nio_count, &req);
1422         if (rc != 0) {
1423                 CERROR("prep_req failed: %d\n", rc);
1424                 GOTO(out, req = ERR_PTR(rc));
1425         }
1426
1427         LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1428         aa = (struct osc_brw_async_args *)&req->rq_async_args;
1429         aa->aa_oa = oa;
1430         aa->aa_requested_nob = requested_nob;
1431         aa->aa_nio_count = nio_count;
1432         aa->aa_page_count = page_count;
1433         aa->aa_pga = pga;
1434         aa->aa_cli = cli;
1435
1436 out:
1437         if (IS_ERR(req)) {
1438                 if (oa)
1439                         obdo_free(oa);
1440                 if (pga)
1441                         OBD_FREE(pga, sizeof(*pga) * page_count);
1442         }
1443         RETURN(req);
1444 }
1445
1446 /* strange write gap too long (15s) */
1447 #define CLI_ODD_WRITE_GAP 15000000
1448
1449 static void lop_update_pending(struct client_obd *cli,
1450                                struct loi_oap_pages *lop, int cmd, int delta)
1451 {
1452         lop->lop_num_pending += delta;
1453         if (cmd == OBD_BRW_WRITE)
1454                 cli->cl_pending_w_pages += delta;
1455         else
1456                 cli->cl_pending_r_pages += delta;
1457 }
1458
1459 static long timeval_sub(struct timeval *large, struct timeval *small)
1460 {
1461         return (large->tv_sec - small->tv_sec) * 1000000 +
1462                 (large->tv_usec - small->tv_usec);
1463 }
1464
1465 /* the loi lock is held across this function but it's allowed to release
1466  * and reacquire it during its work */
1467 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1468                             int cmd, struct loi_oap_pages *lop)
1469 {
1470         struct ptlrpc_request *request;
1471         obd_count page_count = 0;
1472         struct list_head *tmp, *pos;
1473         struct osc_async_page *oap = NULL;
1474         struct osc_brw_async_args *aa;
1475         struct obd_async_page_ops *ops;
1476         LIST_HEAD(rpc_list);
1477         ENTRY;
1478
1479         LASSERT(lop != LP_POISON);
1480         LASSERT(lop->lop_pending.next != LP_POISON);
1481
1482         /* first we find the pages we're allowed to work with */
1483         list_for_each_safe(pos, tmp, &lop->lop_pending) {
1484                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1485                 ops = oap->oap_caller_ops;
1486
1487                 LASSERT(oap->oap_magic == OAP_MAGIC);
1488
1489                 /* in llite being 'ready' equates to the page being locked
1490                  * until completion unlocks it.  commit_write submits a page
1491                  * as not ready because its unlock will happen unconditionally
1492                  * as the call returns.  if we race with commit_write giving
1493                  * us that page we dont' want to create a hole in the page
1494                  * stream, so we stop and leave the rpc to be fired by
1495                  * another dirtier or kupdated interval (the not ready page
1496                  * will still be on the dirty list).  we could call in
1497                  * at the end of ll_file_write to process the queue again. */
1498                 if (!(oap->oap_async_flags & ASYNC_READY)) {
1499                         int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1500                         if (rc < 0)
1501                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
1502                                                 "instead of ready\n", oap,
1503                                                 oap->oap_page, rc);
1504                         switch (rc) {
1505                         case -EAGAIN:
1506                                 /* llite is telling us that the page is still
1507                                  * in commit_write and that we should try
1508                                  * and put it in an rpc again later.  we
1509                                  * break out of the loop so we don't create
1510                                  * a hole in the sequence of pages in the rpc
1511                                  * stream.*/
1512                                 pos = NULL;
1513                                 break;
1514                         case -EINTR:
1515                                 /* the io isn't needed.. tell the checks
1516                                  * below to complete the rpc with EINTR */
1517                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1518                                 oap->oap_count = -EINTR;
1519                                 break;
1520                         case 0:
1521                                 oap->oap_async_flags |= ASYNC_READY;
1522                                 break;
1523                         default:
1524                                 LASSERTF(0, "oap %p page %p returned %d "
1525                                             "from make_ready\n", oap,
1526                                             oap->oap_page, rc);
1527                                 break;
1528                         }
1529                 }
1530                 if (pos == NULL)
1531                         break;
1532                 /*
1533                  * Page submitted for IO has to be locked. Either by
1534                  * ->ap_make_ready() or by higher layers.
1535                  *
1536                  * XXX nikita: this assertion should be adjusted when lustre
1537                  * starts using PG_writeback for pages being written out.
1538                  */
1539                 LASSERT(PageLocked(oap->oap_page));
1540
1541                 /* take the page out of our book-keeping */
1542                 list_del_init(&oap->oap_pending_item);
1543                 lop_update_pending(cli, lop, cmd, -1);
1544                 list_del_init(&oap->oap_urgent_item);
1545
1546                 /* ask the caller for the size of the io as the rpc leaves. */
1547                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1548                         oap->oap_count =
1549                                 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1550                 if (oap->oap_count <= 0) {
1551                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1552                                oap->oap_count);
1553                         osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1554                         continue;
1555                 }
1556
1557                 /* now put the page back in our accounting */
1558                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1559                 if (++page_count >= cli->cl_max_pages_per_rpc)
1560                         break;
1561         }
1562
1563         osc_wake_cache_waiters(cli);
1564
1565         if (page_count == 0)
1566                 RETURN(0);
1567
1568         loi_list_maint(cli, loi);
1569         spin_unlock(&cli->cl_loi_list_lock);
1570
1571         request = osc_build_req(cli, &rpc_list, page_count, cmd);
1572         if (IS_ERR(request)) {
1573                 /* this should happen rarely and is pretty bad, it makes the
1574                  * pending list not follow the dirty order */
1575                 spin_lock(&cli->cl_loi_list_lock);
1576                 list_for_each_safe(pos, tmp, &rpc_list) {
1577                         oap = list_entry(pos, struct osc_async_page,
1578                                          oap_rpc_item);
1579                         list_del_init(&oap->oap_rpc_item);
1580
1581                         /* queued sync pages can be torn down while the pages
1582                          * were between the pending list and the rpc */
1583                         if (oap->oap_interrupted) {
1584                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1585                                 osc_ap_completion(cli, NULL, oap, 0,
1586                                                   oap->oap_count);
1587                                 continue;
1588                         }
1589
1590                         /* put the page back in the loi/lop lists */
1591                         list_add_tail(&oap->oap_pending_item,
1592                                       &lop->lop_pending);
1593                         lop_update_pending(cli, lop, cmd, 1);
1594                         if (oap->oap_async_flags & ASYNC_URGENT)
1595                                 list_add(&oap->oap_urgent_item,
1596                                          &lop->lop_urgent);
1597                 }
1598                 loi_list_maint(cli, loi);
1599                 RETURN(PTR_ERR(request));
1600         }
1601
1602         LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1603         aa = (struct osc_brw_async_args *)&request->rq_async_args;
1604         INIT_LIST_HEAD(&aa->aa_oaps);
1605         list_splice(&rpc_list, &aa->aa_oaps);
1606         INIT_LIST_HEAD(&rpc_list);
1607
1608 #ifdef __KERNEL__
1609         if (cmd == OBD_BRW_READ) {
1610                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1611                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1612         } else {
1613                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1614                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1615                                  cli->cl_w_in_flight);
1616         }
1617 #endif
1618
1619         spin_lock(&cli->cl_loi_list_lock);
1620
1621         /* collect write gaps and sum of them */
1622         if (cmd == OBD_BRW_WRITE && cli->cl_w_in_flight == 0) {
1623                 struct timeval now;
1624                 long diff;
1625
1626                 do_gettimeofday(&now);
1627
1628                 if (cli->cl_last_write_time.tv_sec) {
1629                         diff = timeval_sub(&now, &cli->cl_last_write_time);
1630                         if (diff < CLI_ODD_WRITE_GAP) {
1631                                 cli->cl_write_gap_sum += diff;
1632                                 cli->cl_write_gaps++;
1633                         }
1634                 }
1635         }
1636
1637         if (cmd == OBD_BRW_READ) {
1638                 cli->cl_r_in_flight++;
1639                 cli->cl_read_num++;
1640         } else {
1641                 cli->cl_w_in_flight++;
1642                 cli->cl_write_num++;
1643         }
1644
1645         /* queued sync pages can be torn down while the pages
1646          * were between the pending list and the rpc */
1647         list_for_each(pos, &aa->aa_oaps) {
1648                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1649                 if (oap->oap_interrupted) {
1650                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1651                                oap, request);
1652                         ptlrpc_mark_interrupted(request);
1653                         break;
1654                 }
1655         }
1656
1657         CDEBUG(D_INODE, "req %p: %d pages, aa %p.  now %dr/%dw in flight\n",
1658                         request, page_count, aa, cli->cl_r_in_flight,
1659                         cli->cl_w_in_flight);
1660
1661         oap->oap_request = ptlrpc_request_addref(request);
1662         request->rq_interpret_reply = brw_interpret_oap;
1663
1664         ptlrpcd_add_req(request);
1665         RETURN(1);
1666 }
1667
1668 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1669                          int cmd)
1670 {
1671         int optimal;
1672         ENTRY;
1673
1674         if (lop->lop_num_pending == 0)
1675                 RETURN(0);
1676
1677         /* if we have an invalid import we want to drain the queued pages
1678          * by forcing them through rpcs that immediately fail and complete
1679          * the pages.  recovery relies on this to empty the queued pages
1680          * before canceling the locks and evicting down the llite pages */
1681         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1682                 RETURN(1);
1683
1684         /* stream rpcs in queue order as long as as there is an urgent page
1685          * queued.  this is our cheap solution for good batching in the case
1686          * where writepage marks some random page in the middle of the file as
1687          * urgent because of, say, memory pressure */
1688         if (!list_empty(&lop->lop_urgent))
1689                 RETURN(1);
1690
1691         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1692         optimal = cli->cl_max_pages_per_rpc;
1693         if (cmd == OBD_BRW_WRITE) {
1694                 /* trigger a write rpc stream as long as there are dirtiers
1695                  * waiting for space.  as they're waiting, they're not going to
1696                  * create more pages to coallesce with what's waiting.. */
1697                 if (!list_empty(&cli->cl_cache_waiters))
1698                         RETURN(1);
1699
1700                 /* *2 to avoid triggering rpcs that would want to include pages
1701                  * that are being queued but which can't be made ready until
1702                  * the queuer finishes with the page. this is a wart for
1703                  * llite::commit_write() */
1704                 optimal += 16;
1705         }
1706         if (lop->lop_num_pending >= optimal)
1707                 RETURN(1);
1708
1709         RETURN(0);
1710 }
1711
1712 static void on_list(struct list_head *item, struct list_head *list,
1713                     int should_be_on)
1714 {
1715         if (list_empty(item) && should_be_on)
1716                 list_add_tail(item, list);
1717         else if (!list_empty(item) && !should_be_on)
1718                 list_del_init(item);
1719 }
1720
1721 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1722  * can find pages to build into rpcs quickly */
1723 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1724 {
1725         on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1726                 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1727                 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1728
1729         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1730                 loi->loi_write_lop.lop_num_pending);
1731
1732         on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1733                 loi->loi_read_lop.lop_num_pending);
1734 }
1735
1736 #define LOI_DEBUG(LOI, STR, args...)                                     \
1737         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
1738                !list_empty(&(LOI)->loi_cli_item),                        \
1739                (LOI)->loi_write_lop.lop_num_pending,                     \
1740                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
1741                (LOI)->loi_read_lop.lop_num_pending,                      \
1742                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
1743                args)                                                     \
1744
1745 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1746 {
1747         ENTRY;
1748         /* first return all objects which we already know to have
1749          * pages ready to be stuffed into rpcs */
1750         if (!list_empty(&cli->cl_loi_ready_list))
1751                 RETURN(list_entry(cli->cl_loi_ready_list.next,
1752                                   struct lov_oinfo, loi_cli_item));
1753
1754         /* then if we have cache waiters, return all objects with queued
1755          * writes.  This is especially important when many small files
1756          * have filled up the cache and not been fired into rpcs because
1757          * they don't pass the nr_pending/object threshhold */
1758         if (!list_empty(&cli->cl_cache_waiters) &&
1759             !list_empty(&cli->cl_loi_write_list))
1760                 RETURN(list_entry(cli->cl_loi_write_list.next,
1761                                   struct lov_oinfo, loi_write_item));
1762
1763         /* then return all queued objects when we have an invalid import
1764          * so that they get flushed */
1765         if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1766                 if (!list_empty(&cli->cl_loi_write_list))
1767                         RETURN(list_entry(cli->cl_loi_write_list.next,
1768                                           struct lov_oinfo, loi_write_item));
1769                 if (!list_empty(&cli->cl_loi_read_list))
1770                         RETURN(list_entry(cli->cl_loi_read_list.next,
1771                                           struct lov_oinfo, loi_read_item));
1772         }
1773         RETURN(NULL);
1774 }
1775
1776 /* called with the loi list lock held */
1777 static void osc_check_rpcs(struct client_obd *cli)
1778 {
1779         struct lov_oinfo *loi;
1780         int rc = 0, race_counter = 0;
1781         ENTRY;
1782
1783         while ((loi = osc_next_loi(cli)) != NULL) {
1784                 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
1785                 LASSERT(loi->loi_ost_idx != LL_POISON);
1786
1787                 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
1788                         break;
1789
1790                 /* attempt some read/write balancing by alternating between
1791                  * reads and writes in an object.  The makes_rpc checks here
1792                  * would be redundant if we were getting read/write work items
1793                  * instead of objects.  we don't want send_oap_rpc to drain a
1794                  * partial read pending queue when we're given this object to
1795                  * do io on writes while there are cache waiters */
1796                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1797                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1798                                               &loi->loi_write_lop);
1799                         if (rc < 0)
1800                                 break;
1801                         if (rc > 0)
1802                                 race_counter = 0;
1803                         else
1804                                 race_counter++;
1805                 }
1806                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1807                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1808                                               &loi->loi_read_lop);
1809                         if (rc < 0)
1810                                 break;
1811                         if (rc > 0)
1812                                 race_counter = 0;
1813                         else
1814                                 race_counter++;
1815                 }
1816
1817                 /* attempt some inter-object balancing by issueing rpcs
1818                  * for each object in turn */
1819                 if (!list_empty(&loi->loi_cli_item))
1820                         list_del_init(&loi->loi_cli_item);
1821                 if (!list_empty(&loi->loi_write_item))
1822                         list_del_init(&loi->loi_write_item);
1823                 if (!list_empty(&loi->loi_read_item))
1824                         list_del_init(&loi->loi_read_item);
1825
1826                 loi_list_maint(cli, loi);
1827
1828                 /* send_oap_rpc fails with 0 when make_ready tells it to
1829                  * back off.  llite's make_ready does this when it tries
1830                  * to lock a page queued for write that is already locked.
1831                  * we want to try sending rpcs from many objects, but we
1832                  * don't want to spin failing with 0.  */
1833                 if (race_counter == 10)
1834                         break;
1835         }
1836         EXIT;
1837 }
1838
1839 /* we're trying to queue a page in the osc so we're subject to the
1840  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1841  * If the osc's queued pages are already at that limit, then we want to sleep
1842  * until there is space in the osc's queue for us.  We also may be waiting for
1843  * write credits from the OST if there are RPCs in flight that may return some
1844  * before we fall back to sync writes.
1845  *
1846  * We need this know our allocation was granted in the presence of signals */
1847 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1848 {
1849         int rc;
1850         ENTRY;
1851         spin_lock(&cli->cl_loi_list_lock);
1852         rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
1853         spin_unlock(&cli->cl_loi_list_lock);
1854         RETURN(rc);
1855 };
1856
1857 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1858  * grant or cache space. */
1859 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1860                            struct osc_async_page *oap)
1861 {
1862         struct osc_cache_waiter ocw;
1863         struct l_wait_info lwi = { 0 };
1864         struct timeval start, stop;
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                 return(-EDQUOT);
1872
1873         /* Hopefully normal case - cache space and write credits available */
1874         if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1875             cli->cl_avail_grant >= PAGE_SIZE) {
1876                 /* account for ourselves */
1877                 osc_consume_write_grant(cli, oap);
1878                 return(0);
1879         }
1880
1881         /* Make sure that there are write rpcs in flight to wait for.  This
1882          * is a little silly as this object may not have any pending but
1883          * other objects sure might. */
1884         if (cli->cl_w_in_flight) {
1885                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1886                 init_waitqueue_head(&ocw.ocw_waitq);
1887                 ocw.ocw_oap = oap;
1888                 ocw.ocw_rc = 0;
1889
1890                 loi_list_maint(cli, loi);
1891                 osc_check_rpcs(cli);
1892                 spin_unlock(&cli->cl_loi_list_lock);
1893
1894                 CDEBUG(0, "sleeping for cache space\n");
1895                 do_gettimeofday(&start);
1896                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1897                 do_gettimeofday(&stop);
1898
1899                 cli->cl_cache_wait_num++;
1900                 cli->cl_cache_wait_sum += timeval_sub(&stop, &start);
1901
1902                 spin_lock(&cli->cl_loi_list_lock);
1903                 lprocfs_stime_record(&cli->cl_enter_stime, &stop, &start);
1904                 if (!list_empty(&ocw.ocw_entry)) {
1905                         list_del(&ocw.ocw_entry);
1906                         RETURN(-EINTR);
1907                 }
1908                 RETURN(ocw.ocw_rc);
1909         }
1910
1911         RETURN(-EDQUOT);
1912 }
1913
1914 /* the companion to enter_cache, called when an oap is no longer part of the
1915  * dirty accounting.. so writeback completes or truncate happens before writing
1916  * starts.  must be called with the loi lock held. */
1917 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1918                            int sent)
1919 {
1920         ENTRY;
1921
1922         if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1923                 EXIT;
1924                 return;
1925         }
1926
1927         oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1928         cli->cl_dirty -= PAGE_SIZE;
1929         if (!sent) {
1930                 cli->cl_lost_grant += PAGE_SIZE;
1931                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1932                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1933         }
1934
1935         EXIT;
1936 }
1937
1938 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1939                         struct lov_oinfo *loi, struct page *page,
1940                         obd_off offset, struct obd_async_page_ops *ops,
1941                         void *data, void **res)
1942 {
1943         struct osc_async_page *oap;
1944         ENTRY;
1945
1946         OBD_ALLOC(oap, sizeof(*oap));
1947         if (oap == NULL)
1948                 return -ENOMEM;
1949
1950         oap->oap_magic = OAP_MAGIC;
1951         oap->oap_cli = &exp->exp_obd->u.cli;
1952         oap->oap_loi = loi;
1953
1954         oap->oap_caller_ops = ops;
1955         oap->oap_caller_data = data;
1956
1957         oap->oap_page = page;
1958         oap->oap_obj_off = offset;
1959
1960         INIT_LIST_HEAD(&oap->oap_pending_item);
1961         INIT_LIST_HEAD(&oap->oap_urgent_item);
1962         INIT_LIST_HEAD(&oap->oap_rpc_item);
1963
1964         oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1965
1966         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1967         *res = oap;
1968         RETURN(0);
1969 }
1970
1971 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1972                               struct lov_oinfo *loi, void *cookie,
1973                               int cmd, obd_off off, int count,
1974                               obd_flags brw_flags, enum async_flags async_flags)
1975 {
1976         struct client_obd *cli = &exp->exp_obd->u.cli;
1977         struct osc_async_page *oap;
1978         struct loi_oap_pages *lop;
1979         int rc;
1980         ENTRY;
1981
1982         oap = OAP_FROM_COOKIE(cookie);
1983
1984         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1985                 RETURN(-EIO);
1986
1987         if (!list_empty(&oap->oap_pending_item) ||
1988             !list_empty(&oap->oap_urgent_item) ||
1989             !list_empty(&oap->oap_rpc_item))
1990                 RETURN(-EBUSY);
1991
1992         if (loi == NULL)
1993                 loi = &lsm->lsm_oinfo[0];
1994
1995         spin_lock(&cli->cl_loi_list_lock);
1996
1997         oap->oap_cmd = cmd;
1998         oap->oap_async_flags = async_flags;
1999         oap->oap_page_off = off;
2000         oap->oap_count = count;
2001         oap->oap_brw_flags = brw_flags;
2002
2003         if (cmd == OBD_BRW_WRITE) {
2004                 rc = osc_enter_cache(cli, loi, oap);
2005                 if (rc) {
2006                         spin_unlock(&cli->cl_loi_list_lock);
2007                         RETURN(rc);
2008                 }
2009                 lop = &loi->loi_write_lop;
2010         } else {
2011                 lop = &loi->loi_read_lop;
2012         }
2013
2014         if (oap->oap_async_flags & ASYNC_URGENT)
2015                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2016         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2017         lop_update_pending(cli, lop, cmd, 1);
2018
2019         loi_list_maint(cli, loi);
2020
2021         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2022                   cmd);
2023
2024         osc_check_rpcs(cli);
2025         spin_unlock(&cli->cl_loi_list_lock);
2026
2027         RETURN(0);
2028 }
2029
2030 /* aka (~was & now & flag), but this is more clear :) */
2031 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2032
2033 static int osc_set_async_flags(struct obd_export *exp,
2034                                struct lov_stripe_md *lsm,
2035                                struct lov_oinfo *loi, void *cookie,
2036                                obd_flags async_flags)
2037 {
2038         struct client_obd *cli = &exp->exp_obd->u.cli;
2039         struct loi_oap_pages *lop;
2040         struct osc_async_page *oap;
2041         int rc = 0;
2042         ENTRY;
2043
2044         oap = OAP_FROM_COOKIE(cookie);
2045
2046         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2047                 RETURN(-EIO);
2048
2049         if (loi == NULL)
2050                 loi = &lsm->lsm_oinfo[0];
2051
2052         if (oap->oap_cmd == OBD_BRW_WRITE) {
2053                 lop = &loi->loi_write_lop;
2054         } else {
2055                 lop = &loi->loi_read_lop;
2056         }
2057
2058         spin_lock(&cli->cl_loi_list_lock);
2059
2060         if (list_empty(&oap->oap_pending_item))
2061                 GOTO(out, rc = -EINVAL);
2062
2063         if ((oap->oap_async_flags & async_flags) == async_flags)
2064                 GOTO(out, rc = 0);
2065
2066         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2067                 oap->oap_async_flags |= ASYNC_READY;
2068
2069         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2070                 if (list_empty(&oap->oap_rpc_item)) {
2071                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2072                         loi_list_maint(cli, loi);
2073                 }
2074         }
2075
2076         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2077                         oap->oap_async_flags);
2078 out:
2079         osc_check_rpcs(cli);
2080         spin_unlock(&cli->cl_loi_list_lock);
2081         RETURN(rc);
2082 }
2083
2084 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2085                              struct lov_oinfo *loi,
2086                              struct obd_io_group *oig, void *cookie,
2087                              int cmd, obd_off off, int count,
2088                              obd_flags brw_flags,
2089                              obd_flags async_flags)
2090 {
2091         struct client_obd *cli = &exp->exp_obd->u.cli;
2092         struct osc_async_page *oap;
2093         struct loi_oap_pages *lop;
2094         ENTRY;
2095
2096         oap = OAP_FROM_COOKIE(cookie);
2097
2098         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2099                 RETURN(-EIO);
2100
2101         if (!list_empty(&oap->oap_pending_item) ||
2102             !list_empty(&oap->oap_urgent_item) ||
2103             !list_empty(&oap->oap_rpc_item))
2104                 RETURN(-EBUSY);
2105
2106         if (loi == NULL)
2107                 loi = &lsm->lsm_oinfo[0];
2108
2109         spin_lock(&cli->cl_loi_list_lock);
2110
2111         oap->oap_cmd = cmd;
2112         oap->oap_page_off = off;
2113         oap->oap_count = count;
2114         oap->oap_brw_flags = brw_flags;
2115         oap->oap_async_flags = async_flags;
2116
2117         if (cmd == OBD_BRW_WRITE)
2118                 lop = &loi->loi_write_lop;
2119         else
2120                 lop = &loi->loi_read_lop;
2121
2122         list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2123         if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2124                 oap->oap_oig = oig;
2125                 oig_add_one(oig, &oap->oap_occ);
2126         }
2127
2128         LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
2129
2130         spin_unlock(&cli->cl_loi_list_lock);
2131
2132         RETURN(0);
2133 }
2134
2135 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2136                                  struct loi_oap_pages *lop, int cmd)
2137 {
2138         struct list_head *pos, *tmp;
2139         struct osc_async_page *oap;
2140
2141         list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2142                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2143                 list_del(&oap->oap_pending_item);
2144                 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2145                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2146                 lop_update_pending(cli, lop, cmd, 1);
2147         }
2148         loi_list_maint(cli, loi);
2149 }
2150
2151 static int osc_trigger_group_io(struct obd_export *exp,
2152                                 struct lov_stripe_md *lsm,
2153                                 struct lov_oinfo *loi,
2154                                 struct obd_io_group *oig)
2155 {
2156         struct client_obd *cli = &exp->exp_obd->u.cli;
2157         ENTRY;
2158
2159         if (loi == NULL)
2160                 loi = &lsm->lsm_oinfo[0];
2161
2162         spin_lock(&cli->cl_loi_list_lock);
2163
2164         osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2165         osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2166
2167         osc_check_rpcs(cli);
2168         spin_unlock(&cli->cl_loi_list_lock);
2169
2170         RETURN(0);
2171 }
2172
2173 static int osc_teardown_async_page(struct obd_export *exp,
2174                                    struct lov_stripe_md *lsm,
2175                                    struct lov_oinfo *loi, void *cookie)
2176 {
2177         struct client_obd *cli = &exp->exp_obd->u.cli;
2178         struct loi_oap_pages *lop;
2179         struct osc_async_page *oap;
2180         int rc = 0;
2181         ENTRY;
2182
2183         oap = OAP_FROM_COOKIE(cookie);
2184
2185         if (loi == NULL)
2186                 loi = &lsm->lsm_oinfo[0];
2187
2188         if (oap->oap_cmd == OBD_BRW_WRITE) {
2189                 lop = &loi->loi_write_lop;
2190         } else {
2191                 lop = &loi->loi_read_lop;
2192         }
2193
2194         spin_lock(&cli->cl_loi_list_lock);
2195
2196         if (!list_empty(&oap->oap_rpc_item))
2197                 GOTO(out, rc = -EBUSY);
2198
2199         osc_exit_cache(cli, oap, 0);
2200         osc_wake_cache_waiters(cli);
2201
2202         if (!list_empty(&oap->oap_urgent_item)) {
2203                 list_del_init(&oap->oap_urgent_item);
2204                 oap->oap_async_flags &= ~ASYNC_URGENT;
2205         }
2206         if (!list_empty(&oap->oap_pending_item)) {
2207                 list_del_init(&oap->oap_pending_item);
2208                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2209         }
2210         loi_list_maint(cli, loi);
2211
2212         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2213 out:
2214         spin_unlock(&cli->cl_loi_list_lock);
2215         if (rc == 0)
2216                 OBD_FREE(oap, sizeof(*oap));
2217         RETURN(rc);
2218 }
2219
2220 #ifdef __KERNEL__
2221 /* Note: caller will lock/unlock, and set uptodate on the pages */
2222 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2223 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2224                            struct lov_stripe_md *lsm, obd_count page_count,
2225                            struct brw_page *pga)
2226 {
2227         struct ptlrpc_request *request = NULL;
2228         struct ost_body *body;
2229         struct niobuf_remote *nioptr;
2230         struct obd_ioobj *iooptr;
2231         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2232         int swab;
2233         ENTRY;
2234
2235         /* XXX does not handle 'new' brw protocol */
2236
2237         size[1] = sizeof(struct obd_ioobj);
2238         size[2] = page_count * sizeof(*nioptr);
2239
2240         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2241                                   OST_SAN_READ, 3, size, NULL);
2242         if (!request)
2243                 RETURN(-ENOMEM);
2244
2245         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2246         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2247         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2248                                 sizeof(*nioptr) * page_count);
2249
2250         memcpy(&body->oa, oa, sizeof(body->oa));
2251
2252         obdo_to_ioobj(oa, iooptr);
2253         iooptr->ioo_bufcnt = page_count;
2254
2255         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2256                 LASSERT(PageLocked(pga[mapped].pg));
2257                 LASSERT(mapped == 0 ||
2258                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2259
2260                 nioptr->offset = pga[mapped].disk_offset;
2261                 nioptr->len    = pga[mapped].count;
2262                 nioptr->flags  = pga[mapped].flag;
2263         }
2264
2265         size[1] = page_count * sizeof(*nioptr);
2266         request->rq_replen = lustre_msg_size(2, size);
2267
2268         rc = ptlrpc_queue_wait(request);
2269         if (rc)
2270                 GOTO(out_req, rc);
2271
2272         body = lustre_swab_repbuf(request, 0, sizeof(*body),
2273                                   lustre_swab_ost_body);
2274         if (body == NULL) {
2275                 CERROR("Can't unpack body\n");
2276                 GOTO(out_req, rc = -EPROTO);
2277         }
2278
2279         memcpy(oa, &body->oa, sizeof(*oa));
2280
2281         swab = lustre_msg_swabbed(request->rq_repmsg);
2282         LASSERT_REPSWAB(request, 1);
2283         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2284         if (!nioptr) {
2285                 /* nioptr missing or short */
2286                 GOTO(out_req, rc = -EPROTO);
2287         }
2288
2289         /* actual read */
2290         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2291                 struct page *page = pga[mapped].pg;
2292                 struct buffer_head *bh;
2293                 kdev_t dev;
2294
2295                 if (swab)
2296                         lustre_swab_niobuf_remote (nioptr);
2297
2298                 /* got san device associated */
2299                 LASSERT(exp->exp_obd != NULL);
2300                 dev = exp->exp_obd->u.cli.cl_sandev;
2301
2302                 /* hole */
2303                 if (!nioptr->offset) {
2304                         CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2305                                         page->mapping->host->i_ino,
2306                                         page->index);
2307                         memset(page_address(page), 0, PAGE_SIZE);
2308                         continue;
2309                 }
2310
2311                 if (!page->buffers) {
2312                         create_empty_buffers(page, dev, PAGE_SIZE);
2313                         bh = page->buffers;
2314
2315                         clear_bit(BH_New, &bh->b_state);
2316                         set_bit(BH_Mapped, &bh->b_state);
2317                         bh->b_blocknr = (unsigned long)nioptr->offset;
2318
2319                         clear_bit(BH_Uptodate, &bh->b_state);
2320
2321                         ll_rw_block(READ, 1, &bh);
2322                 } else {
2323                         bh = page->buffers;
2324
2325                         /* if buffer already existed, it must be the
2326                          * one we mapped before, check it */
2327                         LASSERT(!test_bit(BH_New, &bh->b_state));
2328                         LASSERT(test_bit(BH_Mapped, &bh->b_state));
2329                         LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2330
2331                         /* wait it's io completion */
2332                         if (test_bit(BH_Lock, &bh->b_state))
2333                                 wait_on_buffer(bh);
2334
2335                         if (!test_bit(BH_Uptodate, &bh->b_state))
2336                                 ll_rw_block(READ, 1, &bh);
2337                 }
2338
2339
2340                 /* must do syncronous write here */
2341                 wait_on_buffer(bh);
2342                 if (!buffer_uptodate(bh)) {
2343                         /* I/O error */
2344                         rc = -EIO;
2345                         goto out_req;
2346                 }
2347         }
2348
2349 out_req:
2350         ptlrpc_req_finished(request);
2351         RETURN(rc);
2352 }
2353
2354 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2355                             struct lov_stripe_md *lsm, obd_count page_count,
2356                             struct brw_page *pga)
2357 {
2358         struct ptlrpc_request *request = NULL;
2359         struct ost_body *body;
2360         struct niobuf_remote *nioptr;
2361         struct obd_ioobj *iooptr;
2362         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2363         int swab;
2364         ENTRY;
2365
2366         size[1] = sizeof(struct obd_ioobj);
2367         size[2] = page_count * sizeof(*nioptr);
2368
2369         request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2370                                   OST_SAN_WRITE, 3, size, NULL);
2371         if (!request)
2372                 RETURN(-ENOMEM);
2373
2374         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2375         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2376         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2377                                 sizeof (*nioptr) * page_count);
2378
2379         memcpy(&body->oa, oa, sizeof(body->oa));
2380
2381         obdo_to_ioobj(oa, iooptr);
2382         iooptr->ioo_bufcnt = page_count;
2383
2384         /* pack request */
2385         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2386                 LASSERT(PageLocked(pga[mapped].pg));
2387                 LASSERT(mapped == 0 ||
2388                         pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2389
2390                 nioptr->offset = pga[mapped].disk_offset;
2391                 nioptr->len    = pga[mapped].count;
2392                 nioptr->flags  = pga[mapped].flag;
2393         }
2394
2395         size[1] = page_count * sizeof(*nioptr);
2396         request->rq_replen = lustre_msg_size(2, size);
2397
2398         rc = ptlrpc_queue_wait(request);
2399         if (rc)
2400                 GOTO(out_req, rc);
2401
2402         swab = lustre_msg_swabbed (request->rq_repmsg);
2403         LASSERT_REPSWAB (request, 1);
2404         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2405         if (!nioptr) {
2406                 CERROR("absent/short niobuf array\n");
2407                 GOTO(out_req, rc = -EPROTO);
2408         }
2409
2410         /* actual write */
2411         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2412                 struct page *page = pga[mapped].pg;
2413                 struct buffer_head *bh;
2414                 kdev_t dev;
2415
2416                 if (swab)
2417                         lustre_swab_niobuf_remote (nioptr);
2418
2419                 /* got san device associated */
2420                 LASSERT(exp->exp_obd != NULL);
2421                 dev = exp->exp_obd->u.cli.cl_sandev;
2422
2423                 if (!page->buffers) {
2424                         create_empty_buffers(page, dev, PAGE_SIZE);
2425                 } else {
2426                         /* checking */
2427                         LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2428                         LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2429                         LASSERT(page->buffers->b_blocknr ==
2430                                 (unsigned long)nioptr->offset);
2431                 }
2432                 bh = page->buffers;
2433
2434                 LASSERT(bh);
2435
2436                 /* if buffer locked, wait it's io completion */
2437                 if (test_bit(BH_Lock, &bh->b_state))
2438                         wait_on_buffer(bh);
2439
2440                 clear_bit(BH_New, &bh->b_state);
2441                 set_bit(BH_Mapped, &bh->b_state);
2442
2443                 /* override the block nr */
2444                 bh->b_blocknr = (unsigned long)nioptr->offset;
2445
2446                 /* we are about to write it, so set it
2447                  * uptodate/dirty
2448                  * page lock should garentee no race condition here */
2449                 set_bit(BH_Uptodate, &bh->b_state);
2450                 set_bit(BH_Dirty, &bh->b_state);
2451
2452                 ll_rw_block(WRITE, 1, &bh);
2453
2454                 /* must do syncronous write here */
2455                 wait_on_buffer(bh);
2456                 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2457                         /* I/O error */
2458                         rc = -EIO;
2459                         goto out_req;
2460                 }
2461         }
2462
2463 out_req:
2464         ptlrpc_req_finished(request);
2465         RETURN(rc);
2466 }
2467
2468 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2469                       struct lov_stripe_md *lsm, obd_count page_count,
2470                       struct brw_page *pga, struct obd_trans_info *oti)
2471 {
2472         ENTRY;
2473
2474         while (page_count) {
2475                 obd_count pages_per_brw;
2476                 int rc;
2477
2478                 if (page_count > PTLRPC_MAX_BRW_PAGES)
2479                         pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2480                 else
2481                         pages_per_brw = page_count;
2482
2483                 if (cmd & OBD_BRW_WRITE)
2484                         rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2485                 else
2486                         rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2487
2488                 if (rc != 0)
2489                         RETURN(rc);
2490
2491                 page_count -= pages_per_brw;
2492                 pga += pages_per_brw;
2493         }
2494         RETURN(0);
2495 }
2496 #endif
2497 #endif
2498
2499 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2500 {
2501         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2502
2503         if (lock == NULL) {
2504                 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2505                 return;
2506         }
2507
2508         lock_res_and_lock(lock);
2509 #ifdef __KERNEL__
2510         if (lock->l_ast_data && lock->l_ast_data != data) {
2511                 struct inode *new_inode = data;
2512                 struct inode *old_inode = lock->l_ast_data;
2513                 if (!(old_inode->i_state & I_FREEING))
2514                         LDLM_ERROR(lock, "inconsistent l_ast_data found");
2515                 LASSERTF(old_inode->i_state & I_FREEING,
2516                          "Found existing inode %p/%lu/%u state %lu in lock: "
2517                          "setting data to %p/%lu/%u\n", old_inode,
2518                          old_inode->i_ino, old_inode->i_generation,
2519                          old_inode->i_state,
2520                          new_inode, new_inode->i_ino, new_inode->i_generation);
2521         }
2522 #endif
2523         lock->l_ast_data = data;
2524         unlock_res_and_lock(lock);
2525         LDLM_LOCK_PUT(lock);
2526 }
2527
2528 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2529                              ldlm_iterator_t replace, void *data)
2530 {
2531         struct ldlm_res_id res_id = { .name = {0} };
2532         struct obd_device *obd = class_exp2obd(exp);
2533
2534         res_id.name[0] = lsm->lsm_object_id;
2535         res_id.name[2] = lsm->lsm_object_gr;
2536         ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2537         return 0;
2538 }
2539
2540 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2541                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2542                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2543                        void *data, __u32 lvb_len, void *lvb_swabber,
2544                        struct lustre_handle *lockh)
2545 {
2546         struct obd_device *obd = exp->exp_obd;
2547         struct ldlm_res_id res_id = { .name = {0} };
2548         struct ost_lvb lvb;
2549         struct ldlm_reply *rep;
2550         struct ptlrpc_request *req = NULL;
2551         int rc;
2552         ENTRY;
2553
2554         res_id.name[0] = lsm->lsm_object_id;
2555         res_id.name[2] = lsm->lsm_object_gr;
2556
2557         /* Filesystem lock extents are extended to page boundaries so that
2558          * dealing with the page cache is a little smoother.  */
2559         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2560         policy->l_extent.end |= ~PAGE_MASK;
2561
2562         if (lsm->lsm_oinfo->loi_kms_valid == 0)
2563                 goto no_match;
2564
2565         /* Next, search for already existing extent locks that will cover us */
2566         rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2567                              lockh);
2568         if (rc == 1) {
2569                 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2570                         /* return immediately if no credential held */
2571                         ldlm_lock_decref(lockh, mode);
2572                         RETURN(-EACCES);
2573                 }
2574
2575                 osc_set_data_with_check(lockh, data);
2576                 if (*flags & LDLM_FL_HAS_INTENT) {
2577                         /* I would like to be able to ASSERT here that rss <=
2578                          * kms, but I can't, for reasons which are explained in
2579                          * lov_enqueue() */
2580                 }
2581                 /* We already have a lock, and it's referenced */
2582                 RETURN(ELDLM_OK);
2583         }
2584
2585         /* If we're trying to read, we also search for an existing PW lock.  The
2586          * VFS and page cache already protect us locally, so lots of readers/
2587          * writers can share a single PW lock.
2588          *
2589          * There are problems with conversion deadlocks, so instead of
2590          * converting a read lock to a write lock, we'll just enqueue a new
2591          * one.
2592          *
2593          * At some point we should cancel the read lock instead of making them
2594          * send us a blocking callback, but there are problems with canceling
2595          * locks out from other users right now, too. */
2596
2597         if (mode == LCK_PR) {
2598                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2599                                      policy, LCK_PW, lockh);
2600                 if (rc == 1) {
2601                         if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2602                                 /* return immediately if no credential held */
2603                                 ldlm_lock_decref(lockh, LCK_PW);
2604                                 RETURN(-EACCES);
2605                         }
2606
2607                         /* FIXME: This is not incredibly elegant, but it might
2608                          * be more elegant than adding another parameter to
2609                          * lock_match.  I want a second opinion. */
2610                         ldlm_lock_addref(lockh, LCK_PR);
2611                         ldlm_lock_decref(lockh, LCK_PW);
2612                         osc_set_data_with_check(lockh, data);
2613                         RETURN(ELDLM_OK);
2614                 }
2615         }
2616         if (mode == LCK_PW) {
2617                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2618                                      policy, LCK_PR, lockh);
2619                 if (rc == 1) {
2620                         rc = ldlm_cli_convert(lockh, mode, flags);
2621                         if (!rc) {
2622                                 /* Update readers/writers accounting */
2623                                 ldlm_lock_addref(lockh, LCK_PW);
2624                                 ldlm_lock_decref(lockh, LCK_PR);
2625                                 osc_set_data_with_check(lockh, data);
2626                                 RETURN(ELDLM_OK);
2627                         }
2628                         /* If the conversion failed, we need to drop refcount
2629                            on matched lock before we get new one */
2630                         /* XXX Won't it save us some efforts if we cancel PR
2631                            lock here? We are going to take PW lock anyway and it
2632                            will invalidate PR lock */
2633                         ldlm_lock_decref(lockh, LCK_PR);
2634                         if (rc != EDEADLOCK) {
2635                                 RETURN(rc);
2636                         }
2637                 }
2638         }
2639
2640 no_match:
2641         if (*flags & LDLM_FL_HAS_INTENT) {
2642                 int size[2] = {0, sizeof(struct ldlm_request)};
2643
2644                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
2645                                       LDLM_ENQUEUE, 2, size, NULL);
2646                 if (req == NULL)
2647                         RETURN(-ENOMEM);
2648
2649                 size[0] = sizeof(*rep);
2650                 size[1] = sizeof(lvb);
2651                 req->rq_replen = lustre_msg_size(2, size);
2652         }
2653         rc = ldlm_cli_enqueue(exp, req, obd->obd_namespace, res_id, type,
2654                               policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2655                               &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2656         if (req != NULL) {
2657                 if (rc == ELDLM_LOCK_ABORTED) {
2658                         /* swabbed by ldlm_cli_enqueue() */
2659                         LASSERT_REPSWABBED(req, 0);
2660                         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
2661                         LASSERT(rep != NULL);
2662                         if (rep->lock_policy_res1)
2663                                 rc = rep->lock_policy_res1;
2664                 }
2665                 ptlrpc_req_finished(req);
2666         }
2667
2668         if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2669                 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2670                        lvb.lvb_size, lvb.lvb_blocks);
2671                 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2672                 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2673         }
2674
2675         RETURN(rc);
2676 }
2677
2678 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2679                      __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2680                      int *flags, void *data, struct lustre_handle *lockh)
2681 {
2682         struct ldlm_res_id res_id = { .name = {0} };
2683         struct obd_device *obd = exp->exp_obd;
2684         int rc;
2685         ENTRY;
2686
2687         res_id.name[0] = lsm->lsm_object_id;
2688         res_id.name[2] = lsm->lsm_object_gr;
2689
2690         OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2691
2692         /* Filesystem lock extents are extended to page boundaries so that
2693          * dealing with the page cache is a little smoother */
2694         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2695         policy->l_extent.end |= ~PAGE_MASK;
2696
2697         /* Next, search for already existing extent locks that will cover us */
2698         rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2699                              policy, mode, lockh);
2700         if (rc) {
2701                // if (!(*flags & LDLM_FL_TEST_LOCK))
2702                         osc_set_data_with_check(lockh, data);
2703                 RETURN(rc);
2704         }
2705         /* If we're trying to read, we also search for an existing PW lock.  The
2706          * VFS and page cache already protect us locally, so lots of readers/
2707          * writers can share a single PW lock. */
2708         if (mode == LCK_PR) {
2709                 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2710                                      policy, LCK_PW, lockh);
2711                 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2712                         /* FIXME: This is not incredibly elegant, but it might
2713                          * be more elegant than adding another parameter to
2714                          * lock_match.  I want a second opinion. */
2715                         osc_set_data_with_check(lockh, data);
2716                         ldlm_lock_addref(lockh, LCK_PR);
2717                         ldlm_lock_decref(lockh, LCK_PW);
2718                 }
2719         }
2720         RETURN(rc);
2721 }
2722
2723 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2724                       __u32 mode, struct lustre_handle *lockh)
2725 {
2726         ENTRY;
2727
2728         if (mode == LCK_GROUP)
2729                 ldlm_lock_decref_and_cancel(lockh, mode);
2730         else
2731                 ldlm_lock_decref(lockh, mode);
2732
2733         RETURN(0);
2734 }
2735
2736 static int osc_cancel_unused(struct obd_export *exp,
2737                              struct lov_stripe_md *lsm,
2738                              int flags, void *opaque)
2739 {
2740         struct obd_device *obd = class_exp2obd(exp);
2741         struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2742
2743         if (lsm != NULL) {
2744                 res_id.name[0] = lsm->lsm_object_id;
2745                 res_id.name[2] = lsm->lsm_object_gr;
2746                 resp = &res_id;
2747         }
2748
2749         return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2750 }
2751
2752 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2753                       unsigned long max_age)
2754 {
2755         struct obd_statfs *msfs;
2756         struct ptlrpc_request *request;
2757         int rc, size = sizeof(*osfs);
2758         ENTRY;
2759
2760         /* We could possibly pass max_age in the request (as an absolute
2761          * timestamp or a "seconds.usec ago") so the target can avoid doing
2762          * extra calls into the filesystem if that isn't necessary (e.g.
2763          * during mount that would help a bit).  Having relative timestamps
2764          * is not so great if request processing is slow, while absolute
2765          * timestamps are not ideal because they need time synchronization. */
2766         request = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OBD_VERSION,
2767                                   OST_STATFS, 0, NULL, NULL);
2768         if (!request)
2769                 RETURN(-ENOMEM);
2770
2771         request->rq_replen = lustre_msg_size(1, &size);
2772         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2773
2774         rc = ptlrpc_queue_wait(request);
2775         if (rc)
2776                 GOTO(out, rc);
2777
2778         msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2779                                   lustre_swab_obd_statfs);
2780         if (msfs == NULL) {
2781                 CERROR("Can't unpack obd_statfs\n");
2782                 GOTO(out, rc = -EPROTO);
2783         }
2784
2785         memcpy(osfs, msfs, sizeof(*osfs));
2786
2787         EXIT;
2788  out:
2789         ptlrpc_req_finished(request);
2790         return rc;
2791 }
2792
2793 /* Retrieve object striping information.
2794  *
2795  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2796  * the maximum number of OST indices which will fit in the user buffer.
2797  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2798  */
2799 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2800 {
2801         struct lov_user_md lum, *lumk;
2802         int rc, lum_size;
2803         ENTRY;
2804
2805         if (!lsm)
2806                 RETURN(-ENODATA);
2807
2808         rc = copy_from_user(&lum, lump, sizeof(lum));
2809         if (rc)
2810                 RETURN(-EFAULT);
2811
2812         if (lum.lmm_magic != LOV_USER_MAGIC)
2813                 RETURN(-EINVAL);
2814
2815         if (lum.lmm_stripe_count > 0) {
2816                 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2817                 OBD_ALLOC(lumk, lum_size);
2818                 if (!lumk)
2819                         RETURN(-ENOMEM);
2820
2821                 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2822                 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2823         } else {
2824                 lum_size = sizeof(lum);
2825                 lumk = &lum;
2826         }
2827
2828         lumk->lmm_object_id = lsm->lsm_object_id;
2829         lumk->lmm_object_gr = lsm->lsm_object_gr;
2830         lumk->lmm_stripe_count = 1;
2831
2832         if (copy_to_user(lump, lumk, lum_size))
2833                 rc = -EFAULT;
2834
2835         if (lumk != &lum)
2836                 OBD_FREE(lumk, lum_size);
2837
2838         RETURN(rc);
2839 }
2840
2841 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2842                          void *karg, void *uarg)
2843 {
2844         struct obd_device *obd = exp->exp_obd;
2845         struct obd_ioctl_data *data = karg;
2846         int err = 0;
2847         ENTRY;
2848
2849 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2850         MOD_INC_USE_COUNT;
2851 #else
2852         if (!try_module_get(THIS_MODULE)) {
2853                 CERROR("Can't get module. Is it alive?");
2854                 return -EINVAL;
2855         }
2856 #endif
2857         switch (cmd) {
2858         case OBD_IOC_LOV_GET_CONFIG: {
2859                 char *buf;
2860                 struct lov_desc *desc;
2861                 struct obd_uuid uuid;
2862
2863                 buf = NULL;
2864                 len = 0;
2865                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2866                         GOTO(out, err = -EINVAL);
2867
2868                 data = (struct obd_ioctl_data *)buf;
2869
2870                 if (sizeof(*desc) > data->ioc_inllen1) {
2871                         OBD_FREE(buf, len);
2872                         GOTO(out, err = -EINVAL);
2873                 }
2874
2875                 if (data->ioc_inllen2 < sizeof(uuid)) {
2876                         OBD_FREE(buf, len);
2877                         GOTO(out, err = -EINVAL);
2878                 }
2879
2880                 if (data->ioc_inllen3 < sizeof(__u32)) {
2881                         OBD_FREE(buf, len);
2882                         GOTO(out, err = -EINVAL);
2883                 }
2884
2885                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2886                 desc->ld_tgt_count = 1;
2887                 desc->ld_active_tgt_count = 1;
2888                 desc->ld_default_stripe_count = 1;
2889                 desc->ld_default_stripe_size = 0;
2890                 desc->ld_default_stripe_offset = 0;
2891                 desc->ld_pattern = 0;
2892                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2893                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2894                 *((__u32 *)data->ioc_inlbuf3) = 1;
2895
2896                 err = copy_to_user((void *)uarg, buf, len);
2897                 if (err)
2898                         err = -EFAULT;
2899                 obd_ioctl_freedata(buf, len);
2900                 GOTO(out, err);
2901         }
2902         case LL_IOC_LOV_SETSTRIPE:
2903                 err = obd_alloc_memmd(exp, karg);
2904                 if (err > 0)
2905                         err = 0;
2906                 GOTO(out, err);
2907         case LL_IOC_LOV_GETSTRIPE:
2908                 err = osc_getstripe(karg, uarg);
2909                 GOTO(out, err);
2910         case OBD_IOC_CLIENT_RECOVER:
2911                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2912                                             data->ioc_inlbuf1);
2913                 if (err > 0)
2914                         err = 0;
2915                 GOTO(out, err);
2916         case IOC_OSC_SET_ACTIVE:
2917                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2918                                                data->ioc_offset);
2919                 GOTO(out, err);
2920         case IOC_OSC_CTL_RECOVERY:
2921                 err = ptlrpc_import_control_recovery(obd->u.cli.cl_import,
2922                                                      data->ioc_offset);
2923                 GOTO(out, err);
2924         default:
2925                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2926                 GOTO(out, err = -ENOTTY);
2927         }
2928 out:
2929 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2930         MOD_DEC_USE_COUNT;
2931 #else
2932         module_put(THIS_MODULE);
2933 #endif
2934         return err;
2935 }
2936
2937 static int osc_get_info(struct obd_export *exp, __u32 keylen,
2938                         void *key, __u32 *vallen, void *val)
2939 {
2940         ENTRY;
2941         if (!vallen || !val)
2942                 RETURN(-EFAULT);
2943
2944         if (keylen > strlen("lock_to_stripe") &&
2945             strcmp(key, "lock_to_stripe") == 0) {
2946                 __u32 *stripe = val;
2947                 *vallen = sizeof(*stripe);
2948                 *stripe = 0;
2949                 RETURN(0);
2950         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2951                 struct ptlrpc_request *req;
2952                 obd_id *reply;
2953                 char *bufs[1] = {key};
2954                 int rc;
2955                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2956                                       OST_GET_INFO, 1, (int *)&keylen, bufs);
2957                 if (req == NULL)
2958                         RETURN(-ENOMEM);
2959
2960                 req->rq_replen = lustre_msg_size(1, (int *)vallen);
2961                 rc = ptlrpc_queue_wait(req);
2962                 if (rc)
2963                         GOTO(out, rc);
2964
2965                 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2966                                            lustre_swab_ost_last_id);
2967                 if (reply == NULL) {
2968                         CERROR("Can't unpack OST last ID\n");
2969                         GOTO(out, rc = -EPROTO);
2970                 }
2971                 *((obd_id *)val) = *reply;
2972         out:
2973                 ptlrpc_req_finished(req);
2974                 RETURN(rc);
2975         } else if (keylen == 10 && strcmp(key, "client_nid") == 0) {
2976                 struct ptlrpc_connection * conn;
2977                 ptl_nid_t * nid = val;
2978                 ptl_process_id_t id;
2979                 int rc;
2980
2981                 *vallen = sizeof(*nid);
2982                 conn = class_exp2cliimp(exp)->imp_connection;
2983
2984                 if (!conn || !conn->c_peer.peer_ni)
2985                         RETURN(-ENOTCONN);
2986
2987                 rc = PtlGetId(conn->c_peer.peer_ni->pni_ni_h, &id);
2988                 if (rc == PTL_OK)
2989                         *nid = id.nid;
2990
2991                 RETURN(0);
2992         }
2993         RETURN(-EPROTO);
2994 }
2995
2996 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2997                         void *key, obd_count vallen, void *val)
2998 {
2999         struct obd_device  *obd = exp->exp_obd;
3000         struct obd_import *imp = class_exp2cliimp(exp);
3001         struct llog_ctxt *ctxt;
3002         int rc = 0;
3003         ENTRY;
3004
3005         if (keylen == strlen("unlinked") &&
3006             memcmp(key, "unlinked", keylen) == 0) {
3007                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3008                 spin_lock(&oscc->oscc_lock);
3009                 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3010                 spin_unlock(&oscc->oscc_lock);
3011                 RETURN(0);
3012         }
3013
3014         if (keylen == strlen("unrecovery") &&
3015             memcmp(key, "unrecovery", keylen) == 0) {
3016                 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3017                 spin_lock(&oscc->oscc_lock);
3018                 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3019                 spin_unlock(&oscc->oscc_lock);
3020                 RETURN(0);
3021         }
3022
3023         if (keylen == strlen("initial_recov") &&
3024             memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
3025                 struct obd_import *imp = class_exp2cliimp(exp);
3026                 if (vallen != sizeof(int))
3027                         RETURN(-EINVAL);
3028                 imp->imp_initial_recov = *(int *)val;
3029                 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
3030                        exp->exp_obd->obd_name,
3031                        imp->imp_initial_recov);
3032                 RETURN(0);
3033         }
3034
3035         if (keylen == strlen("async") &&
3036             memcmp(key, "async", keylen) == 0) {
3037                 struct client_obd *cl = &obd->u.cli;
3038                 if (vallen != sizeof(int))
3039                         RETURN(-EINVAL);
3040                 cl->cl_async = *(int *)val;
3041                 CDEBUG(D_HA, "%s: set async = %d\n",
3042                        obd->obd_name, cl->cl_async);
3043                 RETURN(0);
3044         }
3045
3046         if (keylen == 5 && strcmp(key, "audit") == 0) {
3047                 struct ptlrpc_request *req;
3048                 char *bufs[2] = {key, val};
3049                 int size[2] = {keylen, vallen};
3050
3051                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3052                                       OST_SET_INFO, 2, size, bufs);
3053                 if (req == NULL)
3054                         RETURN(-ENOMEM);
3055
3056                 req->rq_replen = lustre_msg_size(0, size);
3057                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
3058                                    lustre_swab_audit_attr);
3059                 rc = ptlrpc_queue_wait(req);
3060
3061                 ptlrpc_req_finished(req);
3062                 RETURN(rc);
3063         }
3064
3065         if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
3066                 struct ptlrpc_request *req;
3067                 char *bufs[2] = {key, val};
3068                 int size[2] = {keylen, vallen};
3069
3070                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3071                                       OST_SET_INFO, 2, size, bufs);
3072                 if (req == NULL)
3073                         RETURN(-ENOMEM);
3074
3075                 req->rq_replen = lustre_msg_size(0, size);
3076                 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
3077                                    lustre_swab_obdo);
3078                 rc = ptlrpc_queue_wait(req);
3079
3080                 ptlrpc_req_finished(req);
3081                 RETURN(rc);
3082         }
3083
3084         if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
3085                 struct ptlrpc_request *req;
3086                 char *bufs[2] = {key, val};
3087                 int size[2] = {keylen, vallen};
3088
3089                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3090                                       OST_SET_INFO, 2, size, bufs);
3091                 if (req == NULL)
3092                         RETURN(-ENOMEM);
3093
3094                 req->rq_replen = lustre_msg_size(0, size);
3095                 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
3096                                    lustre_swab_audit_msg);
3097                 rc = ptlrpc_queue_wait(req);
3098
3099                 ptlrpc_req_finished(req);
3100                 RETURN(rc);
3101         }
3102
3103         if (keylen == strlen("sec") && memcmp(key, "sec", keylen) == 0) {
3104                 struct client_obd *cli = &exp->exp_obd->u.cli;
3105
3106                 cli->cl_sec_flavor = ptlrpcs_name2flavor(val);
3107                 if (cli->cl_sec_flavor == PTLRPCS_FLVR_INVALID) {
3108                         CERROR("unrecognized security flavor %s\n", (char*) val);
3109                         RETURN(-EINVAL);
3110                 }
3111
3112                 RETURN(0);
3113         }
3114
3115         if (keylen == strlen("sec_flags") &&
3116             memcmp(key, "sec_flags", keylen) == 0) {
3117                 struct client_obd *cli = &exp->exp_obd->u.cli;
3118
3119                 cli->cl_sec_flags = *((unsigned long *) val);
3120                 RETURN(0);
3121         }
3122
3123         if (keylen == strlen("flush_cred") &&
3124             memcmp(key, "flush_cred", keylen) == 0) {
3125                 struct client_obd *cli = &exp->exp_obd->u.cli;
3126
3127                 if (cli->cl_import)
3128                         ptlrpcs_import_flush_current_creds(cli->cl_import);
3129                 RETURN(0);
3130         }
3131         if (keylen == strlen("crypto_cb") &&
3132             memcmp(key, "crypto_cb", keylen) == 0) {
3133                 LASSERT(vallen == sizeof(crypt_cb_t));
3134                 osc_crypt_cb = (crypt_cb_t)val;
3135                 RETURN(0);
3136         }
3137
3138         if (keylen == 8 && memcmp(key, "capa_key", 8) == 0) {
3139                 struct ptlrpc_request *req;
3140                 char *bufs[2] = {key, val};
3141                 unsigned long irqflags;
3142                 int rc, size[2] = {keylen, vallen};
3143
3144                 LASSERT(vallen == sizeof(struct lustre_capa_key));
3145
3146                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3147                                       OST_SET_INFO, 2, size, bufs);
3148                 if (req == NULL)
3149                         RETURN(-ENOMEM);
3150
3151                 spin_lock_irqsave (&req->rq_lock, irqflags);
3152                 req->rq_replay = 1;
3153                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
3154
3155                 req->rq_replen = lustre_msg_size(0, NULL);
3156                 rc = ptlrpc_queue_wait(req);
3157                 ptlrpc_req_finished(req);
3158                 RETURN(rc);
3159         }
3160
3161         if (keylen == strlen("setext") &&
3162             memcmp(key, "setext", keylen) == 0) {
3163                 struct client_obd *cli = &exp->exp_obd->u.cli;
3164                 struct osc_creator *oscc = &cli->cl_oscc;
3165                 struct fid_extent *ext = val;
3166
3167                 oscc->oscc_next_id = (obd_id)ext->fe_start;
3168                 RETURN(0);
3169         }
3170
3171         if (keylen < strlen("mds_conn") ||
3172             memcmp(key, "mds_conn", keylen) != 0)
3173                 RETURN(-EINVAL);
3174
3175         ctxt = llog_get_context(&exp->exp_obd->obd_llogs,
3176                                 LLOG_UNLINK_ORIG_CTXT);
3177         if (ctxt) {
3178                 if (rc == 0)
3179                         rc = llog_initiator_connect(ctxt);
3180                 else
3181                         CERROR("cannot establish the connect for "
3182                                "ctxt %p: %d\n", ctxt, rc);
3183         }
3184
3185         imp->imp_server_timeout = 1;
3186         CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
3187         imp->imp_pingable = 1;
3188
3189         RETURN(rc);
3190 }
3191
3192
3193 static struct llog_operations osc_size_repl_logops = {
3194         lop_cancel: llog_obd_repl_cancel
3195 };
3196
3197 static struct llog_operations osc_unlink_orig_logops;
3198
3199 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
3200                          struct obd_device *tgt, int count,
3201                          struct llog_catid *catid)
3202 {
3203         int rc;
3204         ENTRY;
3205
3206         osc_unlink_orig_logops = llog_lvfs_ops;
3207         osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
3208         osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
3209         osc_unlink_orig_logops.lop_add = llog_catalog_add;
3210         osc_unlink_orig_logops.lop_connect = llog_origin_connect;
3211
3212         rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
3213                             &catid->lci_logid, &osc_unlink_orig_logops);
3214         if (rc)
3215                 RETURN(rc);
3216
3217         rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
3218                             &osc_size_repl_logops);
3219         RETURN(rc);
3220 }
3221
3222 static int osc_llog_finish(struct obd_device *obd,
3223                            struct obd_llogs *llogs, int count)
3224 {
3225         int rc;
3226         ENTRY;
3227
3228         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
3229         if (rc)
3230                 RETURN(rc);
3231
3232         rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
3233         RETURN(rc);
3234 }
3235
3236 static int osc_connect(struct lustre_handle *exph,
3237                        struct obd_device *obd, struct obd_uuid *cluuid,
3238                        struct obd_connect_data *data,
3239                        unsigned long connect_flags)
3240 {
3241         int rc;
3242         ENTRY;
3243         rc = client_connect_import(exph, obd, cluuid, data, connect_flags);
3244         RETURN(rc);
3245 }
3246
3247 static int osc_disconnect(struct obd_export *exp, unsigned long flags)
3248 {
3249         struct obd_device *obd = class_exp2obd(exp);
3250         struct llog_ctxt *ctxt;
3251         int rc;
3252         ENTRY;
3253
3254         ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
3255         if (obd->u.cli.cl_conn_count == 1)
3256                 /* flush any remaining cancel messages out to the target */
3257                 llog_sync(ctxt, exp);
3258
3259         rc = client_disconnect_export(exp, flags);
3260         RETURN(rc);
3261 }
3262
3263 static int osc_import_event(struct obd_device *obd,
3264                             struct obd_import *imp,
3265                             enum obd_import_event event)
3266 {
3267         struct client_obd *cli;
3268         int rc = 0;
3269         ENTRY;
3270
3271         LASSERT(imp->imp_obd == obd);
3272
3273         switch (event) {
3274         case IMP_EVENT_DISCON: {
3275                 /* Only do this on the MDS OSC's */
3276                 if (imp->imp_server_timeout) {
3277                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3278
3279                         spin_lock(&oscc->oscc_lock);
3280                         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3281                         spin_unlock(&oscc->oscc_lock);
3282                 }
3283                 break;
3284         }
3285         case IMP_EVENT_INACTIVE: {
3286                 if (obd->obd_observer)
3287                         rc = obd_notify(obd->obd_observer, obd, 0, 0);
3288                 break;
3289         }
3290         case IMP_EVENT_INVALIDATE: {
3291                 struct ldlm_namespace *ns = obd->obd_namespace;
3292
3293                 /* Reset grants */
3294                 cli = &obd->u.cli;
3295                 spin_lock(&cli->cl_loi_list_lock);
3296                 cli->cl_avail_grant = 0;
3297                 cli->cl_lost_grant = 0;
3298                 /* all pages go to failing rpcs due to the invalid import */
3299                 osc_check_rpcs(cli);
3300                 spin_unlock(&cli->cl_loi_list_lock);
3301
3302                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3303
3304                 break;
3305         }
3306         case IMP_EVENT_ACTIVE: {
3307                 /* Only do this on the MDS OSC's */
3308                 if (imp->imp_server_timeout) {
3309                         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3310
3311                         spin_lock(&oscc->oscc_lock);
3312                         oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3313                         spin_unlock(&oscc->oscc_lock);
3314                 }
3315
3316                 if (obd->obd_observer)
3317                         rc = obd_notify(obd->obd_observer, obd, 1, 0);
3318                 break;
3319         }
3320         default:
3321                 CERROR("Unknown import event %d\n", event);
3322                 LBUG();
3323         }
3324         RETURN(rc);
3325 }
3326
3327 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
3328 {
3329         struct lprocfs_static_vars lvars;
3330         int rc;
3331         ENTRY;
3332
3333         lprocfs_init_vars(osc,&lvars);
3334         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3335         if (rc < 0)
3336                 RETURN(rc);
3337
3338         rc = lproc_osc_attach_seqstat(dev);
3339         if (rc < 0) {
3340                 lprocfs_obd_detach(dev);
3341                 RETURN(rc);
3342         }
3343
3344         ptlrpc_lprocfs_register_obd(dev);
3345         RETURN(0);
3346 }
3347
3348 static int osc_detach(struct obd_device *dev)
3349 {
3350         ptlrpc_lprocfs_unregister_obd(dev);
3351         return lprocfs_obd_detach(dev);
3352 }
3353
3354 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3355 {
3356         int rc;
3357         ENTRY;
3358         rc = ptlrpcd_addref();
3359         if (rc)
3360                 RETURN(rc);
3361
3362         rc = client_obd_setup(obd, len, buf);
3363         if (rc)
3364                 ptlrpcd_decref();
3365         else
3366                 oscc_init(obd);
3367
3368         RETURN(rc);
3369 }
3370
3371 static int osc_cleanup(struct obd_device *obd, int flags)
3372 {
3373         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3374         int rc;
3375
3376         rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3377                                     LDLM_FL_CONFIG_CHANGE, NULL);
3378         if (rc)
3379                 RETURN(rc);
3380
3381         spin_lock(&oscc->oscc_lock);
3382         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3383         oscc->oscc_flags |= OSCC_FLAG_EXITING;
3384         spin_unlock(&oscc->oscc_lock);
3385
3386         rc = client_obd_cleanup(obd, flags);
3387         ptlrpcd_decref();
3388         RETURN(rc);
3389 }
3390
3391
3392 struct obd_ops osc_obd_ops = {
3393         .o_owner                = THIS_MODULE,
3394         .o_attach               = osc_attach,
3395         .o_detach               = osc_detach,
3396         .o_setup                = osc_setup,
3397         .o_cleanup              = osc_cleanup,
3398         .o_add_conn             = client_import_add_conn,
3399         .o_del_conn             = client_import_del_conn,
3400         .o_connect              = osc_connect,
3401         .o_disconnect           = osc_disconnect,
3402         .o_statfs               = osc_statfs,
3403         .o_packmd               = osc_packmd,
3404         .o_unpackmd             = osc_unpackmd,
3405         .o_create               = osc_create,
3406         .o_destroy              = osc_destroy,
3407         .o_getattr              = osc_getattr,
3408         .o_getattr_async        = osc_getattr_async,
3409         .o_setattr              = osc_setattr,
3410         .o_brw                  = osc_brw,
3411         .o_brw_async            = osc_brw_async,
3412         .o_prep_async_page      = osc_prep_async_page,
3413         .o_queue_async_io       = osc_queue_async_io,
3414         .o_set_async_flags      = osc_set_async_flags,
3415         .o_queue_group_io       = osc_queue_group_io,
3416         .o_trigger_group_io     = osc_trigger_group_io,
3417         .o_teardown_async_page  = osc_teardown_async_page,
3418         .o_punch                = osc_punch,
3419         .o_sync                 = osc_sync,
3420         .o_enqueue              = osc_enqueue,
3421         .o_match                = osc_match,
3422         .o_change_cbdata        = osc_change_cbdata,
3423         .o_cancel               = osc_cancel,
3424         .o_cancel_unused        = osc_cancel_unused,
3425         .o_iocontrol            = osc_iocontrol,
3426         .o_get_info             = osc_get_info,
3427         .o_set_info             = osc_set_info,
3428         .o_import_event         = osc_import_event,
3429         .o_llog_init            = osc_llog_init,
3430         .o_llog_finish          = osc_llog_finish,
3431 };
3432
3433 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3434 struct obd_ops sanosc_obd_ops = {
3435         .o_owner                = THIS_MODULE,
3436         .o_attach               = osc_attach,
3437         .o_detach               = osc_detach,
3438         .o_cleanup              = client_obd_cleanup,
3439         .o_add_conn             = client_import_add_conn,
3440         .o_del_conn             = client_import_del_conn,
3441         .o_connect              = osc_connect,
3442         .o_disconnect           = client_disconnect_export,
3443         .o_statfs               = osc_statfs,
3444         .o_packmd               = osc_packmd,
3445         .o_unpackmd             = osc_unpackmd,
3446         .o_create               = osc_real_create,
3447         .o_destroy              = osc_destroy,
3448         .o_getattr              = osc_getattr,
3449         .o_getattr_async        = osc_getattr_async,
3450         .o_setattr              = osc_setattr,
3451         .o_setup                = client_sanobd_setup,
3452         .o_brw                  = sanosc_brw,
3453         .o_punch                = osc_punch,
3454         .o_sync                 = osc_sync,
3455         .o_enqueue              = osc_enqueue,
3456         .o_match                = osc_match,
3457         .o_change_cbdata        = osc_change_cbdata,
3458         .o_cancel               = osc_cancel,
3459         .o_cancel_unused        = osc_cancel_unused,
3460         .o_iocontrol            = osc_iocontrol,
3461         .o_import_event         = osc_import_event,
3462         .o_llog_init            = osc_llog_init,
3463         .o_llog_finish          = osc_llog_finish,
3464 };
3465 #endif
3466
3467 int __init osc_init(void)
3468 {
3469         struct lprocfs_static_vars lvars;
3470 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3471         struct lprocfs_static_vars sanlvars;
3472 #endif
3473         int rc;
3474         ENTRY;
3475
3476         lprocfs_init_vars(osc, &lvars);
3477 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3478         lprocfs_init_vars(osc, &sanlvars);
3479 #endif
3480
3481         rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3482                                  OBD_OSC_DEVICENAME);
3483         if (rc)
3484                 RETURN(rc);
3485
3486 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3487         rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3488                                  OBD_SANOSC_DEVICENAME);
3489         if (rc)
3490                 class_unregister_type(OBD_OSC_DEVICENAME);
3491 #endif
3492
3493         RETURN(rc);
3494 }
3495
3496 #ifdef __KERNEL__
3497 static void /*__exit*/ osc_exit(void)
3498 {
3499 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3500         class_unregister_type(OBD_SANOSC_DEVICENAME);
3501 #endif
3502         class_unregister_type(OBD_OSC_DEVICENAME);
3503 }
3504
3505 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3506 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3507 MODULE_LICENSE("GPL");
3508
3509 module_init(osc_init);
3510 module_exit(osc_exit);
3511 #endif