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/lustre_dlm.h>
  40 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  41 #  include <linux/workqueue.h>
  42 #  include <linux/smp_lock.h>
  43 # else
  44 #  include <linux/locks.h>
  45 # endif
  46 #else /* __KERNEL__ */
  47 # include <liblustre.h>
  48 #endif
  49
  50 #include <linux/kp30.h>
  51 #include <linux/lustre_net.h>
  52 #include <linux/lustre_user.h>
  53 #include <linux/obd_ost.h>
  54 #include <linux/obd_lov.h>
  55
  56 #ifndef  __CYGWIN__
  57 # include <linux/ctype.h>
  58 # include <linux/init.h>
  59 #else
  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 "osc_internal.h"
  67
  68
  69 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
  70 {
  71         struct lprocfs_static_vars lvars;
  72         int rc;
  73         ENTRY;
  74
  75         lprocfs_init_vars(osc,&lvars);
  76         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
  77         if (rc < 0)
  78                 RETURN(rc);
  79
  80         rc = lproc_osc_attach_seqstat(dev);
  81         if (rc < 0) {
  82                 lprocfs_obd_detach(dev);
  83                 RETURN(rc);
  84         }
  85
  86         ptlrpc_lprocfs_register_obd(dev);
  87         RETURN(0);
  88 }
  89
  90 static int osc_detach(struct obd_device *dev)
  91 {
  92         ptlrpc_lprocfs_unregister_obd(dev);
  93         return lprocfs_obd_detach(dev);
  94 }
  95
  96
  97 /* Pack OSC object metadata for disk storage (LE byte order). */
  98 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
  99                       struct lov_stripe_md *lsm)
 100 {
 101         int lmm_size;
 102         ENTRY;
 103
 104         lmm_size = sizeof(**lmmp);
 105         if (!lmmp)
 106                 RETURN(lmm_size);
 107
 108         if (*lmmp && !lsm) {
 109                 OBD_FREE(*lmmp, lmm_size);
 110                 *lmmp = NULL;
 111                 RETURN(0);
 112         }
 113
 114         if (!*lmmp) {
 115                 OBD_ALLOC(*lmmp, lmm_size);
 116                 if (!*lmmp)
 117                         RETURN(-ENOMEM);
 118         }
 119
 120         if (lsm) {
 121                 LASSERT(lsm->lsm_object_id);
 122                 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
 123         }
 124
 125         RETURN(lmm_size);
 126 }
 127
 128 /* Unpack OSC object metadata from disk storage (LE byte order). */
 129 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
 130                         struct lov_mds_md *lmm, int lmm_bytes)
 131 {
 132         int lsm_size;
 133         ENTRY;
 134
 135         if (lmm != NULL) {
 136                 if (lmm_bytes < sizeof (*lmm)) {
 137                         CERROR("lov_mds_md too small: %d, need %d\n",
 138                                lmm_bytes, (int)sizeof(*lmm));
 139                         RETURN(-EINVAL);
 140                 }
 141                 /* XXX LOV_MAGIC etc check? */
 142
 143                 if (lmm->lmm_object_id == 0) {
 144                         CERROR("lov_mds_md: zero lmm_object_id\n");
 145                         RETURN(-EINVAL);
 146                 }
 147         }
 148
 149         lsm_size = lov_stripe_md_size(1);
 150         if (lsmp == NULL)
 151                 RETURN(lsm_size);
 152
 153         if (*lsmp != NULL && lmm == NULL) {
 154                 OBD_FREE(*lsmp, lsm_size);
 155                 *lsmp = NULL;
 156                 RETURN(0);
 157         }
 158
 159         if (*lsmp == NULL) {
 160                 OBD_ALLOC(*lsmp, lsm_size);
 161                 if (*lsmp == NULL)
 162                         RETURN(-ENOMEM);
 163                 loi_init((*lsmp)->lsm_oinfo);
 164         }
 165
 166         if (lmm != NULL) {
 167                 /* XXX zero *lsmp? */
 168                 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
 169                 LASSERT((*lsmp)->lsm_object_id);
 170         }
 171
 172         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
 173
 174         RETURN(lsm_size);
 175 }
 176
 177 static int osc_getattr_interpret(struct ptlrpc_request *req,
 178                                  struct osc_getattr_async_args *aa, int rc)
 179 {
 180         struct ost_body *body;
 181         ENTRY;
 182
 183         if (rc != 0)
 184                 RETURN(rc);
 185
 186         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 187         if (body) {
 188                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 189                 memcpy(aa->aa_oa, &body->oa, sizeof(*aa->aa_oa));
 190
 191                 /* This should really be sent by the OST */
 192                 aa->aa_oa->o_blksize = OSC_BRW_MAX_SIZE;
 193                 aa->aa_oa->o_valid |= OBD_MD_FLBLKSZ;
 194         } else {
 195                 CERROR("can't unpack ost_body\n");
 196                 rc = -EPROTO;
 197                 aa->aa_oa->o_valid = 0;
 198         }
 199
 200         RETURN(rc);
 201 }
 202
 203 static int osc_getattr_async(struct obd_export *exp, struct obdo *oa,
 204                              struct lov_stripe_md *md,
 205                              struct ptlrpc_request_set *set)
 206 {
 207         struct ptlrpc_request *request;
 208         struct ost_body *body;
 209         int size = sizeof(*body);
 210         struct osc_getattr_async_args *aa;
 211         ENTRY;
 212
 213         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GETATTR, 1,
 214                                   &size, NULL);
 215         if (!request)
 216                 RETURN(-ENOMEM);
 217
 218         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 219         memcpy(&body->oa, oa, sizeof(*oa));
 220
 221         request->rq_replen = lustre_msg_size(1, &size);
 222         request->rq_interpret_reply = osc_getattr_interpret;
 223
 224         LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
 225         aa = (struct osc_getattr_async_args *)&request->rq_async_args;
 226         aa->aa_oa = oa;
 227
 228         ptlrpc_set_add_req (set, request);
 229         RETURN (0);
 230 }
 231
 232 static int osc_getattr(struct obd_export *exp, struct obdo *oa,
 233                        struct lov_stripe_md *md)
 234 {
 235         struct ptlrpc_request *request;
 236         struct ost_body *body;
 237         int rc, size = sizeof(*body);
 238         ENTRY;
 239
 240         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GETATTR, 1,
 241                                   &size, NULL);
 242         if (!request)
 243                 RETURN(-ENOMEM);
 244
 245         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 246         memcpy(&body->oa, oa, sizeof(*oa));
 247
 248         request->rq_replen = lustre_msg_size(1, &size);
 249
 250         rc = ptlrpc_queue_wait(request);
 251         if (rc) {
 252                 CERROR("%s failed: rc = %d\n", __FUNCTION__, rc);
 253                 GOTO(out, rc);
 254         }
 255
 256         body = lustre_swab_repbuf(request, 0, sizeof (*body),
 257                                   lustre_swab_ost_body);
 258         if (body == NULL) {
 259                 CERROR ("can't unpack ost_body\n");
 260                 GOTO (out, rc = -EPROTO);
 261         }
 262
 263         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 264         memcpy(oa, &body->oa, sizeof(*oa));
 265
 266         /* This should really be sent by the OST */
 267         oa->o_blksize = OSC_BRW_MAX_SIZE;
 268         oa->o_valid |= OBD_MD_FLBLKSZ;
 269
 270         EXIT;
 271  out:
 272         ptlrpc_req_finished(request);
 273         return rc;
 274 }
 275
 276 static int osc_setattr(struct obd_export *exp, struct obdo *oa,
 277                        struct lov_stripe_md *md, struct obd_trans_info *oti)
 278 {
 279         struct ptlrpc_request *request;
 280         struct ost_body *body;
 281         int rc, size = sizeof(*body);
 282         ENTRY;
 283
 284         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SETATTR, 1, &size,
 285                                   NULL);
 286         if (!request)
 287                 RETURN(-ENOMEM);
 288
 289         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
 290         memcpy(&body->oa, oa, sizeof(*oa));
 291
 292         request->rq_replen = lustre_msg_size(1, &size);
 293
 294         rc = ptlrpc_queue_wait(request);
 295         if (rc)
 296                 GOTO(out, rc);
 297
 298         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 299                                   lustre_swab_ost_body);
 300         if (body == NULL)
 301                 GOTO(out, rc = -EPROTO);
 302
 303         memcpy(oa, &body->oa, sizeof(*oa));
 304
 305         EXIT;
 306 out:
 307         ptlrpc_req_finished(request);
 308         RETURN(0);
 309 }
 310
 311 int osc_real_create(struct obd_export *exp, struct obdo *oa,
 312                     struct lov_stripe_md **ea, struct obd_trans_info *oti)
 313 {
 314         struct ptlrpc_request *request;
 315         struct ost_body *body;
 316         struct lov_stripe_md *lsm;
 317         int rc, size = sizeof(*body);
 318         ENTRY;
 319
 320         LASSERT(oa);
 321         LASSERT(ea);
 322
 323         lsm = *ea;
 324         if (!lsm) {
 325                 rc = obd_alloc_memmd(exp, &lsm);
 326                 if (rc < 0)
 327                         RETURN(rc);
 328         }
 329
 330         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_CREATE, 1, &size,
 331                                   NULL);
 332         if (!request)
 333                 GOTO(out, rc = -ENOMEM);
 334
 335         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 336         memcpy(&body->oa, oa, sizeof(body->oa));
 337
 338         request->rq_replen = lustre_msg_size(1, &size);
 339         if (oa->o_valid & OBD_MD_FLINLINE) {
 340                 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
 341                         oa->o_flags == OBD_FL_DELORPHAN);
 342                 DEBUG_REQ(D_HA, request,
 343                           "delorphan from OST integration; level == RECOVER");
 344                 request->rq_send_state = LUSTRE_IMP_RECOVER;
 345         }
 346
 347         rc = ptlrpc_queue_wait(request);
 348         if (rc)
 349                 GOTO(out_req, rc);
 350
 351         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 352                                   lustre_swab_ost_body);
 353         if (body == NULL) {
 354                 CERROR ("can't unpack ost_body\n");
 355                 GOTO (out_req, rc = -EPROTO);
 356         }
 357
 358         memcpy(oa, &body->oa, sizeof(*oa));
 359
 360         /* This should really be sent by the OST */
 361         oa->o_blksize = OSC_BRW_MAX_SIZE;
 362         oa->o_valid |= OBD_MD_FLBLKSZ;
 363
 364         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
 365          * have valid lsm_oinfo data structs, so don't go touching that.
 366          * This needs to be fixed in a big way.
 367          */
 368         lsm->lsm_object_id = oa->o_id;
 369         *ea = lsm;
 370
 371         if (oti != NULL) {
 372                 oti->oti_transno = request->rq_repmsg->transno;
 373
 374                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
 375                         if (!oti->oti_logcookies)
 376                                 oti_alloc_cookies(oti, 1);
 377                         memcpy(oti->oti_logcookies, obdo_logcookie(oa),
 378                                sizeof(oti->oti_onecookie));
 379                 }
 380         }
 381
 382         CDEBUG(D_HA, "transno: "LPD64"\n", request->rq_repmsg->transno);
 383         EXIT;
 384 out_req:
 385         ptlrpc_req_finished(request);
 386 out:
 387         if (rc && !*ea)
 388                 obd_free_memmd(exp, &lsm);
 389         return rc;
 390 }
 391
 392 static int osc_punch(struct obd_export *exp, struct obdo *oa,
 393                      struct lov_stripe_md *md, obd_size start,
 394                      obd_size end, struct obd_trans_info *oti)
 395 {
 396         struct ptlrpc_request *request;
 397         struct ost_body *body;
 398         int rc, size = sizeof(*body);
 399         ENTRY;
 400
 401         if (!oa) {
 402                 CERROR("oa NULL\n");
 403                 RETURN(-EINVAL);
 404         }
 405
 406         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_PUNCH, 1, &size,
 407                                   NULL);
 408         if (!request)
 409                 RETURN(-ENOMEM);
 410
 411         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 412         memcpy(&body->oa, oa, sizeof(*oa));
 413
 414         /* overload the size and blocks fields in the oa with start/end */
 415         body->oa.o_size = start;
 416         body->oa.o_blocks = end;
 417         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 418
 419         request->rq_replen = lustre_msg_size(1, &size);
 420
 421         rc = ptlrpc_queue_wait(request);
 422         if (rc)
 423                 GOTO(out, rc);
 424
 425         body = lustre_swab_repbuf (request, 0, sizeof (*body),
 426                                    lustre_swab_ost_body);
 427         if (body == NULL) {
 428                 CERROR ("can't unpack ost_body\n");
 429                 GOTO (out, rc = -EPROTO);
 430         }
 431
 432         memcpy(oa, &body->oa, sizeof(*oa));
 433
 434         EXIT;
 435  out:
 436         ptlrpc_req_finished(request);
 437         return rc;
 438 }
 439
 440 static int osc_sync(struct obd_export *exp, struct obdo *oa,
 441                     struct lov_stripe_md *md, obd_size start, obd_size end)
 442 {
 443         struct ptlrpc_request *request;
 444         struct ost_body *body;
 445         int rc, size = sizeof(*body);
 446         ENTRY;
 447
 448         if (!oa) {
 449                 CERROR("oa NULL\n");
 450                 RETURN(-EINVAL);
 451         }
 452
 453         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SYNC, 1, &size,
 454                                   NULL);
 455         if (!request)
 456                 RETURN(-ENOMEM);
 457
 458         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 459         memcpy(&body->oa, oa, sizeof(*oa));
 460
 461         /* overload the size and blocks fields in the oa with start/end */
 462         body->oa.o_size = start;
 463         body->oa.o_blocks = end;
 464         body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
 465
 466         request->rq_replen = lustre_msg_size(1, &size);
 467
 468         rc = ptlrpc_queue_wait(request);
 469         if (rc)
 470                 GOTO(out, rc);
 471
 472         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 473                                   lustre_swab_ost_body);
 474         if (body == NULL) {
 475                 CERROR ("can't unpack ost_body\n");
 476                 GOTO (out, rc = -EPROTO);
 477         }
 478
 479         memcpy(oa, &body->oa, sizeof(*oa));
 480
 481         EXIT;
 482  out:
 483         ptlrpc_req_finished(request);
 484         return rc;
 485 }
 486
 487 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
 488                        struct lov_stripe_md *ea, struct obd_trans_info *oti)
 489 {
 490         struct ptlrpc_request *request;
 491         struct ost_body *body;
 492         int rc, size = sizeof(*body);
 493         ENTRY;
 494
 495         if (!oa) {
 496                 CERROR("oa NULL\n");
 497                 RETURN(-EINVAL);
 498         }
 499
 500         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_DESTROY, 1,
 501                                   &size, NULL);
 502         if (!request)
 503                 RETURN(-ENOMEM);
 504
 505         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
 506
 507         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
 508                 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
 509                        sizeof(*oti->oti_logcookies));
 510                 oti->oti_logcookies++;
 511         }
 512
 513         memcpy(&body->oa, oa, sizeof(*oa));
 514         request->rq_replen = lustre_msg_size(1, &size);
 515
 516         rc = ptlrpc_queue_wait(request);
 517         if (rc)
 518                 GOTO(out, rc);
 519
 520         body = lustre_swab_repbuf(request, 0, sizeof(*body),
 521                                   lustre_swab_ost_body);
 522         if (body == NULL) {
 523                 CERROR ("Can't unpack body\n");
 524                 GOTO (out, rc = -EPROTO);
 525         }
 526
 527         memcpy(oa, &body->oa, sizeof(*oa));
 528
 529         EXIT;
 530  out:
 531         ptlrpc_req_finished(request);
 532         return rc;
 533 }
 534
 535 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 536                                 long writing_bytes)
 537 {
 538         obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
 539
 540         LASSERT(!(oa->o_valid & bits));
 541
 542         oa->o_valid |= bits;
 543         spin_lock(&cli->cl_loi_list_lock);
 544         oa->o_dirty = cli->cl_dirty;
 545         oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
 546         oa->o_grant = cli->cl_avail_grant;
 547         oa->o_dropped = cli->cl_lost_grant;
 548         cli->cl_lost_grant = 0;
 549         spin_unlock(&cli->cl_loi_list_lock);
 550         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
 551                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 552 }
 553
 554 /* caller must hold loi_list_lock */
 555 static void osc_consume_write_grant(struct client_obd *cli,
 556                                     struct osc_async_page *oap)
 557 {
 558         cli->cl_dirty += PAGE_SIZE;
 559         cli->cl_avail_grant -= PAGE_SIZE;
 560         oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
 561         CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
 562         LASSERT(cli->cl_avail_grant >= 0);
 563 }
 564
 565 /* caller must hold loi_list_lock */
 566 void osc_wake_cache_waiters(struct client_obd *cli)
 567 {
 568         struct list_head *l, *tmp;
 569         struct osc_cache_waiter *ocw;
 570
 571         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 572                 /* if we can't dirty more, we must wait until some is written */
 573                 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
 574                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
 575                                cli->cl_dirty, cli->cl_dirty_max);
 576                         return;
 577                 }
 578
 579                 /* if still dirty cache but no grant wait for pending RPCs that
 580                  * may yet return us some grant before doing sync writes */
 581                 if (cli->cl_brw_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
 582                         CDEBUG(D_CACHE, "%d BRWs in flight, no grant\n",
 583                                cli->cl_brw_in_flight);
 584                         return;
 585                 }
 586
 587                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
 588                 list_del_init(&ocw->ocw_entry);
 589                 if (cli->cl_avail_grant < PAGE_SIZE) {
 590                         /* no more RPCs in flight to return grant, do sync IO */
 591                         ocw->ocw_rc = -EDQUOT;
 592                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
 593                 } else {
 594                         osc_consume_write_grant(cli, ocw->ocw_oap);
 595                 }
 596                 wake_up(&ocw->ocw_waitq);
 597         }
 598
 599         EXIT;
 600 }
 601
 602 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 603 {
 604         spin_lock(&cli->cl_loi_list_lock);
 605         CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
 606         cli->cl_avail_grant += body->oa.o_grant;
 607         /* waiters are woken in brw_interpret_oap */
 608         spin_unlock(&cli->cl_loi_list_lock);
 609 }
 610
 611 /* We assume that the reason this OSC got a short read is because it read
 612  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
 613  * via the LOV, and it _knows_ it's reading inside the file, it's just that
 614  * this stripe never got written at or beyond this stripe offset yet. */
 615 static void handle_short_read(int nob_read, obd_count page_count,
 616                               struct brw_page *pga)
 617 {
 618         char *ptr;
 619
 620         /* skip bytes read OK */
 621         while (nob_read > 0) {
 622                 LASSERT (page_count > 0);
 623
 624                 if (pga->count > nob_read) {
 625                         /* EOF inside this page */
 626                         ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
 627                         memset(ptr + nob_read, 0, pga->count - nob_read);
 628                         kunmap(pga->pg);
 629                         page_count--;
 630                         pga++;
 631                         break;
 632                 }
 633
 634                 nob_read -= pga->count;
 635                 page_count--;
 636                 pga++;
 637         }
 638
 639         /* zero remaining pages */
 640         while (page_count-- > 0) {
 641                 ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
 642                 memset(ptr, 0, pga->count);
 643                 kunmap(pga->pg);
 644                 pga++;
 645         }
 646 }
 647
 648 static int check_write_rcs(struct ptlrpc_request *request,
 649                            int requested_nob, int niocount,
 650                            obd_count page_count, struct brw_page *pga)
 651 {
 652         int    *remote_rcs, i;
 653
 654         /* return error if any niobuf was in error */
 655         remote_rcs = lustre_swab_repbuf(request, 1,
 656                                         sizeof(*remote_rcs) * niocount, NULL);
 657         if (remote_rcs == NULL) {
 658                 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
 659                 return(-EPROTO);
 660         }
 661         if (lustre_msg_swabbed(request->rq_repmsg))
 662                 for (i = 0; i < niocount; i++)
 663                         __swab32s(&remote_rcs[i]);
 664
 665         for (i = 0; i < niocount; i++) {
 666                 if (remote_rcs[i] < 0)
 667                         return(remote_rcs[i]);
 668
 669                 if (remote_rcs[i] != 0) {
 670                         CERROR("rc[%d] invalid (%d) req %p\n",
 671                                 i, remote_rcs[i], request);
 672                         return(-EPROTO);
 673                 }
 674         }
 675
 676         if (request->rq_bulk->bd_nob_transferred != requested_nob) {
 677                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
 678                        requested_nob, request->rq_bulk->bd_nob_transferred);
 679                 return(-EPROTO);
 680         }
 681
 682         return (0);
 683 }
 684
 685 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
 686 {
 687         if (p1->flag != p2->flag) {
 688                 unsigned mask = ~OBD_BRW_FROM_GRANT;
 689
 690                 /* warn if we try to combine flags that we don't know to be
 691                  * safe to combine */
 692                 if ((p1->flag & mask) != (p2->flag & mask))
 693                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
 694                                "same brw?\n", p1->flag, p2->flag);
 695                 return 0;
 696         }
 697
 698         return (p1->off + p1->count == p2->off);
 699 }
 700
 701 #if CHECKSUM_BULK
 702 static obd_count cksum_pages(int nob, obd_count page_count,
 703                              struct brw_page *pga)
 704 {
 705         obd_count cksum = 0;
 706         char *ptr;
 707
 708         while (nob > 0) {
 709                 LASSERT (page_count > 0);
 710
 711                 ptr = kmap(pga->pg);
 712                 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
 713                              pga->count > nob ? nob : pga->count);
 714                 kunmap(pga->pg);
 715
 716                 nob -= pga->count;
 717                 page_count--;
 718                 pga++;
 719         }
 720
 721         return (cksum);
 722 }
 723 #endif
 724
 725 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
 726                                 struct lov_stripe_md *lsm, obd_count page_count,
 727                                 struct brw_page *pga, int *requested_nobp,
 728                                 int *niocountp, struct ptlrpc_request **reqp)
 729 {
 730         struct ptlrpc_request   *req;
 731         struct ptlrpc_bulk_desc *desc;
 732         struct client_obd       *cli = &imp->imp_obd->u.cli;
 733         struct ost_body         *body;
 734         struct obd_ioobj        *ioobj;
 735         struct niobuf_remote    *niobuf;
 736         unsigned long            flags;
 737         int                      niocount;
 738         int                      size[3];
 739         int                      i;
 740         int                      requested_nob;
 741         int                      opc;
 742         int                      rc;
 743
 744         opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
 745
 746         for (niocount = i = 1; i < page_count; i++)
 747                 if (!can_merge_pages(&pga[i - 1], &pga[i]))
 748                         niocount++;
 749
 750         size[0] = sizeof(*body);
 751         size[1] = sizeof(*ioobj);
 752         size[2] = niocount * sizeof(*niobuf);
 753
 754         req = ptlrpc_prep_req(imp, opc, 3, size, NULL);
 755         if (req == NULL)
 756                 return (-ENOMEM);
 757
 758         if (opc == OST_WRITE)
 759                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 760                                              BULK_GET_SOURCE, OST_BULK_PORTAL);
 761         else
 762                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 763                                              BULK_PUT_SINK, OST_BULK_PORTAL);
 764         if (desc == NULL)
 765                 GOTO(out, rc = -ENOMEM);
 766         /* NB request now owns desc and will free it when it gets freed */
 767
 768         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 769         ioobj = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*ioobj));
 770         niobuf = lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf));
 771
 772         memcpy(&body->oa, oa, sizeof(*oa));
 773
 774         obdo_to_ioobj(oa, ioobj);
 775         ioobj->ioo_bufcnt = niocount;
 776
 777         LASSERT (page_count > 0);
 778         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
 779                 struct brw_page *pg = &pga[i];
 780                 struct brw_page *pg_prev = pg - 1;
 781
 782                 LASSERT(pg->count > 0);
 783                 LASSERT((pg->off & ~PAGE_MASK) + pg->count <= PAGE_SIZE);
 784                 LASSERTF(i == 0 || pg->off > pg_prev->off,
 785                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
 786                          " prev_pg %p [pri %lu ind %lu] off "LPU64,
 787                          i, page_count,
 788                          pg->pg, pg->pg->private, pg->pg->index, pg->off,
 789                          pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
 790                                  pg_prev->off);
 791
 792                 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~PAGE_MASK,
 793                                       pg->count);
 794                 requested_nob += pg->count;
 795
 796                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
 797                         niobuf--;
 798                         niobuf->len += pg->count;
 799                 } else {
 800                         niobuf->offset = pg->off;
 801                         niobuf->len    = pg->count;
 802                         niobuf->flags  = pg->flag;
 803                 }
 804         }
 805
 806         LASSERT((void *)(niobuf - niocount) ==
 807                 lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf)));
 808         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
 809         spin_lock_irqsave(&req->rq_lock, flags);
 810         req->rq_no_resend = 1;
 811         spin_unlock_irqrestore(&req->rq_lock, flags);
 812
 813         /* size[0] still sizeof (*body) */
 814         if (opc == OST_WRITE) {
 815 #if CHECKSUM_BULK
 816                 body->oa.o_valid |= OBD_MD_FLCKSUM;
 817                 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
 818 #endif
 819                 /* 1 RC per niobuf */
 820                 size[1] = sizeof(__u32) * niocount;
 821                 req->rq_replen = lustre_msg_size(2, size);
 822         } else {
 823                 /* 1 RC for the whole I/O */
 824                 req->rq_replen = lustre_msg_size(1, size);
 825         }
 826
 827         *niocountp = niocount;
 828         *requested_nobp = requested_nob;
 829         *reqp = req;
 830         return (0);
 831
 832  out:
 833         ptlrpc_req_finished (req);
 834         return (rc);
 835 }
 836
 837 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
 838                                 int requested_nob, int niocount,
 839                                 obd_count page_count, struct brw_page *pga,
 840                                 int rc)
 841 {
 842         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 843         struct ost_body *body;
 844         ENTRY;
 845
 846         if (rc < 0)
 847                 RETURN(rc);
 848
 849         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 850         if (body == NULL) {
 851                 CERROR ("Can't unpack body\n");
 852                 RETURN(-EPROTO);
 853         }
 854
 855         osc_update_grant(cli, body);
 856
 857         if (req->rq_reqmsg->opc == OST_WRITE) {
 858                 if (rc > 0) {
 859                         CERROR ("Unexpected +ve rc %d\n", rc);
 860                         RETURN(-EPROTO);
 861                 }
 862                 LASSERT (req->rq_bulk->bd_nob == requested_nob);
 863
 864                 RETURN(check_write_rcs(req, requested_nob, niocount,
 865                                        page_count, pga));
 866         }
 867
 868         if (rc > requested_nob) {
 869                 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
 870                 RETURN(-EPROTO);
 871         }
 872
 873         if (rc != req->rq_bulk->bd_nob_transferred) {
 874                 CERROR ("Unexpected rc %d (%d transferred)\n",
 875                         rc, req->rq_bulk->bd_nob_transferred);
 876                 return (-EPROTO);
 877         }
 878
 879         if (rc < requested_nob)
 880                 handle_short_read(rc, page_count, pga);
 881
 882         memcpy(oa, &body->oa, sizeof(*oa));
 883
 884 #if CHECKSUM_BULK
 885         if (oa->o_valid & OBD_MD_FLCKSUM) {
 886                 const struct ptlrpc_peer *peer =
 887                         &req->rq_import->imp_connection->c_peer;
 888                 static int cksum_counter;
 889                 obd_count server_cksum = oa->o_cksum;
 890                 obd_count cksum = cksum_pages(rc, page_count, pga);
 891                 char str[PTL_NALFMT_SIZE];
 892
 893                 portals_nid2str(peer->peer_ni->pni_number, peer->peer_nid, str);
 894
 895                 cksum_counter++;
 896                 if (server_cksum != cksum) {
 897                         CERROR("Bad checksum: server %x, client %x, server NID "
 898                                LPX64" (%s)\n", server_cksum, cksum,
 899                                peer->peer_nid, str);
 900                         cksum_counter = 0;
 901                         oa->o_cksum = cksum;
 902                 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
 903                         CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
 904                               cksum_counter, peer->peer_nid, str, cksum);
 905                 }
 906         } else {
 907                 static int cksum_missed;
 908
 909                 cksum_missed++;
 910                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
 911                         CERROR("Request checksum %u from "LPX64", no reply\n",
 912                                cksum_missed,
 913                                req->rq_import->imp_connection->c_peer.peer_nid);
 914         }
 915 #endif
 916         RETURN(0);
 917 }
 918
 919 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
 920                             struct lov_stripe_md *lsm,
 921                             obd_count page_count, struct brw_page *pga)
 922 {
 923         int                    requested_nob;
 924         int                    niocount;
 925         struct ptlrpc_request *request;
 926         int                    rc;
 927         ENTRY;
 928
 929 restart_bulk:
 930         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
 931                                   page_count, pga, &requested_nob, &niocount,
 932                                   &request);
 933         /* NB ^ sets rq_no_resend */
 934
 935         if (rc != 0)
 936                 return (rc);
 937
 938         rc = ptlrpc_queue_wait(request);
 939
 940         if (rc == -ETIMEDOUT && request->rq_resend) {
 941                 DEBUG_REQ(D_HA, request,  "BULK TIMEOUT");
 942                 ptlrpc_req_finished(request);
 943                 goto restart_bulk;
 944         }
 945
 946         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
 947                                   page_count, pga, rc);
 948
 949         ptlrpc_req_finished(request);
 950         RETURN (rc);
 951 }
 952
 953 static int brw_interpret(struct ptlrpc_request *request,
 954                          struct osc_brw_async_args *aa, int rc)
 955 {
 956         struct obdo *oa      = aa->aa_oa;
 957         int requested_nob    = aa->aa_requested_nob;
 958         int niocount         = aa->aa_nio_count;
 959         obd_count page_count = aa->aa_page_count;
 960         struct brw_page *pga = aa->aa_pga;
 961         ENTRY;
 962
 963         /* XXX bug 937 here */
 964         if (rc == -ETIMEDOUT && request->rq_resend) {
 965                 DEBUG_REQ(D_HA, request,  "BULK TIMEOUT");
 966                 LBUG(); /* re-send.  later. */
 967                 //goto restart_bulk;
 968         }
 969
 970         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
 971                                   page_count, pga, rc);
 972         RETURN (rc);
 973 }
 974
 975 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
 976                           struct lov_stripe_md *lsm, obd_count page_count,
 977                           struct brw_page *pga, struct ptlrpc_request_set *set)
 978 {
 979         struct ptlrpc_request     *request;
 980         int                        requested_nob;
 981         int                        nio_count;
 982         struct osc_brw_async_args *aa;
 983         int                        rc;
 984         ENTRY;
 985
 986         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
 987                                   page_count, pga, &requested_nob, &nio_count,
 988                                   &request);
 989         /* NB ^ sets rq_no_resend */
 990
 991         if (rc == 0) {
 992                 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
 993                 aa = (struct osc_brw_async_args *)&request->rq_async_args;
 994                 aa->aa_oa = oa;
 995                 aa->aa_requested_nob = requested_nob;
 996                 aa->aa_nio_count = nio_count;
 997                 aa->aa_page_count = page_count;
 998                 aa->aa_pga = pga;
 999
1000                 request->rq_interpret_reply = brw_interpret;
1001                 ptlrpc_set_add_req(set, request);
1002         }
1003         RETURN (rc);
1004 }
1005
1006 #ifndef min_t
1007 #define min_t(type,x,y) \
1008         ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
1009 #endif
1010
1011 /*
1012  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1013  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1014  * fine for our small page arrays and doesn't require allocation.  its an
1015  * insertion sort that swaps elements that are strides apart, shrinking the
1016  * stride down until its '1' and the array is sorted.
1017  */
1018 static void sort_brw_pages(struct brw_page *array, int num)
1019 {
1020         int stride, i, j;
1021         struct brw_page tmp;
1022
1023         if (num == 1)
1024                 return;
1025         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1026                 ;
1027
1028         do {
1029                 stride /= 3;
1030                 for (i = stride ; i < num ; i++) {
1031                         tmp = array[i];
1032                         j = i;
1033                         while (j >= stride && array[j - stride].off > tmp.off) {
1034                                 array[j] = array[j - stride];
1035                                 j -= stride;
1036                         }
1037                         array[j] = tmp;
1038                 }
1039         } while (stride > 1);
1040 }
1041
1042 /* make sure we the regions we're passing to elan don't violate its '4
1043  * fragments' constraint.  portal headers are a fragment, all full
1044  * PAGE_SIZE long pages count as 1 fragment, and each partial page
1045  * counts as a fragment.  I think.  see bug 934. */
1046 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1047 {
1048         int frags_left = 3;
1049         int saw_whole_frag = 0;
1050         int i;
1051
1052         for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1053                 if (pg->count == PAGE_SIZE) {
1054                         if (!saw_whole_frag) {
1055                                 saw_whole_frag = 1;
1056                                 frags_left--;
1057                         }
1058                 } else {
1059                         frags_left--;
1060                 }
1061         }
1062         return i;
1063 }
1064
1065 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1066                    struct lov_stripe_md *md, obd_count page_count,
1067                    struct brw_page *pga, struct obd_trans_info *oti)
1068 {
1069         ENTRY;
1070
1071         if (cmd == OBD_BRW_CHECK) {
1072                 /* The caller just wants to know if there's a chance that this
1073                  * I/O can succeed */
1074                 struct obd_import *imp = class_exp2cliimp(exp);
1075
1076                 if (imp == NULL || imp->imp_invalid)
1077                         RETURN(-EIO);
1078                 RETURN(0);
1079         }
1080
1081         while (page_count) {
1082                 obd_count pages_per_brw;
1083                 int rc;
1084
1085                 if (page_count > OSC_BRW_MAX_IOV)
1086                         pages_per_brw = OSC_BRW_MAX_IOV;
1087                 else
1088                         pages_per_brw = page_count;
1089
1090                 sort_brw_pages(pga, pages_per_brw);
1091                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1092
1093                 rc = osc_brw_internal(cmd, exp, oa, md, pages_per_brw, pga);
1094
1095                 if (rc != 0)
1096                         RETURN(rc);
1097
1098                 page_count -= pages_per_brw;
1099                 pga += pages_per_brw;
1100         }
1101         RETURN(0);
1102 }
1103
1104 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1105                          struct lov_stripe_md *md, obd_count page_count,
1106                          struct brw_page *pga, struct ptlrpc_request_set *set,
1107                          struct obd_trans_info *oti)
1108 {
1109         ENTRY;
1110
1111         if (cmd == OBD_BRW_CHECK) {
1112                 /* The caller just wants to know if there's a chance that this
1113                  * I/O can succeed */
1114                 struct obd_import *imp = class_exp2cliimp(exp);
1115
1116                 if (imp == NULL || imp->imp_invalid)
1117                         RETURN(-EIO);
1118                 RETURN(0);
1119         }
1120
1121         while (page_count) {
1122                 obd_count pages_per_brw;
1123                 int rc;
1124
1125                 if (page_count > OSC_BRW_MAX_IOV)
1126                         pages_per_brw = OSC_BRW_MAX_IOV;
1127                 else
1128                         pages_per_brw = page_count;
1129
1130                 sort_brw_pages(pga, pages_per_brw);
1131                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1132
1133                 rc = async_internal(cmd, exp, oa, md, pages_per_brw, pga, set);
1134
1135                 if (rc != 0)
1136                         RETURN(rc);
1137
1138                 page_count -= pages_per_brw;
1139                 pga += pages_per_brw;
1140         }
1141         RETURN(0);
1142 }
1143
1144 static void osc_check_rpcs(struct client_obd *cli);
1145 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1146                            int sent);
1147 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1148 static void lop_update_pending(struct client_obd *cli,
1149                                struct loi_oap_pages *lop, int cmd, int delta);
1150
1151 /* this is called when a sync waiter receives an interruption.  Its job is to
1152  * get the caller woken as soon as possible.  If its page hasn't been put in an
1153  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
1154  * desiring interruption which will forcefully complete the rpc once the rpc
1155  * has timed out */
1156 static void osc_occ_interrupted(struct osic_callback_context *occ)
1157 {
1158         struct osc_async_page *oap;
1159         struct loi_oap_pages *lop;
1160         struct lov_oinfo *loi;
1161         ENTRY;
1162
1163         /* XXX member_of() */
1164         oap = list_entry(occ, struct osc_async_page, oap_occ);
1165
1166         spin_lock(&oap->oap_cli->cl_loi_list_lock);
1167
1168         oap->oap_interrupted = 1;
1169
1170         /* ok, it's been put in an rpc. */
1171         if (oap->oap_request != NULL) {
1172                 ptlrpc_mark_interrupted(oap->oap_request);
1173                 ptlrpcd_wake();
1174                 GOTO(unlock, 0);
1175         }
1176
1177         /* we don't get interruption callbacks until osc_trigger_sync_io()
1178          * has been called and put the sync oaps in the pending/urgent lists.*/
1179         if (!list_empty(&oap->oap_pending_item)) {
1180                 list_del_init(&oap->oap_pending_item);
1181                 if (oap->oap_async_flags & ASYNC_URGENT)
1182                         list_del_init(&oap->oap_urgent_item);
1183
1184                 loi = oap->oap_loi;
1185                 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1186                         &loi->loi_write_lop : &loi->loi_read_lop;
1187                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1188                 loi_list_maint(oap->oap_cli, oap->oap_loi);
1189
1190                 osic_complete_one(oap->oap_osic, &oap->oap_occ, 0);
1191                 oap->oap_osic = NULL;
1192         }
1193
1194 unlock:
1195         spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1196 }
1197
1198 /* this must be called holding the loi list lock to give coverage to exit_cache,
1199  * async_flag maintenance, and oap_request */
1200 static void osc_complete_oap(struct client_obd *cli,
1201                              struct osc_async_page *oap, int sent, int rc)
1202 {
1203         osc_exit_cache(cli, oap, sent);
1204         oap->oap_async_flags = 0;
1205         oap->oap_interrupted = 0;
1206
1207         if (oap->oap_request != NULL) {
1208                 ptlrpc_req_finished(oap->oap_request);
1209                 oap->oap_request = NULL;
1210         }
1211
1212         if (oap->oap_osic) {
1213                 osic_complete_one(oap->oap_osic, &oap->oap_occ, rc);
1214                 oap->oap_osic = NULL;
1215                 EXIT;
1216                 return;
1217         }
1218
1219         oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1220                                            rc);
1221 }
1222
1223 static int brw_interpret_oap(struct ptlrpc_request *request,
1224                              struct osc_brw_async_args *aa, int rc)
1225 {
1226         struct osc_async_page *oap;
1227         struct client_obd *cli;
1228         struct list_head *pos, *n;
1229         ENTRY;
1230
1231         CDEBUG(D_INODE, "request %p aa %p\n", request, aa);
1232
1233         rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1234                                   aa->aa_nio_count, aa->aa_page_count,
1235                                   aa->aa_pga, rc);
1236
1237         cli = aa->aa_cli;
1238         /* in failout recovery we ignore writeback failure and want
1239          * to just tell llite to unlock the page and continue */
1240         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1241                 rc = 0;
1242
1243         spin_lock(&cli->cl_loi_list_lock);
1244
1245         /* We need to decrement before osc_complete_oap->osc_wake_cache_waiters
1246          * is called so we know whether to go to sync BRWs or wait for more
1247          * RPCs to complete */
1248         cli->cl_brw_in_flight--;
1249
1250         /* the caller may re-use the oap after the completion call so
1251          * we need to clean it up a little */
1252         list_for_each_safe(pos, n, &aa->aa_oaps) {
1253                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1254
1255                 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1256                        //oap->oap_page, oap->oap_page->index, oap);
1257
1258                 list_del_init(&oap->oap_rpc_item);
1259                 osc_complete_oap(cli, oap, 1, rc);
1260         }
1261
1262         osc_wake_cache_waiters(cli);
1263         osc_check_rpcs(cli);
1264
1265         spin_unlock(&cli->cl_loi_list_lock);
1266
1267         obdo_free(aa->aa_oa);
1268         OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1269
1270         RETURN(0);
1271 }
1272
1273 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1274                                             struct list_head *rpc_list,
1275                                             int page_count, int cmd)
1276 {
1277         struct ptlrpc_request *req;
1278         struct brw_page *pga = NULL;
1279         int requested_nob, nio_count;
1280         struct osc_brw_async_args *aa;
1281         struct obdo *oa = NULL;
1282         struct obd_async_page_ops *ops = NULL;
1283         void *caller_data = NULL;
1284         struct list_head *pos;
1285         int i, rc;
1286
1287         LASSERT(!list_empty(rpc_list));
1288
1289         OBD_ALLOC(pga, sizeof(*pga) * page_count);
1290         if (pga == NULL)
1291                 RETURN(ERR_PTR(-ENOMEM));
1292
1293         oa = obdo_alloc();
1294         if (oa == NULL)
1295                 GOTO(out, req = ERR_PTR(-ENOMEM));
1296
1297         i = 0;
1298         list_for_each(pos, rpc_list) {
1299                 struct osc_async_page *oap;
1300
1301                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1302                 if (ops == NULL) {
1303                         ops = oap->oap_caller_ops;
1304                         caller_data = oap->oap_caller_data;
1305                 }
1306                 pga[i].off = oap->oap_obj_off + oap->oap_page_off;
1307                 pga[i].pg = oap->oap_page;
1308                 pga[i].count = oap->oap_count;
1309                 pga[i].flag = oap->oap_brw_flags;
1310                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1311                        pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1312                 i++;
1313         }
1314
1315         /* always get the data for the obdo for the rpc */
1316         LASSERT(ops != NULL);
1317         ops->ap_fill_obdo(caller_data, cmd, oa);
1318
1319         sort_brw_pages(pga, page_count);
1320         rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1321                                   pga, &requested_nob, &nio_count, &req);
1322         if (rc != 0) {
1323                 CERROR("prep_req failed: %d\n", rc);
1324                 GOTO(out, req = ERR_PTR(rc));
1325         }
1326
1327         LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1328         aa = (struct osc_brw_async_args *)&req->rq_async_args;
1329         aa->aa_oa = oa;
1330         aa->aa_requested_nob = requested_nob;
1331         aa->aa_nio_count = nio_count;
1332         aa->aa_page_count = page_count;
1333         aa->aa_pga = pga;
1334         aa->aa_cli = cli;
1335
1336 out:
1337         if (IS_ERR(req)) {
1338                 if (oa)
1339                         obdo_free(oa);
1340                 if (pga)
1341                         OBD_FREE(pga, sizeof(*pga) * page_count);
1342         }
1343         RETURN(req);
1344 }
1345
1346 static void lop_update_pending(struct client_obd *cli,
1347                                struct loi_oap_pages *lop, int cmd, int delta)
1348 {
1349         lop->lop_num_pending += delta;
1350         if (cmd == OBD_BRW_WRITE)
1351                 cli->cl_pending_w_pages += delta;
1352         else
1353                 cli->cl_pending_r_pages += delta;
1354 }
1355
1356 /* the loi lock is held across this function but it's allowed to release
1357  * and reacquire it during its work */
1358 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1359                             int cmd, struct loi_oap_pages *lop)
1360 {
1361         struct ptlrpc_request *request;
1362         obd_count page_count = 0;
1363         struct list_head *tmp, *pos;
1364         struct osc_async_page *oap = NULL;
1365         struct osc_brw_async_args *aa;
1366         struct obd_async_page_ops *ops;
1367         LIST_HEAD(rpc_list);
1368         ENTRY;
1369
1370         /* first we find the pages we're allowed to work with */
1371         list_for_each_safe(pos, tmp, &lop->lop_pending) {
1372                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1373                 ops = oap->oap_caller_ops;
1374
1375                 LASSERT(oap->oap_magic == OAP_MAGIC);
1376
1377                 /* in llite being 'ready' equates to the page being locked
1378                  * until completion unlocks it.  commit_write submits a page
1379                  * as not ready because its unlock will happen unconditionally
1380                  * as the call returns.  if we race with commit_write giving
1381                  * us that page we dont' want to create a hole in the page
1382                  * stream, so we stop and leave the rpc to be fired by
1383                  * another dirtier or kupdated interval (the not ready page
1384                  * will still be on the dirty list).  we could call in
1385                  * at the end of ll_file_write to process the queue again. */
1386                 if (!(oap->oap_async_flags & ASYNC_READY)) {
1387                         int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1388                         if (rc < 0)
1389                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
1390                                                 "instead of ready\n", oap,
1391                                                 oap->oap_page, rc);
1392                         switch (rc) {
1393                         case -EAGAIN:
1394                                 /* llite is telling us that the page is still
1395                                  * in commit_write and that we should try
1396                                  * and put it in an rpc again later.  we
1397                                  * break out of the loop so we don't create
1398                                  * a hole in the sequence of pages in the rpc
1399                                  * stream.*/
1400                                 pos = NULL;
1401                                 break;
1402                         case -EINTR:
1403                                 /* the io isn't needed.. tell the checks
1404                                  * below to complete the rpc with EINTR */
1405                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1406                                 oap->oap_count = -EINTR;
1407                                 break;
1408                         case 0:
1409                                 oap->oap_async_flags |= ASYNC_READY;
1410                                 break;
1411                         default:
1412                                 LASSERTF(0, "oap %p page %p returned %d "
1413                                             "from make_ready\n", oap,
1414                                             oap->oap_page, rc);
1415                                 break;
1416                         }
1417                 }
1418                 if (pos == NULL)
1419                         break;
1420
1421                 /* take the page out of our book-keeping */
1422                 list_del_init(&oap->oap_pending_item);
1423                 lop_update_pending(cli, lop, cmd, -1);
1424                 if (!list_empty(&oap->oap_urgent_item))
1425                         list_del_init(&oap->oap_urgent_item);
1426
1427                 /* ask the caller for the size of the io as the rpc leaves. */
1428                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1429                         oap->oap_count =
1430                                 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1431                 if (oap->oap_count <= 0) {
1432                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1433                                oap->oap_count);
1434                         osc_complete_oap(cli, oap, 0, oap->oap_count);
1435                         continue;
1436                 }
1437
1438                 /* now put the page back in our accounting */
1439                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1440                 if (++page_count >= cli->cl_max_pages_per_rpc)
1441                         break;
1442         }
1443
1444         osc_wake_cache_waiters(cli);
1445
1446         if (page_count == 0)
1447                 RETURN(0);
1448
1449         loi_list_maint(cli, loi);
1450         spin_unlock(&cli->cl_loi_list_lock);
1451
1452         request = osc_build_req(cli, &rpc_list, page_count, cmd);
1453         if (IS_ERR(request)) {
1454                 /* this should happen rarely and is pretty bad, it makes the
1455                  * pending list not follow the dirty order */
1456                 spin_lock(&cli->cl_loi_list_lock);
1457                 list_for_each_safe(pos, tmp, &rpc_list) {
1458                         oap = list_entry(pos, struct osc_async_page,
1459                                          oap_rpc_item);
1460                         list_del_init(&oap->oap_rpc_item);
1461
1462                         /* queued sync pages can be torn down while the pages
1463                          * were between the pending list and the rpc */
1464                         if (oap->oap_interrupted) {
1465                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1466                                 osc_complete_oap(cli, oap, 0, oap->oap_count);
1467                                 continue;
1468                         }
1469
1470                         /* put the page back in the loi/lop lists */
1471                         list_add_tail(&oap->oap_pending_item,
1472                                       &lop->lop_pending);
1473                         lop_update_pending(cli, lop, cmd, 1);
1474                         if (oap->oap_async_flags & ASYNC_URGENT)
1475                                 list_add(&oap->oap_urgent_item,
1476                                          &lop->lop_urgent);
1477                 }
1478                 loi_list_maint(cli, loi);
1479                 RETURN(PTR_ERR(request));
1480         }
1481
1482         LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1483         aa = (struct osc_brw_async_args *)&request->rq_async_args;
1484         INIT_LIST_HEAD(&aa->aa_oaps);
1485         list_splice(&rpc_list, &aa->aa_oaps);
1486         INIT_LIST_HEAD(&rpc_list);
1487
1488 #ifdef __KERNEL__
1489         if (cmd == OBD_BRW_READ) {
1490                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1491                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_brw_in_flight);
1492         } else {
1493                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1494                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1495                                  cli->cl_brw_in_flight);
1496         }
1497 #endif
1498
1499         spin_lock(&cli->cl_loi_list_lock);
1500
1501         cli->cl_brw_in_flight++;
1502         /* queued sync pages can be torn down while the pages
1503          * were between the pending list and the rpc */
1504         list_for_each(pos, &aa->aa_oaps) {
1505                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1506                 if (oap->oap_interrupted) {
1507                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1508                                oap, request);
1509                         ptlrpc_mark_interrupted(request);
1510                         break;
1511                 }
1512         }
1513
1514         CDEBUG(D_INODE, "req %p: %d pages, aa %p.  now %d in flight\n", request,
1515                page_count, aa, cli->cl_brw_in_flight);
1516
1517         oap->oap_request = ptlrpc_request_addref(request);
1518         request->rq_interpret_reply = brw_interpret_oap;
1519         ptlrpcd_add_req(request);
1520         RETURN(1);
1521 }
1522
1523 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1524                          int cmd)
1525 {
1526         int optimal;
1527         ENTRY;
1528
1529         if (lop->lop_num_pending == 0)
1530                 RETURN(0);
1531
1532         /* stream rpcs in queue order as long as as there is an urgent page
1533          * queued.  this is our cheap solution for good batching in the case
1534          * where writepage marks some random page in the middle of the file as
1535          * urgent because of, say, memory pressure */
1536         if (!list_empty(&lop->lop_urgent))
1537                 RETURN(1);
1538
1539         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1540         optimal = cli->cl_max_pages_per_rpc;
1541         if (cmd == OBD_BRW_WRITE) {
1542                 /* trigger a write rpc stream as long as there are dirtiers
1543                  * waiting for space.  as they're waiting, they're not going to
1544                  * create more pages to coallesce with what's waiting.. */
1545                 if (!list_empty(&cli->cl_cache_waiters))
1546                         RETURN(1);
1547
1548                 /* *2 to avoid triggering rpcs that would want to include pages
1549                  * that are being queued but which can't be made ready until
1550                  * the queuer finishes with the page. this is a wart for
1551                  * llite::commit_write() */
1552                 optimal += 16;
1553         }
1554         if (lop->lop_num_pending >= optimal)
1555                 RETURN(1);
1556
1557         RETURN(0);
1558 }
1559
1560 static void on_list(struct list_head *item, struct list_head *list,
1561                     int should_be_on)
1562 {
1563         if (list_empty(item) && should_be_on)
1564                 list_add_tail(item, list);
1565         else if (!list_empty(item) && !should_be_on)
1566                 list_del_init(item);
1567 }
1568
1569 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1570  * can find pages to build into rpcs quickly */
1571 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1572 {
1573         on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1574                 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1575                 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1576
1577         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1578                 loi->loi_write_lop.lop_num_pending);
1579 }
1580
1581 #define LOI_DEBUG(LOI, STR, args...)                                     \
1582         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
1583                !list_empty(&(LOI)->loi_cli_item),                        \
1584                (LOI)->loi_write_lop.lop_num_pending,                     \
1585                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
1586                (LOI)->loi_read_lop.lop_num_pending,                      \
1587                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
1588                args)                                                     \
1589
1590 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1591 {
1592         ENTRY;
1593         /* first return all objects which we already know to have
1594          * pages ready to be stuffed into rpcs */
1595         if (!list_empty(&cli->cl_loi_ready_list))
1596                 RETURN(list_entry(cli->cl_loi_ready_list.next,
1597                                   struct lov_oinfo, loi_cli_item));
1598
1599         /* then if we have cache waiters, return all objects with queued
1600          * writes.  This is especially important when many small files
1601          * have filled up the cache and not been fired into rpcs because
1602          * they don't pass the nr_pending/object threshhold */
1603         if (!list_empty(&cli->cl_cache_waiters) &&
1604             !list_empty(&cli->cl_loi_write_list))
1605                 RETURN(list_entry(cli->cl_loi_write_list.next,
1606                                   struct lov_oinfo, loi_write_item));
1607         RETURN(NULL);
1608 }
1609
1610 /* called with the loi list lock held */
1611 static void osc_check_rpcs(struct client_obd *cli)
1612 {
1613         struct lov_oinfo *loi;
1614         int rc = 0, race_counter = 0;
1615         ENTRY;
1616
1617         while ((loi = osc_next_loi(cli)) != NULL) {
1618                 LOI_DEBUG(loi, "%d in flight\n", cli->cl_brw_in_flight);
1619
1620                 if (cli->cl_brw_in_flight >= cli->cl_max_rpcs_in_flight)
1621                         break;
1622
1623                 /* attempt some read/write balancing by alternating between
1624                  * reads and writes in an object.  The makes_rpc checks here
1625                  * would be redundant if we were getting read/write work items
1626                  * instead of objects.  we don't want send_oap_rpc to drain a
1627                  * partial read pending queue when we're given this object to
1628                  * do io on writes while there are cache waiters */
1629                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1630                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1631                                               &loi->loi_write_lop);
1632                         if (rc < 0)
1633                                 break;
1634                         if (rc > 0)
1635                                 race_counter = 0;
1636                         else
1637                                 race_counter++;
1638                 }
1639                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1640                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1641                                               &loi->loi_read_lop);
1642                         if (rc < 0)
1643                                 break;
1644                         if (rc > 0)
1645                                 race_counter = 0;
1646                         else
1647                                 race_counter++;
1648                 }
1649
1650                 /* attempt some inter-object balancing by issueing rpcs
1651                  * for each object in turn */
1652                 if (!list_empty(&loi->loi_cli_item))
1653                         list_del_init(&loi->loi_cli_item);
1654                 if (!list_empty(&loi->loi_write_item))
1655                         list_del_init(&loi->loi_write_item);
1656
1657                 loi_list_maint(cli, loi);
1658
1659                 /* send_oap_rpc fails with 0 when make_ready tells it to
1660                  * back off.  llite's make_ready does this when it tries
1661                  * to lock a page queued for write that is already locked.
1662                  * we want to try sending rpcs from many objects, but we
1663                  * don't want to spin failing with 0.  */
1664                 if (race_counter == 10)
1665                         break;
1666         }
1667         EXIT;
1668 }
1669
1670 /* we're trying to queue a page in the osc so we're subject to the
1671  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1672  * If the osc's queued pages are already at that limit, then we want to sleep
1673  * until there is space in the osc's queue for us.  We also may be waiting for
1674  * write credits from the OST if there are RPCs in flight that may return some
1675  * before we fall back to sync writes.
1676  *
1677  * We need this know our allocation was granted in the presence of signals */
1678 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1679 {
1680         int rc;
1681         ENTRY;
1682         spin_lock(&cli->cl_loi_list_lock);
1683         rc = list_empty(&ocw->ocw_entry) || cli->cl_brw_in_flight == 0;
1684         spin_unlock(&cli->cl_loi_list_lock);
1685         RETURN(rc);
1686 };
1687
1688 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1689  * grant or cache space. */
1690 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1691                            struct osc_async_page *oap)
1692 {
1693         struct osc_cache_waiter ocw;
1694         struct l_wait_info lwi = { 0 };
1695
1696         CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1697                cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1698                cli->cl_avail_grant);
1699
1700         if (cli->cl_dirty_max < PAGE_SIZE)
1701                 return(-EDQUOT);
1702
1703
1704         /* Hopefully normal case - cache space and write credits available */
1705         if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1706             cli->cl_avail_grant >= PAGE_SIZE) {
1707                 /* account for ourselves */
1708                 osc_consume_write_grant(cli, oap);
1709                 return(0);
1710         }
1711
1712         /* Make sure that there are write rpcs in flight to wait for.  This
1713          * is a little silly as this object may not have any pending but
1714          * other objects sure might. */
1715         if (cli->cl_brw_in_flight) {
1716                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1717                 init_waitqueue_head(&ocw.ocw_waitq);
1718                 ocw.ocw_oap = oap;
1719                 ocw.ocw_rc = 0;
1720
1721                 loi_list_maint(cli, loi);
1722                 osc_check_rpcs(cli);
1723                 spin_unlock(&cli->cl_loi_list_lock);
1724
1725                 CDEBUG(0, "sleeping for cache space\n");
1726                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1727
1728                 spin_lock(&cli->cl_loi_list_lock);
1729                 if (!list_empty(&ocw.ocw_entry)) {
1730                         list_del(&ocw.ocw_entry);
1731                         RETURN(-EINTR);
1732                 }
1733                 RETURN(ocw.ocw_rc);
1734         }
1735
1736         RETURN(-EDQUOT);
1737 }
1738
1739 /* the companion to enter_cache, called when an oap is no longer part of the
1740  * dirty accounting.. so writeback completes or truncate happens before writing
1741  * starts.  must be called with the loi lock held. */
1742 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1743                            int sent)
1744 {
1745         ENTRY;
1746
1747         if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1748                 EXIT;
1749                 return;
1750         }
1751
1752         oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1753         cli->cl_dirty -= PAGE_SIZE;
1754         if (!sent) {
1755                 cli->cl_lost_grant += PAGE_SIZE;
1756                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1757                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1758         }
1759
1760         EXIT;
1761 }
1762
1763 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1764                         struct lov_oinfo *loi, struct page *page,
1765                         obd_off offset, struct obd_async_page_ops *ops,
1766                         void *data, void **res)
1767 {
1768         struct osc_async_page *oap;
1769         ENTRY;
1770
1771         OBD_ALLOC(oap, sizeof(*oap));
1772         if (oap == NULL)
1773                 return -ENOMEM;
1774
1775         oap->oap_magic = OAP_MAGIC;
1776         oap->oap_cli = &exp->exp_obd->u.cli;
1777         oap->oap_loi = loi;
1778
1779         oap->oap_caller_ops = ops;
1780         oap->oap_caller_data = data;
1781
1782         oap->oap_page = page;
1783         oap->oap_obj_off = offset;
1784
1785         INIT_LIST_HEAD(&oap->oap_pending_item);
1786         INIT_LIST_HEAD(&oap->oap_urgent_item);
1787         INIT_LIST_HEAD(&oap->oap_rpc_item);
1788
1789         oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1790
1791         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1792         *res = oap;
1793         RETURN(0);
1794 }
1795
1796 struct osc_async_page *oap_from_cookie(void *cookie)
1797 {
1798         struct osc_async_page *oap = cookie;
1799         if (oap->oap_magic != OAP_MAGIC)
1800                 return ERR_PTR(-EINVAL);
1801         return oap;
1802 };
1803
1804 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1805                               struct lov_oinfo *loi, void *cookie,
1806                               int cmd, obd_off off, int count,
1807                               obd_flag brw_flags, enum async_flags async_flags)
1808 {
1809         struct client_obd *cli = &exp->exp_obd->u.cli;
1810         struct osc_async_page *oap;
1811         struct loi_oap_pages *lop;
1812         int rc;
1813         ENTRY;
1814
1815         oap = oap_from_cookie(cookie);
1816         if (IS_ERR(oap))
1817                 RETURN(PTR_ERR(oap));
1818
1819         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1820                 RETURN(-EIO);
1821
1822         if (!list_empty(&oap->oap_pending_item) ||
1823             !list_empty(&oap->oap_urgent_item) ||
1824             !list_empty(&oap->oap_rpc_item))
1825                 RETURN(-EBUSY);
1826
1827         if (loi == NULL)
1828                 loi = &lsm->lsm_oinfo[0];
1829
1830         spin_lock(&cli->cl_loi_list_lock);
1831
1832         oap->oap_cmd = cmd;
1833         oap->oap_async_flags = async_flags;
1834         oap->oap_page_off = off;
1835         oap->oap_count = count;
1836         oap->oap_brw_flags = brw_flags;
1837
1838         if (cmd == OBD_BRW_WRITE) {
1839                 rc = osc_enter_cache(cli, loi, oap);
1840                 if (rc) {
1841                         spin_unlock(&cli->cl_loi_list_lock);
1842                         RETURN(rc);
1843                 }
1844                 lop = &loi->loi_write_lop;
1845         } else {
1846                 lop = &loi->loi_read_lop;
1847         }
1848
1849         if (oap->oap_async_flags & ASYNC_URGENT)
1850                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1851         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1852         lop_update_pending(cli, lop, cmd, 1);
1853
1854         loi_list_maint(cli, loi);
1855
1856         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
1857                   cmd);
1858
1859         osc_check_rpcs(cli);
1860         spin_unlock(&cli->cl_loi_list_lock);
1861
1862         RETURN(0);
1863 }
1864
1865 /* aka (~was & now & flag), but this is more clear :) */
1866 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
1867
1868 static int osc_set_async_flags(struct obd_export *exp,
1869                                struct lov_stripe_md *lsm,
1870                                struct lov_oinfo *loi, void *cookie,
1871                                obd_flag async_flags)
1872 {
1873         struct client_obd *cli = &exp->exp_obd->u.cli;
1874         struct loi_oap_pages *lop;
1875         struct osc_async_page *oap;
1876         int rc = 0;
1877         ENTRY;
1878
1879         oap = oap_from_cookie(cookie);
1880         if (IS_ERR(oap))
1881                 RETURN(PTR_ERR(oap));
1882
1883         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1884                 RETURN(-EIO);
1885
1886         if (loi == NULL)
1887                 loi = &lsm->lsm_oinfo[0];
1888
1889         if (oap->oap_cmd == OBD_BRW_WRITE) {
1890                 lop = &loi->loi_write_lop;
1891         } else {
1892                 lop = &loi->loi_read_lop;
1893         }
1894
1895         spin_lock(&cli->cl_loi_list_lock);
1896
1897         if (list_empty(&oap->oap_pending_item))
1898                 GOTO(out, rc = -EINVAL);
1899
1900         if ((oap->oap_async_flags & async_flags) == async_flags)
1901                 GOTO(out, rc = 0);
1902
1903         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
1904                 oap->oap_async_flags |= ASYNC_READY;
1905
1906         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
1907                 if (list_empty(&oap->oap_rpc_item)) {
1908                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1909                         loi_list_maint(cli, loi);
1910                 }
1911         }
1912
1913         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
1914                         oap->oap_async_flags);
1915 out:
1916         osc_check_rpcs(cli);
1917         spin_unlock(&cli->cl_loi_list_lock);
1918         RETURN(rc);
1919 }
1920
1921 static int osc_queue_sync_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1922                              struct lov_oinfo *loi,
1923                              struct obd_sync_io_container *osic, void *cookie,
1924                              int cmd, obd_off off, int count,
1925                              obd_flag brw_flags)
1926 {
1927         struct client_obd *cli = &exp->exp_obd->u.cli;
1928         struct osc_async_page *oap;
1929         struct loi_oap_pages *lop;
1930         ENTRY;
1931
1932         oap = oap_from_cookie(cookie);
1933         if (IS_ERR(oap))
1934                 RETURN(PTR_ERR(oap));
1935
1936         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1937                 RETURN(-EIO);
1938
1939         if (!list_empty(&oap->oap_pending_item) ||
1940             !list_empty(&oap->oap_urgent_item) ||
1941             !list_empty(&oap->oap_rpc_item))
1942                 RETURN(-EBUSY);
1943
1944         if (loi == NULL)
1945                 loi = &lsm->lsm_oinfo[0];
1946
1947         spin_lock(&cli->cl_loi_list_lock);
1948
1949         oap->oap_cmd = cmd;
1950         oap->oap_page_off = off;
1951         oap->oap_count = count;
1952         oap->oap_brw_flags = brw_flags;
1953
1954         if (cmd == OBD_BRW_WRITE)
1955                 lop = &loi->loi_write_lop;
1956         else
1957                 lop = &loi->loi_read_lop;
1958
1959         list_add_tail(&oap->oap_pending_item, &lop->lop_pending_sync);
1960         oap->oap_osic = osic;
1961         osic_add_one(osic, &oap->oap_occ);
1962
1963         LOI_DEBUG(loi, "oap %p page %p on sync pending\n", oap, oap->oap_page);
1964
1965         spin_unlock(&cli->cl_loi_list_lock);
1966
1967         RETURN(0);
1968 }
1969
1970 static void osc_sync_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
1971                                 struct loi_oap_pages *lop, int cmd)
1972 {
1973         struct list_head *pos, *tmp;
1974         struct osc_async_page *oap;
1975
1976         list_for_each_safe(pos, tmp, &lop->lop_pending_sync) {
1977                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1978                 list_del(&oap->oap_pending_item);
1979                 oap->oap_async_flags |= ASYNC_READY | ASYNC_URGENT |
1980                                         ASYNC_COUNT_STABLE;
1981                 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1982                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1983                 lop_update_pending(cli, lop, cmd, 1);
1984         }
1985         loi_list_maint(cli, loi);
1986 }
1987
1988 static int osc_trigger_sync_io(struct obd_export *exp,
1989                                struct lov_stripe_md *lsm,
1990                                struct lov_oinfo *loi,
1991                                struct obd_sync_io_container *osic)
1992 {
1993         struct client_obd *cli = &exp->exp_obd->u.cli;
1994         ENTRY;
1995
1996         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1997                 RETURN(-EIO);
1998
1999         if (loi == NULL)
2000                 loi = &lsm->lsm_oinfo[0];
2001
2002         spin_lock(&cli->cl_loi_list_lock);
2003
2004         osc_sync_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2005         osc_sync_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2006
2007         osc_check_rpcs(cli);
2008         spin_unlock(&cli->cl_loi_list_lock);
2009
2010         RETURN(0);
2011 }
2012
2013 static int osc_teardown_async_page(struct obd_export *exp,
2014                                    struct lov_stripe_md *lsm,
2015                                    struct lov_oinfo *loi, void *cookie)
2016 {
2017         struct client_obd *cli = &exp->exp_obd->u.cli;
2018         struct loi_oap_pages *lop;
2019         struct osc_async_page *oap;
2020         int rc = 0;
2021         ENTRY;
2022
2023         oap = oap_from_cookie(cookie);
2024         if (IS_ERR(oap))
2025                 RETURN(PTR_ERR(oap));
2026
2027         if (loi == NULL)
2028                 loi = &lsm->lsm_oinfo[0];
2029
2030         if (oap->oap_cmd == OBD_BRW_WRITE) {
2031                 lop = &loi->loi_write_lop;
2032         } else {
2033                 lop = &loi->loi_read_lop;
2034         }
2035
2036         spin_lock(&cli->cl_loi_list_lock);
2037
2038         if (!list_empty(&oap->oap_rpc_item))
2039                 GOTO(out, rc = -EBUSY);
2040
2041         osc_exit_cache(cli, oap, 0);
2042         osc_wake_cache_waiters(cli);
2043
2044         if (!list_empty(&oap->oap_urgent_item)) {
2045                 list_del_init(&oap->oap_urgent_item);
2046                 oap->oap_async_flags &= ~ASYNC_URGENT;
2047         }
2048         if (!list_empty(&oap->oap_pending_item)) {
2049                 list_del_init(&oap->oap_pending_item);
2050                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2051         }
2052         loi_list_maint(cli, loi);
2053
2054         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2055 out:
2056         spin_unlock(&cli->cl_loi_list_lock);
2057         if (rc == 0)
2058                 OBD_FREE(oap, sizeof(*oap));
2059         RETURN(rc);
2060 }
2061
2062 #ifdef __KERNEL__
2063 /* Note: caller will lock/unlock, and set uptodate on the pages */
2064 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2065 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2066                            struct lov_stripe_md *lsm, obd_count page_count,
2067                            struct brw_page *pga)
2068 {
2069         struct ptlrpc_request *request = NULL;
2070         struct ost_body *body;
2071         struct niobuf_remote *nioptr;
2072         struct obd_ioobj *iooptr;
2073         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2074         int swab;
2075         ENTRY;
2076
2077         /* XXX does not handle 'new' brw protocol */
2078
2079         size[1] = sizeof(struct obd_ioobj);
2080         size[2] = page_count * sizeof(*nioptr);
2081
2082         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_READ, 3,
2083                                   size, NULL);
2084         if (!request)
2085                 RETURN(-ENOMEM);
2086
2087         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2088         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2089         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2090                                 sizeof(*nioptr) * page_count);
2091
2092         memcpy(&body->oa, oa, sizeof(body->oa));
2093
2094         obdo_to_ioobj(oa, iooptr);
2095         iooptr->ioo_bufcnt = page_count;
2096
2097         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2098                 LASSERT(PageLocked(pga[mapped].pg));
2099                 LASSERT(mapped == 0 || pga[mapped].off > pga[mapped - 1].off);
2100
2101                 nioptr->offset = pga[mapped].off;
2102                 nioptr->len    = pga[mapped].count;
2103                 nioptr->flags  = pga[mapped].flag;
2104         }
2105
2106         size[1] = page_count * sizeof(*nioptr);
2107         request->rq_replen = lustre_msg_size(2, size);
2108
2109         rc = ptlrpc_queue_wait(request);
2110         if (rc)
2111                 GOTO(out_req, rc);
2112
2113         body = lustre_swab_repbuf(request, 0, sizeof(*body),
2114                                   lustre_swab_ost_body);
2115         if (body == NULL) {
2116                 CERROR("Can't unpack body\n");
2117                 GOTO(out_req, rc = -EPROTO);
2118         }
2119
2120         memcpy(oa, &body->oa, sizeof(*oa));
2121
2122         swab = lustre_msg_swabbed(request->rq_repmsg);
2123         LASSERT_REPSWAB(request, 1);
2124         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2125         if (!nioptr) {
2126                 /* nioptr missing or short */
2127                 GOTO(out_req, rc = -EPROTO);
2128         }
2129
2130         /* actual read */
2131         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2132                 struct page *page = pga[mapped].pg;
2133                 struct buffer_head *bh;
2134                 kdev_t dev;
2135
2136                 if (swab)
2137                         lustre_swab_niobuf_remote (nioptr);
2138
2139                 /* got san device associated */
2140                 LASSERT(exp->exp_obd != NULL);
2141                 dev = exp->exp_obd->u.cli.cl_sandev;
2142
2143                 /* hole */
2144                 if (!nioptr->offset) {
2145                         CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2146                                         page->mapping->host->i_ino,
2147                                         page->index);
2148                         memset(page_address(page), 0, PAGE_SIZE);
2149                         continue;
2150                 }
2151
2152                 if (!page->buffers) {
2153                         create_empty_buffers(page, dev, PAGE_SIZE);
2154                         bh = page->buffers;
2155
2156                         clear_bit(BH_New, &bh->b_state);
2157                         set_bit(BH_Mapped, &bh->b_state);
2158                         bh->b_blocknr = (unsigned long)nioptr->offset;
2159
2160                         clear_bit(BH_Uptodate, &bh->b_state);
2161
2162                         ll_rw_block(READ, 1, &bh);
2163                 } else {
2164                         bh = page->buffers;
2165
2166                         /* if buffer already existed, it must be the
2167                          * one we mapped before, check it */
2168                         LASSERT(!test_bit(BH_New, &bh->b_state));
2169                         LASSERT(test_bit(BH_Mapped, &bh->b_state));
2170                         LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2171
2172                         /* wait it's io completion */
2173                         if (test_bit(BH_Lock, &bh->b_state))
2174                                 wait_on_buffer(bh);
2175
2176                         if (!test_bit(BH_Uptodate, &bh->b_state))
2177                                 ll_rw_block(READ, 1, &bh);
2178                 }
2179
2180
2181                 /* must do syncronous write here */
2182                 wait_on_buffer(bh);
2183                 if (!buffer_uptodate(bh)) {
2184                         /* I/O error */
2185                         rc = -EIO;
2186                         goto out_req;
2187                 }
2188         }
2189
2190 out_req:
2191         ptlrpc_req_finished(request);
2192         RETURN(rc);
2193 }
2194
2195 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2196                             struct lov_stripe_md *lsm, obd_count page_count,
2197                             struct brw_page *pga)
2198 {
2199         struct ptlrpc_request *request = NULL;
2200         struct ost_body *body;
2201         struct niobuf_remote *nioptr;
2202         struct obd_ioobj *iooptr;
2203         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2204         int swab;
2205         ENTRY;
2206
2207         size[1] = sizeof(struct obd_ioobj);
2208         size[2] = page_count * sizeof(*nioptr);
2209
2210         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_WRITE,
2211                                   3, size, NULL);
2212         if (!request)
2213                 RETURN(-ENOMEM);
2214
2215         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2216         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2217         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2218                                 sizeof (*nioptr) * page_count);
2219
2220         memcpy(&body->oa, oa, sizeof(body->oa));
2221
2222         obdo_to_ioobj(oa, iooptr);
2223         iooptr->ioo_bufcnt = page_count;
2224
2225         /* pack request */
2226         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2227                 LASSERT(PageLocked(pga[mapped].pg));
2228                 LASSERT(mapped == 0 || pga[mapped].off > pga[mapped - 1].off);
2229
2230                 nioptr->offset = pga[mapped].off;
2231                 nioptr->len    = pga[mapped].count;
2232                 nioptr->flags  = pga[mapped].flag;
2233         }
2234
2235         size[1] = page_count * sizeof(*nioptr);
2236         request->rq_replen = lustre_msg_size(2, size);
2237
2238         rc = ptlrpc_queue_wait(request);
2239         if (rc)
2240                 GOTO(out_req, rc);
2241
2242         swab = lustre_msg_swabbed (request->rq_repmsg);
2243         LASSERT_REPSWAB (request, 1);
2244         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2245         if (!nioptr) {
2246                 CERROR("absent/short niobuf array\n");
2247                 GOTO(out_req, rc = -EPROTO);
2248         }
2249
2250         /* actual write */
2251         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2252                 struct page *page = pga[mapped].pg;
2253                 struct buffer_head *bh;
2254                 kdev_t dev;
2255
2256                 if (swab)
2257                         lustre_swab_niobuf_remote (nioptr);
2258
2259                 /* got san device associated */
2260                 LASSERT(exp->exp_obd != NULL);
2261                 dev = exp->exp_obd->u.cli.cl_sandev;
2262
2263                 if (!page->buffers) {
2264                         create_empty_buffers(page, dev, PAGE_SIZE);
2265                 } else {
2266                         /* checking */
2267                         LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2268                         LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2269                         LASSERT(page->buffers->b_blocknr ==
2270                                 (unsigned long)nioptr->offset);
2271                 }
2272                 bh = page->buffers;
2273
2274                 LASSERT(bh);
2275
2276                 /* if buffer locked, wait it's io completion */
2277                 if (test_bit(BH_Lock, &bh->b_state))
2278                         wait_on_buffer(bh);
2279
2280                 clear_bit(BH_New, &bh->b_state);
2281                 set_bit(BH_Mapped, &bh->b_state);
2282
2283                 /* override the block nr */
2284                 bh->b_blocknr = (unsigned long)nioptr->offset;
2285
2286                 /* we are about to write it, so set it
2287                  * uptodate/dirty
2288                  * page lock should garentee no race condition here */
2289                 set_bit(BH_Uptodate, &bh->b_state);
2290                 set_bit(BH_Dirty, &bh->b_state);
2291
2292                 ll_rw_block(WRITE, 1, &bh);
2293
2294                 /* must do syncronous write here */
2295                 wait_on_buffer(bh);
2296                 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2297                         /* I/O error */
2298                         rc = -EIO;
2299                         goto out_req;
2300                 }
2301         }
2302
2303 out_req:
2304         ptlrpc_req_finished(request);
2305         RETURN(rc);
2306 }
2307
2308 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2309                       struct lov_stripe_md *lsm, obd_count page_count,
2310                       struct brw_page *pga, struct obd_trans_info *oti)
2311 {
2312         ENTRY;
2313
2314         while (page_count) {
2315                 obd_count pages_per_brw;
2316                 int rc;
2317
2318                 if (page_count > OSC_BRW_MAX_IOV)
2319                         pages_per_brw = OSC_BRW_MAX_IOV;
2320                 else
2321                         pages_per_brw = page_count;
2322
2323                 if (cmd & OBD_BRW_WRITE)
2324                         rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2325                 else
2326                         rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2327
2328                 if (rc != 0)
2329                         RETURN(rc);
2330
2331                 page_count -= pages_per_brw;
2332                 pga += pages_per_brw;
2333         }
2334         RETURN(0);
2335 }
2336 #endif
2337 #endif
2338
2339 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2340 {
2341         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2342
2343         LASSERT(lock != NULL);
2344         l_lock(&lock->l_resource->lr_namespace->ns_lock);
2345 #ifdef __KERNEL__
2346         if (lock->l_ast_data && lock->l_ast_data != data) {
2347                 struct inode *new_inode = data;
2348                 struct inode *old_inode = lock->l_ast_data;
2349                 LASSERTF(old_inode->i_state & I_FREEING,
2350                          "Found existing inode %p/%lu/%u state %lu in lock: "
2351                          "setting data to %p/%lu/%u\n", old_inode,
2352                          old_inode->i_ino, old_inode->i_generation,
2353                          old_inode->i_state,
2354                          new_inode, new_inode->i_ino, new_inode->i_generation);
2355         }
2356 #endif
2357         lock->l_ast_data = data;
2358         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
2359         LDLM_LOCK_PUT(lock);
2360 }
2361
2362 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2363                              ldlm_iterator_t replace, void *data)
2364 {
2365         struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2366         struct obd_device *obd = class_exp2obd(exp);
2367
2368         ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2369         return 0;
2370 }
2371
2372 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2373                        struct lustre_handle *parent_lock,
2374                        __u32 type, void *extentp, int extent_len, __u32 mode,
2375                        int *flags, void *callback, void *data,
2376                        struct lustre_handle *lockh)
2377 {
2378         struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2379         struct obd_device *obd = exp->exp_obd;
2380         struct ldlm_extent *extent = extentp;
2381         int rc;
2382         ENTRY;
2383
2384         /* Filesystem lock extents are extended to page boundaries so that
2385          * dealing with the page cache is a little smoother.  */
2386         extent->start -= extent->start & ~PAGE_MASK;
2387         extent->end |= ~PAGE_MASK;
2388
2389         /* Next, search for already existing extent locks that will cover us */
2390         rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id,
2391                              type, extent, sizeof(*extent), mode, lockh);
2392         if (rc == 1) {
2393                 osc_set_data_with_check(lockh, data);
2394                 /* We already have a lock, and it's referenced */
2395                 RETURN(ELDLM_OK);
2396         }
2397
2398         /* If we're trying to read, we also search for an existing PW lock.  The
2399          * VFS and page cache already protect us locally, so lots of readers/
2400          * writers can share a single PW lock.
2401          *
2402          * There are problems with conversion deadlocks, so instead of
2403          * converting a read lock to a write lock, we'll just enqueue a new
2404          * one.
2405          *
2406          * At some point we should cancel the read lock instead of making them
2407          * send us a blocking callback, but there are problems with canceling
2408          * locks out from other users right now, too. */
2409
2410         if (mode == LCK_PR) {
2411                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2412                                      extent, sizeof(*extent), LCK_PW, lockh);
2413                 if (rc == 1) {
2414                         /* FIXME: This is not incredibly elegant, but it might
2415                          * be more elegant than adding another parameter to
2416                          * lock_match.  I want a second opinion. */
2417                         ldlm_lock_addref(lockh, LCK_PR);
2418                         ldlm_lock_decref(lockh, LCK_PW);
2419                         osc_set_data_with_check(lockh, data);
2420                         RETURN(ELDLM_OK);
2421                 }
2422         }
2423
2424         rc = ldlm_cli_enqueue(exp, NULL, obd->obd_namespace, parent_lock,
2425                               res_id, type, extent, sizeof(*extent), mode,
2426                               flags,ldlm_completion_ast, callback, data, lockh);
2427         RETURN(rc);
2428 }
2429
2430 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2431                      __u32 type, void *extentp, int extent_len, __u32 mode,
2432                      int *flags, void *data, struct lustre_handle *lockh)
2433 {
2434         struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2435         struct obd_device *obd = exp->exp_obd;
2436         struct ldlm_extent *extent = extentp;
2437         int rc;
2438         ENTRY;
2439
2440         OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2441
2442         /* Filesystem lock extents are extended to page boundaries so that
2443          * dealing with the page cache is a little smoother */
2444         extent->start -= extent->start & ~PAGE_MASK;
2445         extent->end |= ~PAGE_MASK;
2446
2447         /* Next, search for already existing extent locks that will cover us */
2448         rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2449                              extent, sizeof(*extent), mode, lockh);
2450         if (rc) {
2451                 osc_set_data_with_check(lockh, data);
2452                 RETURN(rc);
2453         }
2454         /* If we're trying to read, we also search for an existing PW lock.  The
2455          * VFS and page cache already protect us locally, so lots of readers/
2456          * writers can share a single PW lock. */
2457         if (mode == LCK_PR) {
2458                 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2459                                      extent, sizeof(*extent), LCK_PW, lockh);
2460                 if (rc == 1) {
2461                         /* FIXME: This is not incredibly elegant, but it might
2462                          * be more elegant than adding another parameter to
2463                          * lock_match.  I want a second opinion. */
2464                         osc_set_data_with_check(lockh, data);
2465                         ldlm_lock_addref(lockh, LCK_PR);
2466                         ldlm_lock_decref(lockh, LCK_PW);
2467                 }
2468         }
2469         RETURN(rc);
2470 }
2471
2472 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2473                       __u32 mode, struct lustre_handle *lockh)
2474 {
2475         ENTRY;
2476
2477         ldlm_lock_decref(lockh, mode);
2478
2479         RETURN(0);
2480 }
2481
2482 static int osc_cancel_unused(struct obd_export *exp,
2483                              struct lov_stripe_md *lsm, int flags, void *opaque)
2484 {
2485         struct obd_device *obd = class_exp2obd(exp);
2486         struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2487
2488         return ldlm_cli_cancel_unused(obd->obd_namespace, &res_id, flags,
2489                                       opaque);
2490 }
2491
2492 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2493                       unsigned long max_age)
2494 {
2495         struct obd_statfs *msfs;
2496         struct ptlrpc_request *request;
2497         int rc, size = sizeof(*osfs);
2498         ENTRY;
2499
2500         /* We could possibly pass max_age in the request (as an absolute
2501          * timestamp or a "seconds.usec ago") so the target can avoid doing
2502          * extra calls into the filesystem if that isn't necessary (e.g.
2503          * during mount that would help a bit).  Having relative timestamps
2504          * is not so great if request processing is slow, while absolute
2505          * timestamps are not ideal because they need time synchronization. */
2506         request = ptlrpc_prep_req(obd->u.cli.cl_import, OST_STATFS,0,NULL,NULL);
2507         if (!request)
2508                 RETURN(-ENOMEM);
2509
2510         request->rq_replen = lustre_msg_size(1, &size);
2511         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2512
2513         rc = ptlrpc_queue_wait(request);
2514         if (rc)
2515                 GOTO(out, rc);
2516
2517         msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2518                                   lustre_swab_obd_statfs);
2519         if (msfs == NULL) {
2520                 CERROR("Can't unpack obd_statfs\n");
2521                 GOTO(out, rc = -EPROTO);
2522         }
2523
2524         memcpy(osfs, msfs, sizeof(*osfs));
2525
2526         EXIT;
2527  out:
2528         ptlrpc_req_finished(request);
2529         return rc;
2530 }
2531
2532 /* Retrieve object striping information.
2533  *
2534  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2535  * the maximum number of OST indices which will fit in the user buffer.
2536  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2537  */
2538 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2539 {
2540         struct lov_user_md lum;
2541         struct lov_mds_md *lmmk;
2542         int rc, lmm_size;
2543         ENTRY;
2544
2545         if (!lsm)
2546                 RETURN(-ENODATA);
2547
2548         rc = copy_from_user(&lum, lump, sizeof(lum));
2549         if (rc)
2550                 RETURN(-EFAULT);
2551
2552         if (lum.lmm_magic != LOV_USER_MAGIC)
2553                 RETURN(-EINVAL);
2554
2555         if (lum.lmm_stripe_count < 1)
2556                 RETURN(-EOVERFLOW);
2557
2558         lmm_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2559         OBD_ALLOC(lmmk, lmm_size);
2560         if (!lmmk)
2561                 RETURN(-ENOMEM);
2562
2563         lmmk->lmm_stripe_count = 1;
2564         lmmk->lmm_object_id = lsm->lsm_object_id;
2565         lmmk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2566
2567         if (copy_to_user(lump, lmmk, lmm_size))
2568                 rc = -EFAULT;
2569
2570         OBD_FREE(lmmk, lmm_size);
2571
2572         RETURN(rc);
2573 }
2574
2575 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2576                          void *karg, void *uarg)
2577 {
2578         struct obd_device *obd = exp->exp_obd;
2579         struct obd_ioctl_data *data = karg;
2580         int err = 0;
2581         ENTRY;
2582
2583         MOD_INC_USE_COUNT;
2584
2585         switch (cmd) {
2586         case OBD_IOC_LOV_GET_CONFIG: {
2587                 char *buf;
2588                 struct lov_desc *desc;
2589                 struct obd_uuid uuid;
2590
2591                 buf = NULL;
2592                 len = 0;
2593                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2594                         GOTO(out, err = -EINVAL);
2595
2596                 data = (struct obd_ioctl_data *)buf;
2597
2598                 if (sizeof(*desc) > data->ioc_inllen1) {
2599                         OBD_FREE(buf, len);
2600                         GOTO(out, err = -EINVAL);
2601                 }
2602
2603                 if (data->ioc_inllen2 < sizeof(uuid)) {
2604                         OBD_FREE(buf, len);
2605                         GOTO(out, err = -EINVAL);
2606                 }
2607
2608                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2609                 desc->ld_tgt_count = 1;
2610                 desc->ld_active_tgt_count = 1;
2611                 desc->ld_default_stripe_count = 1;
2612                 desc->ld_default_stripe_size = 0;
2613                 desc->ld_default_stripe_offset = 0;
2614                 desc->ld_pattern = 0;
2615                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2616
2617                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2618
2619                 err = copy_to_user((void *)uarg, buf, len);
2620                 if (err)
2621                         err = -EFAULT;
2622                 obd_ioctl_freedata(buf, len);
2623                 GOTO(out, err);
2624         }
2625         case LL_IOC_LOV_SETSTRIPE:
2626                 err = obd_alloc_memmd(exp, karg);
2627                 if (err > 0)
2628                         err = 0;
2629                 GOTO(out, err);
2630         case LL_IOC_LOV_GETSTRIPE:
2631                 err = osc_getstripe(karg, uarg);
2632                 GOTO(out, err);
2633         case OBD_IOC_CLIENT_RECOVER:
2634                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2635                                             data->ioc_inlbuf1);
2636                 if (err > 0)
2637                         err = 0;
2638                 GOTO(out, err);
2639         case IOC_OSC_SET_ACTIVE:
2640                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2641                                                data->ioc_offset);
2642                 GOTO(out, err);
2643         default:
2644                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2645                 GOTO(out, err = -ENOTTY);
2646         }
2647 out:
2648         MOD_DEC_USE_COUNT;
2649         return err;
2650 }
2651
2652 static int osc_get_info(struct obd_export *exp, obd_count keylen,
2653                         void *key, __u32 *vallen, void *val)
2654 {
2655         ENTRY;
2656         if (!vallen || !val)
2657                 RETURN(-EFAULT);
2658
2659         if (keylen > strlen("lock_to_stripe") &&
2660             strcmp(key, "lock_to_stripe") == 0) {
2661                 __u32 *stripe = val;
2662                 *vallen = sizeof(*stripe);
2663                 *stripe = 0;
2664                 RETURN(0);
2665         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2666                 struct ptlrpc_request *req;
2667                 obd_id *reply;
2668                 char *bufs[1] = {key};
2669                 int rc;
2670                 req = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GET_INFO, 1,
2671                                       &keylen, bufs);
2672                 if (req == NULL)
2673                         RETURN(-ENOMEM);
2674
2675                 req->rq_replen = lustre_msg_size(1, vallen);
2676                 rc = ptlrpc_queue_wait(req);
2677                 if (rc)
2678                         GOTO(out, rc);
2679
2680                 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2681                                            lustre_swab_ost_last_id);
2682                 if (reply == NULL) {
2683                         CERROR("Can't unpack OST last ID\n");
2684                         GOTO(out, rc = -EPROTO);
2685                 }
2686                 *((obd_id *)val) = *reply;
2687         out:
2688                 ptlrpc_req_finished(req);
2689                 RETURN(rc);
2690         }
2691         RETURN(-EINVAL);
2692 }
2693
2694 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2695                         void *key, obd_count vallen, void *val)
2696 {
2697         struct ptlrpc_request *req;
2698         struct obd_import *imp = class_exp2cliimp(exp);
2699         struct llog_ctxt *ctxt;
2700         int rc, size = keylen;
2701         char *bufs[1] = {key};
2702         ENTRY;
2703
2704         if (keylen == strlen("next_id") &&
2705             memcmp(key, "next_id", strlen("next_id")) == 0) {
2706                 if (vallen != sizeof(obd_id))
2707                         RETURN(-EINVAL);
2708                 exp->u.eu_osc_data.oed_oscc.oscc_next_id = *((obd_id*)val) + 1;
2709                 CDEBUG(D_INODE, "%s: set oscc_next_id = "LPU64"\n",
2710                        exp->exp_obd->obd_name,
2711                        exp->u.eu_osc_data.oed_oscc.oscc_next_id);
2712
2713                 RETURN(0);
2714         }
2715
2716         if (keylen == strlen("growth_count") &&
2717             memcmp(key, "growth_count", strlen("growth_count")) == 0) {
2718                 if (vallen != sizeof(int))
2719                         RETURN(-EINVAL);
2720                 exp->u.eu_osc_data.oed_oscc.oscc_grow_count = *((int*)val);
2721                 RETURN(0);
2722         }
2723
2724         if (keylen == strlen("unlinked") &&
2725             memcmp(key, "unlinked", keylen) == 0) {
2726                 struct osc_creator *oscc = &exp->u.eu_osc_data.oed_oscc;
2727                 spin_lock(&oscc->oscc_lock);
2728                 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2729                 spin_unlock(&oscc->oscc_lock);
2730                 RETURN(0);
2731         }
2732
2733
2734         if (keylen == strlen("initial_recov") &&
2735             memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
2736                 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2737                 if (vallen != sizeof(int))
2738                         RETURN(-EINVAL);
2739                 imp->imp_initial_recov = *(int *)val;
2740                 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
2741                        exp->exp_obd->obd_name,
2742                        imp->imp_initial_recov);
2743                 RETURN(0);
2744         }
2745
2746         if (keylen < strlen("mds_conn") ||
2747             memcmp(key, "mds_conn", strlen("mds_conn")) != 0)
2748                 RETURN(-EINVAL);
2749
2750
2751         req = ptlrpc_prep_req(imp, OST_SET_INFO, 1, &size, bufs);
2752         if (req == NULL)
2753                 RETURN(-ENOMEM);
2754
2755         req->rq_replen = lustre_msg_size(0, NULL);
2756         rc = ptlrpc_queue_wait(req);
2757         ptlrpc_req_finished(req);
2758
2759         ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_ORIG_CTXT);
2760         if (ctxt) {
2761                 rc = llog_initiator_connect(ctxt);
2762                 if (rc)
2763                         RETURN(rc);
2764         }
2765
2766         imp->imp_server_timeout = 1;
2767         CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
2768         ptlrpc_pinger_add_import(imp);
2769
2770         RETURN(rc);
2771 }
2772
2773
2774 static struct llog_operations osc_size_repl_logops = {
2775         lop_cancel: llog_obd_repl_cancel
2776 };
2777
2778 static struct llog_operations osc_unlink_orig_logops;
2779 static int osc_llog_init(struct obd_device *obd, struct obd_device *tgt,
2780                         int count, struct llog_logid *logid)
2781 {
2782         int rc;
2783         ENTRY;
2784
2785         osc_unlink_orig_logops = llog_lvfs_ops;
2786         osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
2787         osc_unlink_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
2788         osc_unlink_orig_logops.lop_add = llog_obd_origin_add;
2789         osc_unlink_orig_logops.lop_connect = llog_origin_connect;
2790
2791         rc = llog_setup(obd, LLOG_UNLINK_ORIG_CTXT, tgt, count, logid,
2792                         &osc_unlink_orig_logops);
2793         if (rc)
2794                 RETURN(rc);
2795
2796         rc = llog_setup(obd, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
2797                         &osc_size_repl_logops);
2798         RETURN(rc);
2799 }
2800
2801 static int osc_llog_finish(struct obd_device *obd, int count)
2802 {
2803         int rc;
2804         ENTRY;
2805
2806         rc = llog_cleanup(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT));
2807         if (rc)
2808                 RETURN(rc);
2809
2810         rc = llog_cleanup(llog_get_context(obd, LLOG_SIZE_REPL_CTXT));
2811         RETURN(rc);
2812 }
2813
2814
2815 static int osc_connect(struct lustre_handle *exph,
2816                        struct obd_device *obd, struct obd_uuid *cluuid)
2817 {
2818         int rc;
2819         struct obd_export *exp;
2820
2821         rc = client_connect_import(exph, obd, cluuid);
2822
2823         if (obd->u.cli.cl_conn_count == 1) {
2824                 exp = class_conn2export(exph);
2825                 oscc_init(exp);
2826         }
2827
2828         return rc;
2829 }
2830
2831 static int osc_disconnect(struct obd_export *exp, int flags)
2832 {
2833         struct obd_device *obd = class_exp2obd(exp);
2834         struct llog_ctxt *ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
2835         int rc;
2836
2837         if (obd->u.cli.cl_conn_count == 1) {
2838                 /* flush any remaining cancel messages out to the target */
2839                 llog_sync(ctxt, exp);
2840
2841                 /* balance the conn2export for oscc in osc_connect */
2842                 class_export_put(exp);
2843         }
2844
2845         rc = client_disconnect_export(exp, flags);
2846         return rc;
2847 }
2848
2849 static int osc_lock_contains(struct obd_export *exp, struct lov_stripe_md *lsm,
2850                              struct ldlm_lock *lock, obd_off offset)
2851 {
2852         ENTRY;
2853         if (exp == NULL)
2854                 RETURN(-ENODEV);
2855
2856         if (lock->l_policy_data.l_extent.start <= offset &&
2857             lock->l_policy_data.l_extent.end >= offset)
2858                 RETURN(1);
2859         RETURN(0);
2860 }
2861
2862 static int osc_invalidate_import(struct obd_device *obd,
2863                                  struct obd_import *imp)
2864 {
2865         struct client_obd *cli;
2866         LASSERT(imp->imp_obd == obd);
2867         /* this used to try and tear down queued pages, but it was
2868          * not correctly implemented.  We'll have to do it again once
2869          * we call obd_invalidate_import() agian */
2870         /* XXX And we still need to do this */
2871
2872         /* Reset grants, too */
2873         cli = &obd->u.cli;
2874         spin_lock(&cli->cl_loi_list_lock);
2875         cli->cl_avail_grant = 0;
2876         cli->cl_lost_grant = 0;
2877         spin_unlock(&cli->cl_loi_list_lock);
2878
2879         RETURN(0);
2880 }
2881
2882 int osc_setup(struct obd_device *obd, obd_count len, void *buf)
2883 {
2884         int rc;
2885
2886         rc = ptlrpcd_addref();
2887         if (rc)
2888                 return rc;
2889
2890         rc = client_obd_setup(obd, len, buf);
2891         if (rc)
2892                 ptlrpcd_decref();
2893         RETURN(rc);
2894 }
2895
2896 int osc_cleanup(struct obd_device *obd, int flags)
2897 {
2898         int rc;
2899
2900         rc = client_obd_cleanup(obd, flags);
2901         ptlrpcd_decref();
2902         RETURN(rc);
2903 }
2904
2905
2906 struct obd_ops osc_obd_ops = {
2907         o_owner:        THIS_MODULE,
2908         o_attach:       osc_attach,
2909         o_detach:       osc_detach,
2910         o_setup:        osc_setup,
2911         o_cleanup:      osc_cleanup,
2912         o_connect:      osc_connect,
2913         o_disconnect:   osc_disconnect,
2914         o_statfs:       osc_statfs,
2915         o_packmd:       osc_packmd,
2916         o_unpackmd:     osc_unpackmd,
2917         o_create:       osc_create,
2918         o_destroy:      osc_destroy,
2919         o_getattr:      osc_getattr,
2920         o_getattr_async:osc_getattr_async,
2921         o_setattr:      osc_setattr,
2922         o_brw:          osc_brw,
2923         o_brw_async:    osc_brw_async,
2924         .o_prep_async_page =            osc_prep_async_page,
2925         .o_queue_async_io =             osc_queue_async_io,
2926         .o_set_async_flags =            osc_set_async_flags,
2927         .o_queue_sync_io =              osc_queue_sync_io,
2928         .o_trigger_sync_io =            osc_trigger_sync_io,
2929         .o_teardown_async_page =        osc_teardown_async_page,
2930         o_punch:        osc_punch,
2931         o_sync:         osc_sync,
2932         o_enqueue:      osc_enqueue,
2933         o_match:        osc_match,
2934         o_change_cbdata:osc_change_cbdata,
2935         o_cancel:       osc_cancel,
2936         o_cancel_unused:osc_cancel_unused,
2937         o_iocontrol:    osc_iocontrol,
2938         o_get_info:     osc_get_info,
2939         o_set_info:     osc_set_info,
2940         o_lock_contains:osc_lock_contains,
2941         o_invalidate_import: osc_invalidate_import,
2942         o_llog_init:    osc_llog_init,
2943         o_llog_finish:  osc_llog_finish,
2944 };
2945
2946 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2947 struct obd_ops sanosc_obd_ops = {
2948         o_owner:        THIS_MODULE,
2949         o_attach:       osc_attach,
2950         o_detach:       osc_detach,
2951         o_cleanup:      client_obd_cleanup,
2952         o_connect:      osc_connect,
2953         o_disconnect:   client_disconnect_export,
2954         o_statfs:       osc_statfs,
2955         o_packmd:       osc_packmd,
2956         o_unpackmd:     osc_unpackmd,
2957         o_create:       osc_real_create,
2958         o_destroy:      osc_destroy,
2959         o_getattr:      osc_getattr,
2960         o_getattr_async:osc_getattr_async,
2961         o_setattr:      osc_setattr,
2962         o_setup:        client_sanobd_setup,
2963         o_brw:          sanosc_brw,
2964         o_punch:        osc_punch,
2965         o_sync:         osc_sync,
2966         o_enqueue:      osc_enqueue,
2967         o_match:        osc_match,
2968         o_change_cbdata:osc_change_cbdata,
2969         o_cancel:       osc_cancel,
2970         o_cancel_unused:osc_cancel_unused,
2971         o_iocontrol:    osc_iocontrol,
2972         o_lock_contains:osc_lock_contains,
2973         o_invalidate_import: osc_invalidate_import,
2974         o_llog_init:    osc_llog_init,
2975         o_llog_finish:  osc_llog_finish,
2976 };
2977 #endif
2978
2979 int __init osc_init(void)
2980 {
2981         struct lprocfs_static_vars lvars, sanlvars;
2982         int rc;
2983         ENTRY;
2984
2985         lprocfs_init_vars(osc, &lvars);
2986         lprocfs_init_vars(osc, &sanlvars);
2987
2988         rc = class_register_type(&osc_obd_ops, lvars.module_vars,
2989                                  LUSTRE_OSC_NAME);
2990         if (rc)
2991                 RETURN(rc);
2992
2993 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2994         rc = class_register_type(&sanosc_obd_ops, sanlvars.module_vars,
2995                                  LUSTRE_SANOSC_NAME);
2996         if (rc)
2997                 class_unregister_type(LUSTRE_OSC_NAME);
2998 #endif
2999
3000         RETURN(rc);
3001 }
3002
3003 static void /*__exit*/ osc_exit(void)
3004 {
3005 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3006         class_unregister_type(LUSTRE_SANOSC_NAME);
3007 #endif
3008         class_unregister_type(LUSTRE_OSC_NAME);
3009 }
3010
3011 #ifdef __KERNEL__
3012 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3013 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3014 MODULE_LICENSE("GPL");
3015
3016 module_init(osc_init);
3017 module_exit(osc_exit);
3018 #endif