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         /* XXX obd_brw_internal() might reuse obdo in it's loop thus
 541          * hit the following assert. any actual meaning of this? temporarily
 542          * disable it.
 543          * in kernel mode, probably VFS will prevent it happen.
 544          */
 545         //LASSERT(!(oa->o_valid & bits));
 546
 547         oa->o_valid |= bits;
 548         spin_lock(&cli->cl_loi_list_lock);
 549         oa->o_dirty = cli->cl_dirty;
 550         oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
 551         oa->o_grant = cli->cl_avail_grant;
 552         oa->o_dropped = cli->cl_lost_grant;
 553         cli->cl_lost_grant = 0;
 554         spin_unlock(&cli->cl_loi_list_lock);
 555         CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
 556                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 557 }
 558
 559 /* caller must hold loi_list_lock */
 560 static void osc_consume_write_grant(struct client_obd *cli,
 561                                     struct osc_async_page *oap)
 562 {
 563         cli->cl_dirty += PAGE_SIZE;
 564         cli->cl_avail_grant -= PAGE_SIZE;
 565         oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
 566         CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
 567         LASSERT(cli->cl_avail_grant >= 0);
 568 }
 569
 570 /* caller must hold loi_list_lock */
 571 void osc_wake_cache_waiters(struct client_obd *cli)
 572 {
 573         struct list_head *l, *tmp;
 574         struct osc_cache_waiter *ocw;
 575
 576         list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
 577                 /* if we can't dirty more, we must wait until some is written */
 578                 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
 579                         CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
 580                                cli->cl_dirty, cli->cl_dirty_max);
 581                         return;
 582                 }
 583
 584                 /* if still dirty cache but no grant wait for pending RPCs that
 585                  * may yet return us some grant before doing sync writes */
 586                 if (cli->cl_brw_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
 587                         CDEBUG(D_CACHE, "%d BRWs in flight, no grant\n",
 588                                cli->cl_brw_in_flight);
 589                         return;
 590                 }
 591
 592                 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
 593                 list_del_init(&ocw->ocw_entry);
 594                 if (cli->cl_avail_grant < PAGE_SIZE) {
 595                         /* no more RPCs in flight to return grant, do sync IO */
 596                         ocw->ocw_rc = -EDQUOT;
 597                         CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
 598                 } else {
 599                         osc_consume_write_grant(cli, ocw->ocw_oap);
 600                 }
 601                 wake_up(&ocw->ocw_waitq);
 602         }
 603
 604         EXIT;
 605 }
 606
 607 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 608 {
 609         spin_lock(&cli->cl_loi_list_lock);
 610         CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
 611         cli->cl_avail_grant += body->oa.o_grant;
 612         /* waiters are woken in brw_interpret_oap */
 613         spin_unlock(&cli->cl_loi_list_lock);
 614 }
 615
 616 /* We assume that the reason this OSC got a short read is because it read
 617  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
 618  * via the LOV, and it _knows_ it's reading inside the file, it's just that
 619  * this stripe never got written at or beyond this stripe offset yet. */
 620 static void handle_short_read(int nob_read, obd_count page_count,
 621                               struct brw_page *pga)
 622 {
 623         char *ptr;
 624
 625         /* skip bytes read OK */
 626         while (nob_read > 0) {
 627                 LASSERT (page_count > 0);
 628
 629                 if (pga->count > nob_read) {
 630                         /* EOF inside this page */
 631                         ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
 632                         memset(ptr + nob_read, 0, pga->count - nob_read);
 633                         kunmap(pga->pg);
 634                         page_count--;
 635                         pga++;
 636                         break;
 637                 }
 638
 639                 nob_read -= pga->count;
 640                 page_count--;
 641                 pga++;
 642         }
 643
 644         /* zero remaining pages */
 645         while (page_count-- > 0) {
 646                 ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
 647                 memset(ptr, 0, pga->count);
 648                 kunmap(pga->pg);
 649                 pga++;
 650         }
 651 }
 652
 653 static int check_write_rcs(struct ptlrpc_request *request,
 654                            int requested_nob, int niocount,
 655                            obd_count page_count, struct brw_page *pga)
 656 {
 657         int    *remote_rcs, i;
 658
 659         /* return error if any niobuf was in error */
 660         remote_rcs = lustre_swab_repbuf(request, 1,
 661                                         sizeof(*remote_rcs) * niocount, NULL);
 662         if (remote_rcs == NULL) {
 663                 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
 664                 return(-EPROTO);
 665         }
 666         if (lustre_msg_swabbed(request->rq_repmsg))
 667                 for (i = 0; i < niocount; i++)
 668                         __swab32s(&remote_rcs[i]);
 669
 670         for (i = 0; i < niocount; i++) {
 671                 if (remote_rcs[i] < 0)
 672                         return(remote_rcs[i]);
 673
 674                 if (remote_rcs[i] != 0) {
 675                         CERROR("rc[%d] invalid (%d) req %p\n",
 676                                 i, remote_rcs[i], request);
 677                         return(-EPROTO);
 678                 }
 679         }
 680
 681         if (request->rq_bulk->bd_nob_transferred != requested_nob) {
 682                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
 683                        requested_nob, request->rq_bulk->bd_nob_transferred);
 684                 return(-EPROTO);
 685         }
 686
 687         return (0);
 688 }
 689
 690 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
 691 {
 692         if (p1->flag != p2->flag) {
 693                 unsigned mask = ~OBD_BRW_FROM_GRANT;
 694
 695                 /* warn if we try to combine flags that we don't know to be
 696                  * safe to combine */
 697                 if ((p1->flag & mask) != (p2->flag & mask))
 698                         CERROR("is it ok to have flags 0x%x and 0x%x in the "
 699                                "same brw?\n", p1->flag, p2->flag);
 700                 return 0;
 701         }
 702
 703         return (p1->off + p1->count == p2->off);
 704 }
 705
 706 #if CHECKSUM_BULK
 707 static obd_count cksum_pages(int nob, obd_count page_count,
 708                              struct brw_page *pga)
 709 {
 710         obd_count cksum = 0;
 711         char *ptr;
 712
 713         while (nob > 0) {
 714                 LASSERT (page_count > 0);
 715
 716                 ptr = kmap(pga->pg);
 717                 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
 718                              pga->count > nob ? nob : pga->count);
 719                 kunmap(pga->pg);
 720
 721                 nob -= pga->count;
 722                 page_count--;
 723                 pga++;
 724         }
 725
 726         return (cksum);
 727 }
 728 #endif
 729
 730 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
 731                                 struct lov_stripe_md *lsm, obd_count page_count,
 732                                 struct brw_page *pga, int *requested_nobp,
 733                                 int *niocountp, struct ptlrpc_request **reqp)
 734 {
 735         struct ptlrpc_request   *req;
 736         struct ptlrpc_bulk_desc *desc;
 737         struct client_obd       *cli = &imp->imp_obd->u.cli;
 738         struct ost_body         *body;
 739         struct obd_ioobj        *ioobj;
 740         struct niobuf_remote    *niobuf;
 741         unsigned long            flags;
 742         int                      niocount;
 743         int                      size[3];
 744         int                      i;
 745         int                      requested_nob;
 746         int                      opc;
 747         int                      rc;
 748
 749         opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
 750
 751         for (niocount = i = 1; i < page_count; i++)
 752                 if (!can_merge_pages(&pga[i - 1], &pga[i]))
 753                         niocount++;
 754
 755         size[0] = sizeof(*body);
 756         size[1] = sizeof(*ioobj);
 757         size[2] = niocount * sizeof(*niobuf);
 758
 759         req = ptlrpc_prep_req(imp, opc, 3, size, NULL);
 760         if (req == NULL)
 761                 return (-ENOMEM);
 762
 763         if (opc == OST_WRITE)
 764                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 765                                              BULK_GET_SOURCE, OST_BULK_PORTAL);
 766         else
 767                 desc = ptlrpc_prep_bulk_imp (req, page_count,
 768                                              BULK_PUT_SINK, OST_BULK_PORTAL);
 769         if (desc == NULL)
 770                 GOTO(out, rc = -ENOMEM);
 771         /* NB request now owns desc and will free it when it gets freed */
 772
 773         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 774         ioobj = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*ioobj));
 775         niobuf = lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf));
 776
 777         memcpy(&body->oa, oa, sizeof(*oa));
 778
 779         obdo_to_ioobj(oa, ioobj);
 780         ioobj->ioo_bufcnt = niocount;
 781
 782         LASSERT (page_count > 0);
 783         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
 784                 struct brw_page *pg = &pga[i];
 785                 struct brw_page *pg_prev = pg - 1;
 786
 787                 LASSERT(pg->count > 0);
 788                 LASSERT((pg->off & ~PAGE_MASK) + pg->count <= PAGE_SIZE);
 789                 LASSERTF(i == 0 || pg->off > pg_prev->off,
 790                          "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
 791                          " prev_pg %p [pri %lu ind %lu] off "LPU64,
 792                          i, page_count,
 793                          pg->pg, pg->pg->private, pg->pg->index, pg->off,
 794                          pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
 795                                  pg_prev->off);
 796
 797                 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~PAGE_MASK,
 798                                       pg->count);
 799                 requested_nob += pg->count;
 800
 801                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
 802                         niobuf--;
 803                         niobuf->len += pg->count;
 804                 } else {
 805                         niobuf->offset = pg->off;
 806                         niobuf->len    = pg->count;
 807                         niobuf->flags  = pg->flag;
 808                 }
 809         }
 810
 811         LASSERT((void *)(niobuf - niocount) ==
 812                 lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf)));
 813         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
 814         spin_lock_irqsave(&req->rq_lock, flags);
 815         req->rq_no_resend = 1;
 816         spin_unlock_irqrestore(&req->rq_lock, flags);
 817
 818         /* size[0] still sizeof (*body) */
 819         if (opc == OST_WRITE) {
 820 #if CHECKSUM_BULK
 821                 body->oa.o_valid |= OBD_MD_FLCKSUM;
 822                 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
 823 #endif
 824                 /* 1 RC per niobuf */
 825                 size[1] = sizeof(__u32) * niocount;
 826                 req->rq_replen = lustre_msg_size(2, size);
 827         } else {
 828                 /* 1 RC for the whole I/O */
 829                 req->rq_replen = lustre_msg_size(1, size);
 830         }
 831
 832         *niocountp = niocount;
 833         *requested_nobp = requested_nob;
 834         *reqp = req;
 835         return (0);
 836
 837  out:
 838         ptlrpc_req_finished (req);
 839         return (rc);
 840 }
 841
 842 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
 843                                 int requested_nob, int niocount,
 844                                 obd_count page_count, struct brw_page *pga,
 845                                 int rc)
 846 {
 847         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 848         struct ost_body *body;
 849         ENTRY;
 850
 851         if (rc < 0)
 852                 RETURN(rc);
 853
 854         body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
 855         if (body == NULL) {
 856                 CERROR ("Can't unpack body\n");
 857                 RETURN(-EPROTO);
 858         }
 859
 860         osc_update_grant(cli, body);
 861
 862         if (req->rq_reqmsg->opc == OST_WRITE) {
 863                 if (rc > 0) {
 864                         CERROR ("Unexpected +ve rc %d\n", rc);
 865                         RETURN(-EPROTO);
 866                 }
 867                 LASSERT (req->rq_bulk->bd_nob == requested_nob);
 868
 869                 RETURN(check_write_rcs(req, requested_nob, niocount,
 870                                        page_count, pga));
 871         }
 872
 873         if (rc > requested_nob) {
 874                 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
 875                 RETURN(-EPROTO);
 876         }
 877
 878         if (rc != req->rq_bulk->bd_nob_transferred) {
 879                 CERROR ("Unexpected rc %d (%d transferred)\n",
 880                         rc, req->rq_bulk->bd_nob_transferred);
 881                 return (-EPROTO);
 882         }
 883
 884         if (rc < requested_nob)
 885                 handle_short_read(rc, page_count, pga);
 886
 887         memcpy(oa, &body->oa, sizeof(*oa));
 888
 889 #if CHECKSUM_BULK
 890         if (oa->o_valid & OBD_MD_FLCKSUM) {
 891                 const struct ptlrpc_peer *peer =
 892                         &req->rq_import->imp_connection->c_peer;
 893                 static int cksum_counter;
 894                 obd_count server_cksum = oa->o_cksum;
 895                 obd_count cksum = cksum_pages(rc, page_count, pga);
 896                 char str[PTL_NALFMT_SIZE];
 897
 898                 portals_nid2str(peer->peer_ni->pni_number, peer->peer_nid, str);
 899
 900                 cksum_counter++;
 901                 if (server_cksum != cksum) {
 902                         CERROR("Bad checksum: server %x, client %x, server NID "
 903                                LPX64" (%s)\n", server_cksum, cksum,
 904                                peer->peer_nid, str);
 905                         cksum_counter = 0;
 906                         oa->o_cksum = cksum;
 907                 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
 908                         CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
 909                               cksum_counter, peer->peer_nid, str, cksum);
 910                 }
 911         } else {
 912                 static int cksum_missed;
 913
 914                 cksum_missed++;
 915                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
 916                         CERROR("Request checksum %u from "LPX64", no reply\n",
 917                                cksum_missed,
 918                                req->rq_import->imp_connection->c_peer.peer_nid);
 919         }
 920 #endif
 921         RETURN(0);
 922 }
 923
 924 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
 925                             struct lov_stripe_md *lsm,
 926                             obd_count page_count, struct brw_page *pga)
 927 {
 928         int                    requested_nob;
 929         int                    niocount;
 930         struct ptlrpc_request *request;
 931         int                    rc;
 932         ENTRY;
 933
 934 restart_bulk:
 935         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
 936                                   page_count, pga, &requested_nob, &niocount,
 937                                   &request);
 938         /* NB ^ sets rq_no_resend */
 939
 940         if (rc != 0)
 941                 return (rc);
 942
 943         rc = ptlrpc_queue_wait(request);
 944
 945         if (rc == -ETIMEDOUT && request->rq_resend) {
 946                 DEBUG_REQ(D_HA, request,  "BULK TIMEOUT");
 947                 ptlrpc_req_finished(request);
 948                 goto restart_bulk;
 949         }
 950
 951         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
 952                                   page_count, pga, rc);
 953
 954         ptlrpc_req_finished(request);
 955         RETURN (rc);
 956 }
 957
 958 static int brw_interpret(struct ptlrpc_request *request,
 959                          struct osc_brw_async_args *aa, int rc)
 960 {
 961         struct obdo *oa      = aa->aa_oa;
 962         int requested_nob    = aa->aa_requested_nob;
 963         int niocount         = aa->aa_nio_count;
 964         obd_count page_count = aa->aa_page_count;
 965         struct brw_page *pga = aa->aa_pga;
 966         ENTRY;
 967
 968         /* XXX bug 937 here */
 969         if (rc == -ETIMEDOUT && request->rq_resend) {
 970                 DEBUG_REQ(D_HA, request,  "BULK TIMEOUT");
 971                 LBUG(); /* re-send.  later. */
 972                 //goto restart_bulk;
 973         }
 974
 975         rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
 976                                   page_count, pga, rc);
 977         RETURN (rc);
 978 }
 979
 980 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
 981                           struct lov_stripe_md *lsm, obd_count page_count,
 982                           struct brw_page *pga, struct ptlrpc_request_set *set)
 983 {
 984         struct ptlrpc_request     *request;
 985         int                        requested_nob;
 986         int                        nio_count;
 987         struct osc_brw_async_args *aa;
 988         int                        rc;
 989         ENTRY;
 990
 991         rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
 992                                   page_count, pga, &requested_nob, &nio_count,
 993                                   &request);
 994         /* NB ^ sets rq_no_resend */
 995
 996         if (rc == 0) {
 997                 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
 998                 aa = (struct osc_brw_async_args *)&request->rq_async_args;
 999                 aa->aa_oa = oa;
1000                 aa->aa_requested_nob = requested_nob;
1001                 aa->aa_nio_count = nio_count;
1002                 aa->aa_page_count = page_count;
1003                 aa->aa_pga = pga;
1004
1005                 request->rq_interpret_reply = brw_interpret;
1006                 ptlrpc_set_add_req(set, request);
1007         }
1008         RETURN (rc);
1009 }
1010
1011 #ifndef min_t
1012 #define min_t(type,x,y) \
1013         ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
1014 #endif
1015
1016 /*
1017  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1018  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1019  * fine for our small page arrays and doesn't require allocation.  its an
1020  * insertion sort that swaps elements that are strides apart, shrinking the
1021  * stride down until its '1' and the array is sorted.
1022  */
1023 static void sort_brw_pages(struct brw_page *array, int num)
1024 {
1025         int stride, i, j;
1026         struct brw_page tmp;
1027
1028         if (num == 1)
1029                 return;
1030         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1031                 ;
1032
1033         do {
1034                 stride /= 3;
1035                 for (i = stride ; i < num ; i++) {
1036                         tmp = array[i];
1037                         j = i;
1038                         while (j >= stride && array[j - stride].off > tmp.off) {
1039                                 array[j] = array[j - stride];
1040                                 j -= stride;
1041                         }
1042                         array[j] = tmp;
1043                 }
1044         } while (stride > 1);
1045 }
1046
1047 /* make sure we the regions we're passing to elan don't violate its '4
1048  * fragments' constraint.  portal headers are a fragment, all full
1049  * PAGE_SIZE long pages count as 1 fragment, and each partial page
1050  * counts as a fragment.  I think.  see bug 934. */
1051 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1052 {
1053         int frags_left = 3;
1054         int saw_whole_frag = 0;
1055         int i;
1056
1057         for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1058                 if (pg->count == PAGE_SIZE) {
1059                         if (!saw_whole_frag) {
1060                                 saw_whole_frag = 1;
1061                                 frags_left--;
1062                         }
1063                 } else {
1064                         frags_left--;
1065                 }
1066         }
1067         return i;
1068 }
1069
1070 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1071                    struct lov_stripe_md *md, obd_count page_count,
1072                    struct brw_page *pga, struct obd_trans_info *oti)
1073 {
1074         ENTRY;
1075
1076         if (cmd == OBD_BRW_CHECK) {
1077                 /* The caller just wants to know if there's a chance that this
1078                  * I/O can succeed */
1079                 struct obd_import *imp = class_exp2cliimp(exp);
1080
1081                 if (imp == NULL || imp->imp_invalid)
1082                         RETURN(-EIO);
1083                 RETURN(0);
1084         }
1085
1086         while (page_count) {
1087                 obd_count pages_per_brw;
1088                 int rc;
1089
1090                 if (page_count > OSC_BRW_MAX_IOV)
1091                         pages_per_brw = OSC_BRW_MAX_IOV;
1092                 else
1093                         pages_per_brw = page_count;
1094
1095                 sort_brw_pages(pga, pages_per_brw);
1096                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1097
1098                 rc = osc_brw_internal(cmd, exp, oa, md, pages_per_brw, pga);
1099
1100                 if (rc != 0)
1101                         RETURN(rc);
1102
1103                 page_count -= pages_per_brw;
1104                 pga += pages_per_brw;
1105         }
1106         RETURN(0);
1107 }
1108
1109 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1110                          struct lov_stripe_md *md, obd_count page_count,
1111                          struct brw_page *pga, struct ptlrpc_request_set *set,
1112                          struct obd_trans_info *oti)
1113 {
1114         ENTRY;
1115
1116         if (cmd == OBD_BRW_CHECK) {
1117                 /* The caller just wants to know if there's a chance that this
1118                  * I/O can succeed */
1119                 struct obd_import *imp = class_exp2cliimp(exp);
1120
1121                 if (imp == NULL || imp->imp_invalid)
1122                         RETURN(-EIO);
1123                 RETURN(0);
1124         }
1125
1126         while (page_count) {
1127                 obd_count pages_per_brw;
1128                 int rc;
1129
1130                 if (page_count > OSC_BRW_MAX_IOV)
1131                         pages_per_brw = OSC_BRW_MAX_IOV;
1132                 else
1133                         pages_per_brw = page_count;
1134
1135                 sort_brw_pages(pga, pages_per_brw);
1136                 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1137
1138                 rc = async_internal(cmd, exp, oa, md, pages_per_brw, pga, set);
1139
1140                 if (rc != 0)
1141                         RETURN(rc);
1142
1143                 page_count -= pages_per_brw;
1144                 pga += pages_per_brw;
1145         }
1146         RETURN(0);
1147 }
1148
1149 static void osc_check_rpcs(struct client_obd *cli);
1150 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1151                            int sent);
1152 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1153 static void lop_update_pending(struct client_obd *cli,
1154                                struct loi_oap_pages *lop, int cmd, int delta);
1155
1156 /* this is called when a sync waiter receives an interruption.  Its job is to
1157  * get the caller woken as soon as possible.  If its page hasn't been put in an
1158  * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
1159  * desiring interruption which will forcefully complete the rpc once the rpc
1160  * has timed out */
1161 static void osc_occ_interrupted(struct osic_callback_context *occ)
1162 {
1163         struct osc_async_page *oap;
1164         struct loi_oap_pages *lop;
1165         struct lov_oinfo *loi;
1166         ENTRY;
1167
1168         /* XXX member_of() */
1169         oap = list_entry(occ, struct osc_async_page, oap_occ);
1170
1171         spin_lock(&oap->oap_cli->cl_loi_list_lock);
1172
1173         oap->oap_interrupted = 1;
1174
1175         /* ok, it's been put in an rpc. */
1176         if (oap->oap_request != NULL) {
1177                 ptlrpc_mark_interrupted(oap->oap_request);
1178                 ptlrpcd_wake();
1179                 GOTO(unlock, 0);
1180         }
1181
1182         /* we don't get interruption callbacks until osc_trigger_sync_io()
1183          * has been called and put the sync oaps in the pending/urgent lists.*/
1184         if (!list_empty(&oap->oap_pending_item)) {
1185                 list_del_init(&oap->oap_pending_item);
1186                 if (oap->oap_async_flags & ASYNC_URGENT)
1187                         list_del_init(&oap->oap_urgent_item);
1188
1189                 loi = oap->oap_loi;
1190                 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1191                         &loi->loi_write_lop : &loi->loi_read_lop;
1192                 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1193                 loi_list_maint(oap->oap_cli, oap->oap_loi);
1194
1195                 osic_complete_one(oap->oap_osic, &oap->oap_occ, 0);
1196                 oap->oap_osic = NULL;
1197         }
1198
1199 unlock:
1200         spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1201 }
1202
1203 /* this must be called holding the loi list lock to give coverage to exit_cache,
1204  * async_flag maintenance, and oap_request */
1205 static void osc_complete_oap(struct client_obd *cli,
1206                              struct osc_async_page *oap, int sent, int rc)
1207 {
1208         osc_exit_cache(cli, oap, sent);
1209         oap->oap_async_flags = 0;
1210         oap->oap_interrupted = 0;
1211
1212         if (oap->oap_request != NULL) {
1213                 ptlrpc_req_finished(oap->oap_request);
1214                 oap->oap_request = NULL;
1215         }
1216
1217         if (oap->oap_osic) {
1218                 osic_complete_one(oap->oap_osic, &oap->oap_occ, rc);
1219                 oap->oap_osic = NULL;
1220                 EXIT;
1221                 return;
1222         }
1223
1224         oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1225                                            rc);
1226 }
1227
1228 static int brw_interpret_oap(struct ptlrpc_request *request,
1229                              struct osc_brw_async_args *aa, int rc)
1230 {
1231         struct osc_async_page *oap;
1232         struct client_obd *cli;
1233         struct list_head *pos, *n;
1234         ENTRY;
1235
1236         CDEBUG(D_INODE, "request %p aa %p\n", request, aa);
1237
1238         rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1239                                   aa->aa_nio_count, aa->aa_page_count,
1240                                   aa->aa_pga, rc);
1241
1242         cli = aa->aa_cli;
1243         /* in failout recovery we ignore writeback failure and want
1244          * to just tell llite to unlock the page and continue */
1245         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1246                 rc = 0;
1247
1248         spin_lock(&cli->cl_loi_list_lock);
1249
1250         /* We need to decrement before osc_complete_oap->osc_wake_cache_waiters
1251          * is called so we know whether to go to sync BRWs or wait for more
1252          * RPCs to complete */
1253         cli->cl_brw_in_flight--;
1254
1255         /* the caller may re-use the oap after the completion call so
1256          * we need to clean it up a little */
1257         list_for_each_safe(pos, n, &aa->aa_oaps) {
1258                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1259
1260                 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1261                        //oap->oap_page, oap->oap_page->index, oap);
1262
1263                 list_del_init(&oap->oap_rpc_item);
1264                 osc_complete_oap(cli, oap, 1, rc);
1265         }
1266
1267         osc_wake_cache_waiters(cli);
1268         osc_check_rpcs(cli);
1269
1270         spin_unlock(&cli->cl_loi_list_lock);
1271
1272         obdo_free(aa->aa_oa);
1273         OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1274
1275         RETURN(0);
1276 }
1277
1278 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1279                                             struct list_head *rpc_list,
1280                                             int page_count, int cmd)
1281 {
1282         struct ptlrpc_request *req;
1283         struct brw_page *pga = NULL;
1284         int requested_nob, nio_count;
1285         struct osc_brw_async_args *aa;
1286         struct obdo *oa = NULL;
1287         struct obd_async_page_ops *ops = NULL;
1288         void *caller_data = NULL;
1289         struct list_head *pos;
1290         int i, rc;
1291
1292         LASSERT(!list_empty(rpc_list));
1293
1294         OBD_ALLOC(pga, sizeof(*pga) * page_count);
1295         if (pga == NULL)
1296                 RETURN(ERR_PTR(-ENOMEM));
1297
1298         oa = obdo_alloc();
1299         if (oa == NULL)
1300                 GOTO(out, req = ERR_PTR(-ENOMEM));
1301
1302         i = 0;
1303         list_for_each(pos, rpc_list) {
1304                 struct osc_async_page *oap;
1305
1306                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1307                 if (ops == NULL) {
1308                         ops = oap->oap_caller_ops;
1309                         caller_data = oap->oap_caller_data;
1310                 }
1311                 pga[i].off = oap->oap_obj_off + oap->oap_page_off;
1312                 pga[i].pg = oap->oap_page;
1313                 pga[i].count = oap->oap_count;
1314                 pga[i].flag = oap->oap_brw_flags;
1315                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1316                        pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1317                 i++;
1318         }
1319
1320         /* always get the data for the obdo for the rpc */
1321         LASSERT(ops != NULL);
1322         ops->ap_fill_obdo(caller_data, cmd, oa);
1323
1324         sort_brw_pages(pga, page_count);
1325         rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1326                                   pga, &requested_nob, &nio_count, &req);
1327         if (rc != 0) {
1328                 CERROR("prep_req failed: %d\n", rc);
1329                 GOTO(out, req = ERR_PTR(rc));
1330         }
1331
1332         LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1333         aa = (struct osc_brw_async_args *)&req->rq_async_args;
1334         aa->aa_oa = oa;
1335         aa->aa_requested_nob = requested_nob;
1336         aa->aa_nio_count = nio_count;
1337         aa->aa_page_count = page_count;
1338         aa->aa_pga = pga;
1339         aa->aa_cli = cli;
1340
1341 out:
1342         if (IS_ERR(req)) {
1343                 if (oa)
1344                         obdo_free(oa);
1345                 if (pga)
1346                         OBD_FREE(pga, sizeof(*pga) * page_count);
1347         }
1348         RETURN(req);
1349 }
1350
1351 static void lop_update_pending(struct client_obd *cli,
1352                                struct loi_oap_pages *lop, int cmd, int delta)
1353 {
1354         lop->lop_num_pending += delta;
1355         if (cmd == OBD_BRW_WRITE)
1356                 cli->cl_pending_w_pages += delta;
1357         else
1358                 cli->cl_pending_r_pages += delta;
1359 }
1360
1361 /* the loi lock is held across this function but it's allowed to release
1362  * and reacquire it during its work */
1363 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1364                             int cmd, struct loi_oap_pages *lop)
1365 {
1366         struct ptlrpc_request *request;
1367         obd_count page_count = 0;
1368         struct list_head *tmp, *pos;
1369         struct osc_async_page *oap = NULL;
1370         struct osc_brw_async_args *aa;
1371         struct obd_async_page_ops *ops;
1372         LIST_HEAD(rpc_list);
1373         ENTRY;
1374
1375         /* first we find the pages we're allowed to work with */
1376         list_for_each_safe(pos, tmp, &lop->lop_pending) {
1377                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1378                 ops = oap->oap_caller_ops;
1379
1380                 LASSERT(oap->oap_magic == OAP_MAGIC);
1381
1382                 /* in llite being 'ready' equates to the page being locked
1383                  * until completion unlocks it.  commit_write submits a page
1384                  * as not ready because its unlock will happen unconditionally
1385                  * as the call returns.  if we race with commit_write giving
1386                  * us that page we dont' want to create a hole in the page
1387                  * stream, so we stop and leave the rpc to be fired by
1388                  * another dirtier or kupdated interval (the not ready page
1389                  * will still be on the dirty list).  we could call in
1390                  * at the end of ll_file_write to process the queue again. */
1391                 if (!(oap->oap_async_flags & ASYNC_READY)) {
1392                         int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1393                         if (rc < 0)
1394                                 CDEBUG(D_INODE, "oap %p page %p returned %d "
1395                                                 "instead of ready\n", oap,
1396                                                 oap->oap_page, rc);
1397                         switch (rc) {
1398                         case -EAGAIN:
1399                                 /* llite is telling us that the page is still
1400                                  * in commit_write and that we should try
1401                                  * and put it in an rpc again later.  we
1402                                  * break out of the loop so we don't create
1403                                  * a hole in the sequence of pages in the rpc
1404                                  * stream.*/
1405                                 pos = NULL;
1406                                 break;
1407                         case -EINTR:
1408                                 /* the io isn't needed.. tell the checks
1409                                  * below to complete the rpc with EINTR */
1410                                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1411                                 oap->oap_count = -EINTR;
1412                                 break;
1413                         case 0:
1414                                 oap->oap_async_flags |= ASYNC_READY;
1415                                 break;
1416                         default:
1417                                 LASSERTF(0, "oap %p page %p returned %d "
1418                                             "from make_ready\n", oap,
1419                                             oap->oap_page, rc);
1420                                 break;
1421                         }
1422                 }
1423                 if (pos == NULL)
1424                         break;
1425
1426                 /* take the page out of our book-keeping */
1427                 list_del_init(&oap->oap_pending_item);
1428                 lop_update_pending(cli, lop, cmd, -1);
1429                 if (!list_empty(&oap->oap_urgent_item))
1430                         list_del_init(&oap->oap_urgent_item);
1431
1432                 /* ask the caller for the size of the io as the rpc leaves. */
1433                 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1434                         oap->oap_count =
1435                                 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1436                 if (oap->oap_count <= 0) {
1437                         CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1438                                oap->oap_count);
1439                         osc_complete_oap(cli, oap, 0, oap->oap_count);
1440                         continue;
1441                 }
1442
1443                 /* now put the page back in our accounting */
1444                 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1445                 if (++page_count >= cli->cl_max_pages_per_rpc)
1446                         break;
1447         }
1448
1449         osc_wake_cache_waiters(cli);
1450
1451         if (page_count == 0)
1452                 RETURN(0);
1453
1454         loi_list_maint(cli, loi);
1455         spin_unlock(&cli->cl_loi_list_lock);
1456
1457         request = osc_build_req(cli, &rpc_list, page_count, cmd);
1458         if (IS_ERR(request)) {
1459                 /* this should happen rarely and is pretty bad, it makes the
1460                  * pending list not follow the dirty order */
1461                 spin_lock(&cli->cl_loi_list_lock);
1462                 list_for_each_safe(pos, tmp, &rpc_list) {
1463                         oap = list_entry(pos, struct osc_async_page,
1464                                          oap_rpc_item);
1465                         list_del_init(&oap->oap_rpc_item);
1466
1467                         /* queued sync pages can be torn down while the pages
1468                          * were between the pending list and the rpc */
1469                         if (oap->oap_interrupted) {
1470                                 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1471                                 osc_complete_oap(cli, oap, 0, oap->oap_count);
1472                                 continue;
1473                         }
1474
1475                         /* put the page back in the loi/lop lists */
1476                         list_add_tail(&oap->oap_pending_item,
1477                                       &lop->lop_pending);
1478                         lop_update_pending(cli, lop, cmd, 1);
1479                         if (oap->oap_async_flags & ASYNC_URGENT)
1480                                 list_add(&oap->oap_urgent_item,
1481                                          &lop->lop_urgent);
1482                 }
1483                 loi_list_maint(cli, loi);
1484                 RETURN(PTR_ERR(request));
1485         }
1486
1487         LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1488         aa = (struct osc_brw_async_args *)&request->rq_async_args;
1489         INIT_LIST_HEAD(&aa->aa_oaps);
1490         list_splice(&rpc_list, &aa->aa_oaps);
1491         INIT_LIST_HEAD(&rpc_list);
1492
1493 #ifdef __KERNEL__
1494         if (cmd == OBD_BRW_READ) {
1495                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1496                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_brw_in_flight);
1497         } else {
1498                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1499                 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1500                                  cli->cl_brw_in_flight);
1501         }
1502 #endif
1503
1504         spin_lock(&cli->cl_loi_list_lock);
1505
1506         cli->cl_brw_in_flight++;
1507         /* queued sync pages can be torn down while the pages
1508          * were between the pending list and the rpc */
1509         list_for_each(pos, &aa->aa_oaps) {
1510                 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1511                 if (oap->oap_interrupted) {
1512                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1513                                oap, request);
1514                         ptlrpc_mark_interrupted(request);
1515                         break;
1516                 }
1517         }
1518
1519         CDEBUG(D_INODE, "req %p: %d pages, aa %p.  now %d in flight\n", request,
1520                page_count, aa, cli->cl_brw_in_flight);
1521
1522         oap->oap_request = ptlrpc_request_addref(request);
1523         request->rq_interpret_reply = brw_interpret_oap;
1524         ptlrpcd_add_req(request);
1525         RETURN(1);
1526 }
1527
1528 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1529                          int cmd)
1530 {
1531         int optimal;
1532         ENTRY;
1533
1534         if (lop->lop_num_pending == 0)
1535                 RETURN(0);
1536
1537         /* stream rpcs in queue order as long as as there is an urgent page
1538          * queued.  this is our cheap solution for good batching in the case
1539          * where writepage marks some random page in the middle of the file as
1540          * urgent because of, say, memory pressure */
1541         if (!list_empty(&lop->lop_urgent))
1542                 RETURN(1);
1543
1544         /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1545         optimal = cli->cl_max_pages_per_rpc;
1546         if (cmd == OBD_BRW_WRITE) {
1547                 /* trigger a write rpc stream as long as there are dirtiers
1548                  * waiting for space.  as they're waiting, they're not going to
1549                  * create more pages to coallesce with what's waiting.. */
1550                 if (!list_empty(&cli->cl_cache_waiters))
1551                         RETURN(1);
1552
1553                 /* *2 to avoid triggering rpcs that would want to include pages
1554                  * that are being queued but which can't be made ready until
1555                  * the queuer finishes with the page. this is a wart for
1556                  * llite::commit_write() */
1557                 optimal += 16;
1558         }
1559         if (lop->lop_num_pending >= optimal)
1560                 RETURN(1);
1561
1562         RETURN(0);
1563 }
1564
1565 static void on_list(struct list_head *item, struct list_head *list,
1566                     int should_be_on)
1567 {
1568         if (list_empty(item) && should_be_on)
1569                 list_add_tail(item, list);
1570         else if (!list_empty(item) && !should_be_on)
1571                 list_del_init(item);
1572 }
1573
1574 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1575  * can find pages to build into rpcs quickly */
1576 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1577 {
1578         on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1579                 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1580                 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1581
1582         on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1583                 loi->loi_write_lop.lop_num_pending);
1584 }
1585
1586 #define LOI_DEBUG(LOI, STR, args...)                                     \
1587         CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR,           \
1588                !list_empty(&(LOI)->loi_cli_item),                        \
1589                (LOI)->loi_write_lop.lop_num_pending,                     \
1590                !list_empty(&(LOI)->loi_write_lop.lop_urgent),            \
1591                (LOI)->loi_read_lop.lop_num_pending,                      \
1592                !list_empty(&(LOI)->loi_read_lop.lop_urgent),             \
1593                args)                                                     \
1594
1595 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1596 {
1597         ENTRY;
1598         /* first return all objects which we already know to have
1599          * pages ready to be stuffed into rpcs */
1600         if (!list_empty(&cli->cl_loi_ready_list))
1601                 RETURN(list_entry(cli->cl_loi_ready_list.next,
1602                                   struct lov_oinfo, loi_cli_item));
1603
1604         /* then if we have cache waiters, return all objects with queued
1605          * writes.  This is especially important when many small files
1606          * have filled up the cache and not been fired into rpcs because
1607          * they don't pass the nr_pending/object threshhold */
1608         if (!list_empty(&cli->cl_cache_waiters) &&
1609             !list_empty(&cli->cl_loi_write_list))
1610                 RETURN(list_entry(cli->cl_loi_write_list.next,
1611                                   struct lov_oinfo, loi_write_item));
1612         RETURN(NULL);
1613 }
1614
1615 /* called with the loi list lock held */
1616 static void osc_check_rpcs(struct client_obd *cli)
1617 {
1618         struct lov_oinfo *loi;
1619         int rc = 0, race_counter = 0;
1620         ENTRY;
1621
1622         while ((loi = osc_next_loi(cli)) != NULL) {
1623                 LOI_DEBUG(loi, "%d in flight\n", cli->cl_brw_in_flight);
1624
1625                 if (cli->cl_brw_in_flight >= cli->cl_max_rpcs_in_flight)
1626                         break;
1627
1628                 /* attempt some read/write balancing by alternating between
1629                  * reads and writes in an object.  The makes_rpc checks here
1630                  * would be redundant if we were getting read/write work items
1631                  * instead of objects.  we don't want send_oap_rpc to drain a
1632                  * partial read pending queue when we're given this object to
1633                  * do io on writes while there are cache waiters */
1634                 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1635                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1636                                               &loi->loi_write_lop);
1637                         if (rc < 0)
1638                                 break;
1639                         if (rc > 0)
1640                                 race_counter = 0;
1641                         else
1642                                 race_counter++;
1643                 }
1644                 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1645                         rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1646                                               &loi->loi_read_lop);
1647                         if (rc < 0)
1648                                 break;
1649                         if (rc > 0)
1650                                 race_counter = 0;
1651                         else
1652                                 race_counter++;
1653                 }
1654
1655                 /* attempt some inter-object balancing by issueing rpcs
1656                  * for each object in turn */
1657                 if (!list_empty(&loi->loi_cli_item))
1658                         list_del_init(&loi->loi_cli_item);
1659                 if (!list_empty(&loi->loi_write_item))
1660                         list_del_init(&loi->loi_write_item);
1661
1662                 loi_list_maint(cli, loi);
1663
1664                 /* send_oap_rpc fails with 0 when make_ready tells it to
1665                  * back off.  llite's make_ready does this when it tries
1666                  * to lock a page queued for write that is already locked.
1667                  * we want to try sending rpcs from many objects, but we
1668                  * don't want to spin failing with 0.  */
1669                 if (race_counter == 10)
1670                         break;
1671         }
1672         EXIT;
1673 }
1674
1675 /* we're trying to queue a page in the osc so we're subject to the
1676  * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1677  * If the osc's queued pages are already at that limit, then we want to sleep
1678  * until there is space in the osc's queue for us.  We also may be waiting for
1679  * write credits from the OST if there are RPCs in flight that may return some
1680  * before we fall back to sync writes.
1681  *
1682  * We need this know our allocation was granted in the presence of signals */
1683 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1684 {
1685         int rc;
1686         ENTRY;
1687         spin_lock(&cli->cl_loi_list_lock);
1688         rc = list_empty(&ocw->ocw_entry) || cli->cl_brw_in_flight == 0;
1689         spin_unlock(&cli->cl_loi_list_lock);
1690         RETURN(rc);
1691 };
1692
1693 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1694  * grant or cache space. */
1695 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1696                            struct osc_async_page *oap)
1697 {
1698         struct osc_cache_waiter ocw;
1699         struct l_wait_info lwi = { 0 };
1700
1701         CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1702                cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1703                cli->cl_avail_grant);
1704
1705         if (cli->cl_dirty_max < PAGE_SIZE)
1706                 return(-EDQUOT);
1707
1708
1709         /* Hopefully normal case - cache space and write credits available */
1710         if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1711             cli->cl_avail_grant >= PAGE_SIZE) {
1712                 /* account for ourselves */
1713                 osc_consume_write_grant(cli, oap);
1714                 return(0);
1715         }
1716
1717         /* Make sure that there are write rpcs in flight to wait for.  This
1718          * is a little silly as this object may not have any pending but
1719          * other objects sure might. */
1720         if (cli->cl_brw_in_flight) {
1721                 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1722                 init_waitqueue_head(&ocw.ocw_waitq);
1723                 ocw.ocw_oap = oap;
1724                 ocw.ocw_rc = 0;
1725
1726                 loi_list_maint(cli, loi);
1727                 osc_check_rpcs(cli);
1728                 spin_unlock(&cli->cl_loi_list_lock);
1729
1730                 CDEBUG(0, "sleeping for cache space\n");
1731                 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1732
1733                 spin_lock(&cli->cl_loi_list_lock);
1734                 if (!list_empty(&ocw.ocw_entry)) {
1735                         list_del(&ocw.ocw_entry);
1736                         RETURN(-EINTR);
1737                 }
1738                 RETURN(ocw.ocw_rc);
1739         }
1740
1741         RETURN(-EDQUOT);
1742 }
1743
1744 /* the companion to enter_cache, called when an oap is no longer part of the
1745  * dirty accounting.. so writeback completes or truncate happens before writing
1746  * starts.  must be called with the loi lock held. */
1747 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1748                            int sent)
1749 {
1750         ENTRY;
1751
1752         if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1753                 EXIT;
1754                 return;
1755         }
1756
1757         oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1758         cli->cl_dirty -= PAGE_SIZE;
1759         if (!sent) {
1760                 cli->cl_lost_grant += PAGE_SIZE;
1761                 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1762                        cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1763         }
1764
1765         EXIT;
1766 }
1767
1768 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1769                         struct lov_oinfo *loi, struct page *page,
1770                         obd_off offset, struct obd_async_page_ops *ops,
1771                         void *data, void **res)
1772 {
1773         struct osc_async_page *oap;
1774         ENTRY;
1775
1776         OBD_ALLOC(oap, sizeof(*oap));
1777         if (oap == NULL)
1778                 return -ENOMEM;
1779
1780         oap->oap_magic = OAP_MAGIC;
1781         oap->oap_cli = &exp->exp_obd->u.cli;
1782         oap->oap_loi = loi;
1783
1784         oap->oap_caller_ops = ops;
1785         oap->oap_caller_data = data;
1786
1787         oap->oap_page = page;
1788         oap->oap_obj_off = offset;
1789
1790         INIT_LIST_HEAD(&oap->oap_pending_item);
1791         INIT_LIST_HEAD(&oap->oap_urgent_item);
1792         INIT_LIST_HEAD(&oap->oap_rpc_item);
1793
1794         oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1795
1796         CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1797         *res = oap;
1798         RETURN(0);
1799 }
1800
1801 struct osc_async_page *oap_from_cookie(void *cookie)
1802 {
1803         struct osc_async_page *oap = cookie;
1804         if (oap->oap_magic != OAP_MAGIC)
1805                 return ERR_PTR(-EINVAL);
1806         return oap;
1807 };
1808
1809 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1810                               struct lov_oinfo *loi, void *cookie,
1811                               int cmd, obd_off off, int count,
1812                               obd_flag brw_flags, enum async_flags async_flags)
1813 {
1814         struct client_obd *cli = &exp->exp_obd->u.cli;
1815         struct osc_async_page *oap;
1816         struct loi_oap_pages *lop;
1817         int rc;
1818         ENTRY;
1819
1820         oap = oap_from_cookie(cookie);
1821         if (IS_ERR(oap))
1822                 RETURN(PTR_ERR(oap));
1823
1824         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1825                 RETURN(-EIO);
1826
1827         if (!list_empty(&oap->oap_pending_item) ||
1828             !list_empty(&oap->oap_urgent_item) ||
1829             !list_empty(&oap->oap_rpc_item))
1830                 RETURN(-EBUSY);
1831
1832         if (loi == NULL)
1833                 loi = &lsm->lsm_oinfo[0];
1834
1835         spin_lock(&cli->cl_loi_list_lock);
1836
1837         oap->oap_cmd = cmd;
1838         oap->oap_async_flags = async_flags;
1839         oap->oap_page_off = off;
1840         oap->oap_count = count;
1841         oap->oap_brw_flags = brw_flags;
1842
1843         if (cmd == OBD_BRW_WRITE) {
1844                 rc = osc_enter_cache(cli, loi, oap);
1845                 if (rc) {
1846                         spin_unlock(&cli->cl_loi_list_lock);
1847                         RETURN(rc);
1848                 }
1849                 lop = &loi->loi_write_lop;
1850         } else {
1851                 lop = &loi->loi_read_lop;
1852         }
1853
1854         if (oap->oap_async_flags & ASYNC_URGENT)
1855                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1856         list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1857         lop_update_pending(cli, lop, cmd, 1);
1858
1859         loi_list_maint(cli, loi);
1860
1861         LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
1862                   cmd);
1863
1864         osc_check_rpcs(cli);
1865         spin_unlock(&cli->cl_loi_list_lock);
1866
1867         RETURN(0);
1868 }
1869
1870 /* aka (~was & now & flag), but this is more clear :) */
1871 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
1872
1873 static int osc_set_async_flags(struct obd_export *exp,
1874                                struct lov_stripe_md *lsm,
1875                                struct lov_oinfo *loi, void *cookie,
1876                                obd_flag async_flags)
1877 {
1878         struct client_obd *cli = &exp->exp_obd->u.cli;
1879         struct loi_oap_pages *lop;
1880         struct osc_async_page *oap;
1881         int rc = 0;
1882         ENTRY;
1883
1884         oap = oap_from_cookie(cookie);
1885         if (IS_ERR(oap))
1886                 RETURN(PTR_ERR(oap));
1887
1888         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1889                 RETURN(-EIO);
1890
1891         if (loi == NULL)
1892                 loi = &lsm->lsm_oinfo[0];
1893
1894         if (oap->oap_cmd == OBD_BRW_WRITE) {
1895                 lop = &loi->loi_write_lop;
1896         } else {
1897                 lop = &loi->loi_read_lop;
1898         }
1899
1900         spin_lock(&cli->cl_loi_list_lock);
1901
1902         if (list_empty(&oap->oap_pending_item))
1903                 GOTO(out, rc = -EINVAL);
1904
1905         if ((oap->oap_async_flags & async_flags) == async_flags)
1906                 GOTO(out, rc = 0);
1907
1908         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
1909                 oap->oap_async_flags |= ASYNC_READY;
1910
1911         if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
1912                 if (list_empty(&oap->oap_rpc_item)) {
1913                         list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1914                         loi_list_maint(cli, loi);
1915                 }
1916         }
1917
1918         LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
1919                         oap->oap_async_flags);
1920 out:
1921         osc_check_rpcs(cli);
1922         spin_unlock(&cli->cl_loi_list_lock);
1923         RETURN(rc);
1924 }
1925
1926 static int osc_queue_sync_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1927                              struct lov_oinfo *loi,
1928                              struct obd_sync_io_container *osic, void *cookie,
1929                              int cmd, obd_off off, int count,
1930                              obd_flag brw_flags)
1931 {
1932         struct client_obd *cli = &exp->exp_obd->u.cli;
1933         struct osc_async_page *oap;
1934         struct loi_oap_pages *lop;
1935         ENTRY;
1936
1937         oap = oap_from_cookie(cookie);
1938         if (IS_ERR(oap))
1939                 RETURN(PTR_ERR(oap));
1940
1941         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1942                 RETURN(-EIO);
1943
1944         if (!list_empty(&oap->oap_pending_item) ||
1945             !list_empty(&oap->oap_urgent_item) ||
1946             !list_empty(&oap->oap_rpc_item))
1947                 RETURN(-EBUSY);
1948
1949         if (loi == NULL)
1950                 loi = &lsm->lsm_oinfo[0];
1951
1952         spin_lock(&cli->cl_loi_list_lock);
1953
1954         oap->oap_cmd = cmd;
1955         oap->oap_page_off = off;
1956         oap->oap_count = count;
1957         oap->oap_brw_flags = brw_flags;
1958
1959         if (cmd == OBD_BRW_WRITE)
1960                 lop = &loi->loi_write_lop;
1961         else
1962                 lop = &loi->loi_read_lop;
1963
1964         list_add_tail(&oap->oap_pending_item, &lop->lop_pending_sync);
1965         oap->oap_osic = osic;
1966         osic_add_one(osic, &oap->oap_occ);
1967
1968         LOI_DEBUG(loi, "oap %p page %p on sync pending\n", oap, oap->oap_page);
1969
1970         spin_unlock(&cli->cl_loi_list_lock);
1971
1972         RETURN(0);
1973 }
1974
1975 static void osc_sync_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
1976                                 struct loi_oap_pages *lop, int cmd)
1977 {
1978         struct list_head *pos, *tmp;
1979         struct osc_async_page *oap;
1980
1981         list_for_each_safe(pos, tmp, &lop->lop_pending_sync) {
1982                 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1983                 list_del(&oap->oap_pending_item);
1984                 oap->oap_async_flags |= ASYNC_READY | ASYNC_URGENT |
1985                                         ASYNC_COUNT_STABLE;
1986                 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1987                 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1988                 lop_update_pending(cli, lop, cmd, 1);
1989         }
1990         loi_list_maint(cli, loi);
1991 }
1992
1993 static int osc_trigger_sync_io(struct obd_export *exp,
1994                                struct lov_stripe_md *lsm,
1995                                struct lov_oinfo *loi,
1996                                struct obd_sync_io_container *osic)
1997 {
1998         struct client_obd *cli = &exp->exp_obd->u.cli;
1999         ENTRY;
2000
2001         if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2002                 RETURN(-EIO);
2003
2004         if (loi == NULL)
2005                 loi = &lsm->lsm_oinfo[0];
2006
2007         spin_lock(&cli->cl_loi_list_lock);
2008
2009         osc_sync_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2010         osc_sync_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2011
2012         osc_check_rpcs(cli);
2013         spin_unlock(&cli->cl_loi_list_lock);
2014
2015         RETURN(0);
2016 }
2017
2018 static int osc_teardown_async_page(struct obd_export *exp,
2019                                    struct lov_stripe_md *lsm,
2020                                    struct lov_oinfo *loi, void *cookie)
2021 {
2022         struct client_obd *cli = &exp->exp_obd->u.cli;
2023         struct loi_oap_pages *lop;
2024         struct osc_async_page *oap;
2025         int rc = 0;
2026         ENTRY;
2027
2028         oap = oap_from_cookie(cookie);
2029         if (IS_ERR(oap))
2030                 RETURN(PTR_ERR(oap));
2031
2032         if (loi == NULL)
2033                 loi = &lsm->lsm_oinfo[0];
2034
2035         if (oap->oap_cmd == OBD_BRW_WRITE) {
2036                 lop = &loi->loi_write_lop;
2037         } else {
2038                 lop = &loi->loi_read_lop;
2039         }
2040
2041         spin_lock(&cli->cl_loi_list_lock);
2042
2043         if (!list_empty(&oap->oap_rpc_item))
2044                 GOTO(out, rc = -EBUSY);
2045
2046         osc_exit_cache(cli, oap, 0);
2047         osc_wake_cache_waiters(cli);
2048
2049         if (!list_empty(&oap->oap_urgent_item)) {
2050                 list_del_init(&oap->oap_urgent_item);
2051                 oap->oap_async_flags &= ~ASYNC_URGENT;
2052         }
2053         if (!list_empty(&oap->oap_pending_item)) {
2054                 list_del_init(&oap->oap_pending_item);
2055                 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2056         }
2057         loi_list_maint(cli, loi);
2058
2059         LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2060 out:
2061         spin_unlock(&cli->cl_loi_list_lock);
2062         if (rc == 0)
2063                 OBD_FREE(oap, sizeof(*oap));
2064         RETURN(rc);
2065 }
2066
2067 #ifdef __KERNEL__
2068 /* Note: caller will lock/unlock, and set uptodate on the pages */
2069 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2070 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2071                            struct lov_stripe_md *lsm, obd_count page_count,
2072                            struct brw_page *pga)
2073 {
2074         struct ptlrpc_request *request = NULL;
2075         struct ost_body *body;
2076         struct niobuf_remote *nioptr;
2077         struct obd_ioobj *iooptr;
2078         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2079         int swab;
2080         ENTRY;
2081
2082         /* XXX does not handle 'new' brw protocol */
2083
2084         size[1] = sizeof(struct obd_ioobj);
2085         size[2] = page_count * sizeof(*nioptr);
2086
2087         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_READ, 3,
2088                                   size, NULL);
2089         if (!request)
2090                 RETURN(-ENOMEM);
2091
2092         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2093         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2094         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2095                                 sizeof(*nioptr) * page_count);
2096
2097         memcpy(&body->oa, oa, sizeof(body->oa));
2098
2099         obdo_to_ioobj(oa, iooptr);
2100         iooptr->ioo_bufcnt = page_count;
2101
2102         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2103                 LASSERT(PageLocked(pga[mapped].pg));
2104                 LASSERT(mapped == 0 || pga[mapped].off > pga[mapped - 1].off);
2105
2106                 nioptr->offset = pga[mapped].off;
2107                 nioptr->len    = pga[mapped].count;
2108                 nioptr->flags  = pga[mapped].flag;
2109         }
2110
2111         size[1] = page_count * sizeof(*nioptr);
2112         request->rq_replen = lustre_msg_size(2, size);
2113
2114         rc = ptlrpc_queue_wait(request);
2115         if (rc)
2116                 GOTO(out_req, rc);
2117
2118         body = lustre_swab_repbuf(request, 0, sizeof(*body),
2119                                   lustre_swab_ost_body);
2120         if (body == NULL) {
2121                 CERROR("Can't unpack body\n");
2122                 GOTO(out_req, rc = -EPROTO);
2123         }
2124
2125         memcpy(oa, &body->oa, sizeof(*oa));
2126
2127         swab = lustre_msg_swabbed(request->rq_repmsg);
2128         LASSERT_REPSWAB(request, 1);
2129         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2130         if (!nioptr) {
2131                 /* nioptr missing or short */
2132                 GOTO(out_req, rc = -EPROTO);
2133         }
2134
2135         /* actual read */
2136         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2137                 struct page *page = pga[mapped].pg;
2138                 struct buffer_head *bh;
2139                 kdev_t dev;
2140
2141                 if (swab)
2142                         lustre_swab_niobuf_remote (nioptr);
2143
2144                 /* got san device associated */
2145                 LASSERT(exp->exp_obd != NULL);
2146                 dev = exp->exp_obd->u.cli.cl_sandev;
2147
2148                 /* hole */
2149                 if (!nioptr->offset) {
2150                         CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2151                                         page->mapping->host->i_ino,
2152                                         page->index);
2153                         memset(page_address(page), 0, PAGE_SIZE);
2154                         continue;
2155                 }
2156
2157                 if (!page->buffers) {
2158                         create_empty_buffers(page, dev, PAGE_SIZE);
2159                         bh = page->buffers;
2160
2161                         clear_bit(BH_New, &bh->b_state);
2162                         set_bit(BH_Mapped, &bh->b_state);
2163                         bh->b_blocknr = (unsigned long)nioptr->offset;
2164
2165                         clear_bit(BH_Uptodate, &bh->b_state);
2166
2167                         ll_rw_block(READ, 1, &bh);
2168                 } else {
2169                         bh = page->buffers;
2170
2171                         /* if buffer already existed, it must be the
2172                          * one we mapped before, check it */
2173                         LASSERT(!test_bit(BH_New, &bh->b_state));
2174                         LASSERT(test_bit(BH_Mapped, &bh->b_state));
2175                         LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2176
2177                         /* wait it's io completion */
2178                         if (test_bit(BH_Lock, &bh->b_state))
2179                                 wait_on_buffer(bh);
2180
2181                         if (!test_bit(BH_Uptodate, &bh->b_state))
2182                                 ll_rw_block(READ, 1, &bh);
2183                 }
2184
2185
2186                 /* must do syncronous write here */
2187                 wait_on_buffer(bh);
2188                 if (!buffer_uptodate(bh)) {
2189                         /* I/O error */
2190                         rc = -EIO;
2191                         goto out_req;
2192                 }
2193         }
2194
2195 out_req:
2196         ptlrpc_req_finished(request);
2197         RETURN(rc);
2198 }
2199
2200 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2201                             struct lov_stripe_md *lsm, obd_count page_count,
2202                             struct brw_page *pga)
2203 {
2204         struct ptlrpc_request *request = NULL;
2205         struct ost_body *body;
2206         struct niobuf_remote *nioptr;
2207         struct obd_ioobj *iooptr;
2208         int rc, size[3] = {sizeof(*body)}, mapped = 0;
2209         int swab;
2210         ENTRY;
2211
2212         size[1] = sizeof(struct obd_ioobj);
2213         size[2] = page_count * sizeof(*nioptr);
2214
2215         request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_WRITE,
2216                                   3, size, NULL);
2217         if (!request)
2218                 RETURN(-ENOMEM);
2219
2220         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2221         iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2222         nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2223                                 sizeof (*nioptr) * page_count);
2224
2225         memcpy(&body->oa, oa, sizeof(body->oa));
2226
2227         obdo_to_ioobj(oa, iooptr);
2228         iooptr->ioo_bufcnt = page_count;
2229
2230         /* pack request */
2231         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2232                 LASSERT(PageLocked(pga[mapped].pg));
2233                 LASSERT(mapped == 0 || pga[mapped].off > pga[mapped - 1].off);
2234
2235                 nioptr->offset = pga[mapped].off;
2236                 nioptr->len    = pga[mapped].count;
2237                 nioptr->flags  = pga[mapped].flag;
2238         }
2239
2240         size[1] = page_count * sizeof(*nioptr);
2241         request->rq_replen = lustre_msg_size(2, size);
2242
2243         rc = ptlrpc_queue_wait(request);
2244         if (rc)
2245                 GOTO(out_req, rc);
2246
2247         swab = lustre_msg_swabbed (request->rq_repmsg);
2248         LASSERT_REPSWAB (request, 1);
2249         nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2250         if (!nioptr) {
2251                 CERROR("absent/short niobuf array\n");
2252                 GOTO(out_req, rc = -EPROTO);
2253         }
2254
2255         /* actual write */
2256         for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2257                 struct page *page = pga[mapped].pg;
2258                 struct buffer_head *bh;
2259                 kdev_t dev;
2260
2261                 if (swab)
2262                         lustre_swab_niobuf_remote (nioptr);
2263
2264                 /* got san device associated */
2265                 LASSERT(exp->exp_obd != NULL);
2266                 dev = exp->exp_obd->u.cli.cl_sandev;
2267
2268                 if (!page->buffers) {
2269                         create_empty_buffers(page, dev, PAGE_SIZE);
2270                 } else {
2271                         /* checking */
2272                         LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2273                         LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2274                         LASSERT(page->buffers->b_blocknr ==
2275                                 (unsigned long)nioptr->offset);
2276                 }
2277                 bh = page->buffers;
2278
2279                 LASSERT(bh);
2280
2281                 /* if buffer locked, wait it's io completion */
2282                 if (test_bit(BH_Lock, &bh->b_state))
2283                         wait_on_buffer(bh);
2284
2285                 clear_bit(BH_New, &bh->b_state);
2286                 set_bit(BH_Mapped, &bh->b_state);
2287
2288                 /* override the block nr */
2289                 bh->b_blocknr = (unsigned long)nioptr->offset;
2290
2291                 /* we are about to write it, so set it
2292                  * uptodate/dirty
2293                  * page lock should garentee no race condition here */
2294                 set_bit(BH_Uptodate, &bh->b_state);
2295                 set_bit(BH_Dirty, &bh->b_state);
2296
2297                 ll_rw_block(WRITE, 1, &bh);
2298
2299                 /* must do syncronous write here */
2300                 wait_on_buffer(bh);
2301                 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2302                         /* I/O error */
2303                         rc = -EIO;
2304                         goto out_req;
2305                 }
2306         }
2307
2308 out_req:
2309         ptlrpc_req_finished(request);
2310         RETURN(rc);
2311 }
2312
2313 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2314                       struct lov_stripe_md *lsm, obd_count page_count,
2315                       struct brw_page *pga, struct obd_trans_info *oti)
2316 {
2317         ENTRY;
2318
2319         while (page_count) {
2320                 obd_count pages_per_brw;
2321                 int rc;
2322
2323                 if (page_count > OSC_BRW_MAX_IOV)
2324                         pages_per_brw = OSC_BRW_MAX_IOV;
2325                 else
2326                         pages_per_brw = page_count;
2327
2328                 if (cmd & OBD_BRW_WRITE)
2329                         rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2330                 else
2331                         rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2332
2333                 if (rc != 0)
2334                         RETURN(rc);
2335
2336                 page_count -= pages_per_brw;
2337                 pga += pages_per_brw;
2338         }
2339         RETURN(0);
2340 }
2341 #endif
2342 #endif
2343
2344 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2345 {
2346         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2347
2348         LASSERT(lock != NULL);
2349         l_lock(&lock->l_resource->lr_namespace->ns_lock);
2350 #ifdef __KERNEL__
2351         if (lock->l_ast_data && lock->l_ast_data != data) {
2352                 struct inode *new_inode = data;
2353                 struct inode *old_inode = lock->l_ast_data;
2354                 LASSERTF(old_inode->i_state & I_FREEING,
2355                          "Found existing inode %p/%lu/%u state %lu in lock: "
2356                          "setting data to %p/%lu/%u\n", old_inode,
2357                          old_inode->i_ino, old_inode->i_generation,
2358                          old_inode->i_state,
2359                          new_inode, new_inode->i_ino, new_inode->i_generation);
2360         }
2361 #endif
2362         lock->l_ast_data = data;
2363         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
2364         LDLM_LOCK_PUT(lock);
2365 }
2366
2367 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2368                              ldlm_iterator_t replace, void *data)
2369 {
2370         struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2371         struct obd_device *obd = class_exp2obd(exp);
2372
2373         ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2374         return 0;
2375 }
2376
2377 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2378                        struct lustre_handle *parent_lock,
2379                        __u32 type, void *extentp, int extent_len, __u32 mode,
2380                        int *flags, void *callback, void *data,
2381                        struct lustre_handle *lockh)
2382 {
2383         struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2384         struct obd_device *obd = exp->exp_obd;
2385         struct ldlm_extent *extent = extentp;
2386         int rc;
2387         ENTRY;
2388
2389         /* Filesystem lock extents are extended to page boundaries so that
2390          * dealing with the page cache is a little smoother.  */
2391         extent->start -= extent->start & ~PAGE_MASK;
2392         extent->end |= ~PAGE_MASK;
2393
2394         /* Next, search for already existing extent locks that will cover us */
2395         rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id,
2396                              type, extent, sizeof(*extent), mode, lockh);
2397         if (rc == 1) {
2398                 osc_set_data_with_check(lockh, data);
2399                 /* We already have a lock, and it's referenced */
2400                 RETURN(ELDLM_OK);
2401         }
2402
2403         /* If we're trying to read, we also search for an existing PW lock.  The
2404          * VFS and page cache already protect us locally, so lots of readers/
2405          * writers can share a single PW lock.
2406          *
2407          * There are problems with conversion deadlocks, so instead of
2408          * converting a read lock to a write lock, we'll just enqueue a new
2409          * one.
2410          *
2411          * At some point we should cancel the read lock instead of making them
2412          * send us a blocking callback, but there are problems with canceling
2413          * locks out from other users right now, too. */
2414
2415         if (mode == LCK_PR) {
2416                 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2417                                      extent, sizeof(*extent), LCK_PW, lockh);
2418                 if (rc == 1) {
2419                         /* FIXME: This is not incredibly elegant, but it might
2420                          * be more elegant than adding another parameter to
2421                          * lock_match.  I want a second opinion. */
2422                         ldlm_lock_addref(lockh, LCK_PR);
2423                         ldlm_lock_decref(lockh, LCK_PW);
2424                         osc_set_data_with_check(lockh, data);
2425                         RETURN(ELDLM_OK);
2426                 }
2427         }
2428
2429         rc = ldlm_cli_enqueue(exp, NULL, obd->obd_namespace, parent_lock,
2430                               res_id, type, extent, sizeof(*extent), mode,
2431                               flags,ldlm_completion_ast, callback, data, lockh);
2432         RETURN(rc);
2433 }
2434
2435 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2436                      __u32 type, void *extentp, int extent_len, __u32 mode,
2437                      int *flags, void *data, struct lustre_handle *lockh)
2438 {
2439         struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2440         struct obd_device *obd = exp->exp_obd;
2441         struct ldlm_extent *extent = extentp;
2442         int rc;
2443         ENTRY;
2444
2445         OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2446
2447         /* Filesystem lock extents are extended to page boundaries so that
2448          * dealing with the page cache is a little smoother */
2449         extent->start -= extent->start & ~PAGE_MASK;
2450         extent->end |= ~PAGE_MASK;
2451
2452         /* Next, search for already existing extent locks that will cover us */
2453         rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2454                              extent, sizeof(*extent), mode, lockh);
2455         if (rc) {
2456                 osc_set_data_with_check(lockh, data);
2457                 RETURN(rc);
2458         }
2459         /* If we're trying to read, we also search for an existing PW lock.  The
2460          * VFS and page cache already protect us locally, so lots of readers/
2461          * writers can share a single PW lock. */
2462         if (mode == LCK_PR) {
2463                 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2464                                      extent, sizeof(*extent), LCK_PW, lockh);
2465                 if (rc == 1) {
2466                         /* FIXME: This is not incredibly elegant, but it might
2467                          * be more elegant than adding another parameter to
2468                          * lock_match.  I want a second opinion. */
2469                         osc_set_data_with_check(lockh, data);
2470                         ldlm_lock_addref(lockh, LCK_PR);
2471                         ldlm_lock_decref(lockh, LCK_PW);
2472                 }
2473         }
2474         RETURN(rc);
2475 }
2476
2477 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2478                       __u32 mode, struct lustre_handle *lockh)
2479 {
2480         ENTRY;
2481
2482         ldlm_lock_decref(lockh, mode);
2483
2484         RETURN(0);
2485 }
2486
2487 static int osc_cancel_unused(struct obd_export *exp,
2488                              struct lov_stripe_md *lsm, int flags, void *opaque)
2489 {
2490         struct obd_device *obd = class_exp2obd(exp);
2491         struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2492
2493         return ldlm_cli_cancel_unused(obd->obd_namespace, &res_id, flags,
2494                                       opaque);
2495 }
2496
2497 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2498                       unsigned long max_age)
2499 {
2500         struct obd_statfs *msfs;
2501         struct ptlrpc_request *request;
2502         int rc, size = sizeof(*osfs);
2503         ENTRY;
2504
2505         /* We could possibly pass max_age in the request (as an absolute
2506          * timestamp or a "seconds.usec ago") so the target can avoid doing
2507          * extra calls into the filesystem if that isn't necessary (e.g.
2508          * during mount that would help a bit).  Having relative timestamps
2509          * is not so great if request processing is slow, while absolute
2510          * timestamps are not ideal because they need time synchronization. */
2511         request = ptlrpc_prep_req(obd->u.cli.cl_import, OST_STATFS,0,NULL,NULL);
2512         if (!request)
2513                 RETURN(-ENOMEM);
2514
2515         request->rq_replen = lustre_msg_size(1, &size);
2516         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2517
2518         rc = ptlrpc_queue_wait(request);
2519         if (rc)
2520                 GOTO(out, rc);
2521
2522         msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2523                                   lustre_swab_obd_statfs);
2524         if (msfs == NULL) {
2525                 CERROR("Can't unpack obd_statfs\n");
2526                 GOTO(out, rc = -EPROTO);
2527         }
2528
2529         memcpy(osfs, msfs, sizeof(*osfs));
2530
2531         EXIT;
2532  out:
2533         ptlrpc_req_finished(request);
2534         return rc;
2535 }
2536
2537 /* Retrieve object striping information.
2538  *
2539  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2540  * the maximum number of OST indices which will fit in the user buffer.
2541  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2542  */
2543 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2544 {
2545         struct lov_user_md lum;
2546         struct lov_mds_md *lmmk;
2547         int rc, lmm_size;
2548         ENTRY;
2549
2550         if (!lsm)
2551                 RETURN(-ENODATA);
2552
2553         rc = copy_from_user(&lum, lump, sizeof(lum));
2554         if (rc)
2555                 RETURN(-EFAULT);
2556
2557         if (lum.lmm_magic != LOV_USER_MAGIC)
2558                 RETURN(-EINVAL);
2559
2560         if (lum.lmm_stripe_count < 1)
2561                 RETURN(-EOVERFLOW);
2562
2563         lmm_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2564         OBD_ALLOC(lmmk, lmm_size);
2565         if (!lmmk)
2566                 RETURN(-ENOMEM);
2567
2568         lmmk->lmm_stripe_count = 1;
2569         lmmk->lmm_object_id = lsm->lsm_object_id;
2570         lmmk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2571
2572         if (copy_to_user(lump, lmmk, lmm_size))
2573                 rc = -EFAULT;
2574
2575         OBD_FREE(lmmk, lmm_size);
2576
2577         RETURN(rc);
2578 }
2579
2580 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2581                          void *karg, void *uarg)
2582 {
2583         struct obd_device *obd = exp->exp_obd;
2584         struct obd_ioctl_data *data = karg;
2585         int err = 0;
2586         ENTRY;
2587
2588         MOD_INC_USE_COUNT;
2589
2590         switch (cmd) {
2591         case OBD_IOC_LOV_GET_CONFIG: {
2592                 char *buf;
2593                 struct lov_desc *desc;
2594                 struct obd_uuid uuid;
2595
2596                 buf = NULL;
2597                 len = 0;
2598                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2599                         GOTO(out, err = -EINVAL);
2600
2601                 data = (struct obd_ioctl_data *)buf;
2602
2603                 if (sizeof(*desc) > data->ioc_inllen1) {
2604                         OBD_FREE(buf, len);
2605                         GOTO(out, err = -EINVAL);
2606                 }
2607
2608                 if (data->ioc_inllen2 < sizeof(uuid)) {
2609                         OBD_FREE(buf, len);
2610                         GOTO(out, err = -EINVAL);
2611                 }
2612
2613                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2614                 desc->ld_tgt_count = 1;
2615                 desc->ld_active_tgt_count = 1;
2616                 desc->ld_default_stripe_count = 1;
2617                 desc->ld_default_stripe_size = 0;
2618                 desc->ld_default_stripe_offset = 0;
2619                 desc->ld_pattern = 0;
2620                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2621
2622                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2623
2624                 err = copy_to_user((void *)uarg, buf, len);
2625                 if (err)
2626                         err = -EFAULT;
2627                 obd_ioctl_freedata(buf, len);
2628                 GOTO(out, err);
2629         }
2630         case LL_IOC_LOV_SETSTRIPE:
2631                 err = obd_alloc_memmd(exp, karg);
2632                 if (err > 0)
2633                         err = 0;
2634                 GOTO(out, err);
2635         case LL_IOC_LOV_GETSTRIPE:
2636                 err = osc_getstripe(karg, uarg);
2637                 GOTO(out, err);
2638         case OBD_IOC_CLIENT_RECOVER:
2639                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2640                                             data->ioc_inlbuf1);
2641                 if (err > 0)
2642                         err = 0;
2643                 GOTO(out, err);
2644         case IOC_OSC_SET_ACTIVE:
2645                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2646                                                data->ioc_offset);
2647                 GOTO(out, err);
2648         default:
2649                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2650                 GOTO(out, err = -ENOTTY);
2651         }
2652 out:
2653         MOD_DEC_USE_COUNT;
2654         return err;
2655 }
2656
2657 static int osc_get_info(struct obd_export *exp, obd_count keylen,
2658                         void *key, __u32 *vallen, void *val)
2659 {
2660         ENTRY;
2661         if (!vallen || !val)
2662                 RETURN(-EFAULT);
2663
2664         if (keylen > strlen("lock_to_stripe") &&
2665             strcmp(key, "lock_to_stripe") == 0) {
2666                 __u32 *stripe = val;
2667                 *vallen = sizeof(*stripe);
2668                 *stripe = 0;
2669                 RETURN(0);
2670         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2671                 struct ptlrpc_request *req;
2672                 obd_id *reply;
2673                 char *bufs[1] = {key};
2674                 int rc;
2675                 req = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GET_INFO, 1,
2676                                       &keylen, bufs);
2677                 if (req == NULL)
2678                         RETURN(-ENOMEM);
2679
2680                 req->rq_replen = lustre_msg_size(1, vallen);
2681                 rc = ptlrpc_queue_wait(req);
2682                 if (rc)
2683                         GOTO(out, rc);
2684
2685                 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2686                                            lustre_swab_ost_last_id);
2687                 if (reply == NULL) {
2688                         CERROR("Can't unpack OST last ID\n");
2689                         GOTO(out, rc = -EPROTO);
2690                 }
2691                 *((obd_id *)val) = *reply;
2692         out:
2693                 ptlrpc_req_finished(req);
2694                 RETURN(rc);
2695         }
2696         RETURN(-EINVAL);
2697 }
2698
2699 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2700                         void *key, obd_count vallen, void *val)
2701 {
2702         struct ptlrpc_request *req;
2703         struct obd_import *imp = class_exp2cliimp(exp);
2704         struct llog_ctxt *ctxt;
2705         int rc, size = keylen;
2706         char *bufs[1] = {key};
2707         ENTRY;
2708
2709         if (keylen == strlen("next_id") &&
2710             memcmp(key, "next_id", strlen("next_id")) == 0) {
2711                 if (vallen != sizeof(obd_id))
2712                         RETURN(-EINVAL);
2713                 exp->u.eu_osc_data.oed_oscc.oscc_next_id = *((obd_id*)val) + 1;
2714                 CDEBUG(D_INODE, "%s: set oscc_next_id = "LPU64"\n",
2715                        exp->exp_obd->obd_name,
2716                        exp->u.eu_osc_data.oed_oscc.oscc_next_id);
2717
2718                 RETURN(0);
2719         }
2720
2721         if (keylen == strlen("growth_count") &&
2722             memcmp(key, "growth_count", strlen("growth_count")) == 0) {
2723                 if (vallen != sizeof(int))
2724                         RETURN(-EINVAL);
2725                 exp->u.eu_osc_data.oed_oscc.oscc_grow_count = *((int*)val);
2726                 RETURN(0);
2727         }
2728
2729         if (keylen == strlen("unlinked") &&
2730             memcmp(key, "unlinked", keylen) == 0) {
2731                 struct osc_creator *oscc = &exp->u.eu_osc_data.oed_oscc;
2732                 spin_lock(&oscc->oscc_lock);
2733                 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2734                 spin_unlock(&oscc->oscc_lock);
2735                 RETURN(0);
2736         }
2737
2738
2739         if (keylen == strlen("initial_recov") &&
2740             memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
2741                 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2742                 if (vallen != sizeof(int))
2743                         RETURN(-EINVAL);
2744                 imp->imp_initial_recov = *(int *)val;
2745                 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
2746                        exp->exp_obd->obd_name,
2747                        imp->imp_initial_recov);
2748                 RETURN(0);
2749         }
2750
2751         if (keylen < strlen("mds_conn") ||
2752             memcmp(key, "mds_conn", strlen("mds_conn")) != 0)
2753                 RETURN(-EINVAL);
2754
2755
2756         req = ptlrpc_prep_req(imp, OST_SET_INFO, 1, &size, bufs);
2757         if (req == NULL)
2758                 RETURN(-ENOMEM);
2759
2760         req->rq_replen = lustre_msg_size(0, NULL);
2761         rc = ptlrpc_queue_wait(req);
2762         ptlrpc_req_finished(req);
2763
2764         ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_ORIG_CTXT);
2765         if (ctxt) {
2766                 rc = llog_initiator_connect(ctxt);
2767                 if (rc)
2768                         RETURN(rc);
2769         }
2770
2771         imp->imp_server_timeout = 1;
2772         CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
2773         ptlrpc_pinger_add_import(imp);
2774
2775         RETURN(rc);
2776 }
2777
2778
2779 static struct llog_operations osc_size_repl_logops = {
2780         lop_cancel: llog_obd_repl_cancel
2781 };
2782
2783 static struct llog_operations osc_unlink_orig_logops;
2784 static int osc_llog_init(struct obd_device *obd, struct obd_device *tgt,
2785                         int count, struct llog_logid *logid)
2786 {
2787         int rc;
2788         ENTRY;
2789
2790         osc_unlink_orig_logops = llog_lvfs_ops;
2791         osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
2792         osc_unlink_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
2793         osc_unlink_orig_logops.lop_add = llog_obd_origin_add;
2794         osc_unlink_orig_logops.lop_connect = llog_origin_connect;
2795
2796         rc = llog_setup(obd, LLOG_UNLINK_ORIG_CTXT, tgt, count, logid,
2797                         &osc_unlink_orig_logops);
2798         if (rc)
2799                 RETURN(rc);
2800
2801         rc = llog_setup(obd, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
2802                         &osc_size_repl_logops);
2803         RETURN(rc);
2804 }
2805
2806 static int osc_llog_finish(struct obd_device *obd, int count)
2807 {
2808         int rc;
2809         ENTRY;
2810
2811         rc = llog_cleanup(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT));
2812         if (rc)
2813                 RETURN(rc);
2814
2815         rc = llog_cleanup(llog_get_context(obd, LLOG_SIZE_REPL_CTXT));
2816         RETURN(rc);
2817 }
2818
2819
2820 static int osc_connect(struct lustre_handle *exph,
2821                        struct obd_device *obd, struct obd_uuid *cluuid)
2822 {
2823         int rc;
2824         struct obd_export *exp;
2825
2826         rc = client_connect_import(exph, obd, cluuid);
2827
2828         if (obd->u.cli.cl_conn_count == 1) {
2829                 exp = class_conn2export(exph);
2830                 oscc_init(exp);
2831         }
2832
2833         return rc;
2834 }
2835
2836 static int osc_disconnect(struct obd_export *exp, int flags)
2837 {
2838         struct obd_device *obd = class_exp2obd(exp);
2839         struct llog_ctxt *ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
2840         int rc;
2841
2842         if (obd->u.cli.cl_conn_count == 1) {
2843                 /* flush any remaining cancel messages out to the target */
2844                 llog_sync(ctxt, exp);
2845
2846                 /* balance the conn2export for oscc in osc_connect */
2847                 class_export_put(exp);
2848         }
2849
2850         rc = client_disconnect_export(exp, flags);
2851         return rc;
2852 }
2853
2854 static int osc_lock_contains(struct obd_export *exp, struct lov_stripe_md *lsm,
2855                              struct ldlm_lock *lock, obd_off offset)
2856 {
2857         ENTRY;
2858         if (exp == NULL)
2859                 RETURN(-ENODEV);
2860
2861         if (lock->l_policy_data.l_extent.start <= offset &&
2862             lock->l_policy_data.l_extent.end >= offset)
2863                 RETURN(1);
2864         RETURN(0);
2865 }
2866
2867 static int osc_invalidate_import(struct obd_device *obd,
2868                                  struct obd_import *imp)
2869 {
2870         struct client_obd *cli;
2871         LASSERT(imp->imp_obd == obd);
2872         /* this used to try and tear down queued pages, but it was
2873          * not correctly implemented.  We'll have to do it again once
2874          * we call obd_invalidate_import() agian */
2875         /* XXX And we still need to do this */
2876
2877         /* Reset grants, too */
2878         cli = &obd->u.cli;
2879         spin_lock(&cli->cl_loi_list_lock);
2880         cli->cl_avail_grant = 0;
2881         cli->cl_lost_grant = 0;
2882         spin_unlock(&cli->cl_loi_list_lock);
2883
2884         RETURN(0);
2885 }
2886
2887 int osc_setup(struct obd_device *obd, obd_count len, void *buf)
2888 {
2889         int rc;
2890
2891         rc = ptlrpcd_addref();
2892         if (rc)
2893                 return rc;
2894
2895         rc = client_obd_setup(obd, len, buf);
2896         if (rc)
2897                 ptlrpcd_decref();
2898         RETURN(rc);
2899 }
2900
2901 int osc_cleanup(struct obd_device *obd, int flags)
2902 {
2903         int rc;
2904
2905         rc = client_obd_cleanup(obd, flags);
2906         ptlrpcd_decref();
2907         RETURN(rc);
2908 }
2909
2910
2911 struct obd_ops osc_obd_ops = {
2912         o_owner:        THIS_MODULE,
2913         o_attach:       osc_attach,
2914         o_detach:       osc_detach,
2915         o_setup:        osc_setup,
2916         o_cleanup:      osc_cleanup,
2917         o_connect:      osc_connect,
2918         o_disconnect:   osc_disconnect,
2919         o_statfs:       osc_statfs,
2920         o_packmd:       osc_packmd,
2921         o_unpackmd:     osc_unpackmd,
2922         o_create:       osc_create,
2923         o_destroy:      osc_destroy,
2924         o_getattr:      osc_getattr,
2925         o_getattr_async:osc_getattr_async,
2926         o_setattr:      osc_setattr,
2927         o_brw:          osc_brw,
2928         o_brw_async:    osc_brw_async,
2929         .o_prep_async_page =            osc_prep_async_page,
2930         .o_queue_async_io =             osc_queue_async_io,
2931         .o_set_async_flags =            osc_set_async_flags,
2932         .o_queue_sync_io =              osc_queue_sync_io,
2933         .o_trigger_sync_io =            osc_trigger_sync_io,
2934         .o_teardown_async_page =        osc_teardown_async_page,
2935         o_punch:        osc_punch,
2936         o_sync:         osc_sync,
2937         o_enqueue:      osc_enqueue,
2938         o_match:        osc_match,
2939         o_change_cbdata:osc_change_cbdata,
2940         o_cancel:       osc_cancel,
2941         o_cancel_unused:osc_cancel_unused,
2942         o_iocontrol:    osc_iocontrol,
2943         o_get_info:     osc_get_info,
2944         o_set_info:     osc_set_info,
2945         o_lock_contains:osc_lock_contains,
2946         o_invalidate_import: osc_invalidate_import,
2947         o_llog_init:    osc_llog_init,
2948         o_llog_finish:  osc_llog_finish,
2949 };
2950
2951 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2952 struct obd_ops sanosc_obd_ops = {
2953         o_owner:        THIS_MODULE,
2954         o_attach:       osc_attach,
2955         o_detach:       osc_detach,
2956         o_cleanup:      client_obd_cleanup,
2957         o_connect:      osc_connect,
2958         o_disconnect:   client_disconnect_export,
2959         o_statfs:       osc_statfs,
2960         o_packmd:       osc_packmd,
2961         o_unpackmd:     osc_unpackmd,
2962         o_create:       osc_real_create,
2963         o_destroy:      osc_destroy,
2964         o_getattr:      osc_getattr,
2965         o_getattr_async:osc_getattr_async,
2966         o_setattr:      osc_setattr,
2967         o_setup:        client_sanobd_setup,
2968         o_brw:          sanosc_brw,
2969         o_punch:        osc_punch,
2970         o_sync:         osc_sync,
2971         o_enqueue:      osc_enqueue,
2972         o_match:        osc_match,
2973         o_change_cbdata:osc_change_cbdata,
2974         o_cancel:       osc_cancel,
2975         o_cancel_unused:osc_cancel_unused,
2976         o_iocontrol:    osc_iocontrol,
2977         o_lock_contains:osc_lock_contains,
2978         o_invalidate_import: osc_invalidate_import,
2979         o_llog_init:    osc_llog_init,
2980         o_llog_finish:  osc_llog_finish,
2981 };
2982 #endif
2983
2984 int __init osc_init(void)
2985 {
2986         struct lprocfs_static_vars lvars, sanlvars;
2987         int rc;
2988         ENTRY;
2989
2990         lprocfs_init_vars(osc, &lvars);
2991         lprocfs_init_vars(osc, &sanlvars);
2992
2993         rc = class_register_type(&osc_obd_ops, lvars.module_vars,
2994                                  LUSTRE_OSC_NAME);
2995         if (rc)
2996                 RETURN(rc);
2997
2998 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2999         rc = class_register_type(&sanosc_obd_ops, sanlvars.module_vars,
3000                                  LUSTRE_SANOSC_NAME);
3001         if (rc)
3002                 class_unregister_type(LUSTRE_OSC_NAME);
3003 #endif
3004
3005         RETURN(rc);
3006 }
3007
3008 static void /*__exit*/ osc_exit(void)
3009 {
3010 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3011         class_unregister_type(LUSTRE_SANOSC_NAME);
3012 #endif
3013         class_unregister_type(LUSTRE_OSC_NAME);
3014 }
3015
3016 #ifdef __KERNEL__
3017 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3018 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3019 MODULE_LICENSE("GPL");
3020
3021 module_init(osc_init);
3022 module_exit(osc_exit);
3023 #endif