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