lustre/obdecho/echo_client.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  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #define DEBUG_SUBSYSTEM S_ECHO
  26 #ifdef __KERNEL__
  27 #include <libcfs/libcfs.h>
  28 #else
  29 #include <liblustre.h>
  30 #endif
  31
  32 #include <obd.h>
  33 #include <obd_support.h>
  34 #include <obd_class.h>
  35 #include <obd_echo.h>
  36 #include <lustre_debug.h>
  37 #include <lprocfs_status.h>
  38
  39 static obd_id last_object_id;
  40
  41 #if 0
  42 static void
  43 echo_printk_object (char *msg, struct ec_object *eco)
  44 {
  45         struct lov_stripe_md *lsm = eco->eco_lsm;
  46         int                   i;
  47
  48         printk (KERN_INFO "Lustre: %s: object %p: "LPX64", refs %d%s: "LPX64
  49                 "=%u!%u\n", msg, eco, eco->eco_id, eco->eco_refcount,
  50                 eco->eco_deleted ? "(deleted) " : "",
  51                 lsm->lsm_object_id, lsm->lsm_stripe_size,
  52                 lsm->lsm_stripe_count);
  53
  54         for (i = 0; i < lsm->lsm_stripe_count; i++)
  55                 printk (KERN_INFO "Lustre:   @%2u:"LPX64"\n",
  56                         lsm->lsm_oinfo[i].loi_ost_idx,
  57                         lsm->lsm_oinfo[i].loi_id);
  58 }
  59 #endif
  60
  61 static struct ec_object *
  62 echo_find_object_locked (struct obd_device *obd, obd_id id)
  63 {
  64         struct echo_client_obd *ec = &obd->u.echo_client;
  65         struct ec_object       *eco = NULL;
  66         struct list_head       *el;
  67
  68         list_for_each (el, &ec->ec_objects) {
  69                 eco = list_entry (el, struct ec_object, eco_obj_chain);
  70
  71                 if (eco->eco_id == id)
  72                         return (eco);
  73         }
  74         return (NULL);
  75 }
  76
  77 static int
  78 echo_copyout_lsm (struct lov_stripe_md *lsm, void *_ulsm, int ulsm_nob)
  79 {
  80         struct lov_stripe_md *ulsm = _ulsm;
  81         int nob, i;
  82
  83         nob = offsetof (struct lov_stripe_md, lsm_oinfo[lsm->lsm_stripe_count]);
  84         if (nob > ulsm_nob)
  85                 return (-EINVAL);
  86
  87         if (copy_to_user (ulsm, lsm, sizeof(ulsm)))
  88                 return (-EFAULT);
  89
  90         for (i = 0; i < lsm->lsm_stripe_count; i++) {
  91                 if (copy_to_user (ulsm->lsm_oinfo[i], lsm->lsm_oinfo[i],
  92                                   sizeof(lsm->lsm_oinfo[0])))
  93                         return (-EFAULT);
  94         }
  95         return 0;
  96 }
  97
  98 static int
  99 echo_copyin_lsm (struct obd_device *obd, struct lov_stripe_md *lsm,
 100                  void *ulsm, int ulsm_nob)
 101 {
 102         struct echo_client_obd *ec = &obd->u.echo_client;
 103         int                     i;
 104
 105         if (ulsm_nob < sizeof (*lsm))
 106                 return (-EINVAL);
 107
 108         if (copy_from_user (lsm, ulsm, sizeof (*lsm)))
 109                 return (-EFAULT);
 110
 111         if (lsm->lsm_stripe_count > ec->ec_nstripes ||
 112             lsm->lsm_magic != LOV_MAGIC ||
 113             (lsm->lsm_stripe_size & (~CFS_PAGE_MASK)) != 0 ||
 114             ((__u64)lsm->lsm_stripe_size * lsm->lsm_stripe_count > ~0UL))
 115                 return (-EINVAL);
 116
 117
 118         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 119                 if (copy_from_user(lsm->lsm_oinfo[i],
 120                                    ((struct lov_stripe_md *)ulsm)->lsm_oinfo[i],
 121                                    sizeof(lsm->lsm_oinfo[0])))
 122                         return (-EFAULT);
 123         }
 124         return (0);
 125 }
 126
 127 static struct ec_object *
 128 echo_allocate_object (struct obd_device *obd)
 129 {
 130         struct echo_client_obd *ec = &obd->u.echo_client;
 131         struct ec_object       *eco;
 132         int rc;
 133
 134         OBD_ALLOC(eco, sizeof (*eco));
 135         if (eco == NULL)
 136                 return NULL;
 137
 138         rc = obd_alloc_memmd(ec->ec_exp, &eco->eco_lsm);
 139         if (rc < 0) {
 140                 OBD_FREE(eco, sizeof (*eco));
 141                 return NULL;
 142         }
 143
 144         eco->eco_device = obd;
 145         eco->eco_deleted = 0;
 146         eco->eco_refcount = 0;
 147         eco->eco_lsm->lsm_magic = LOV_MAGIC;
 148         /* leave stripe count 0 by default */
 149
 150         return (eco);
 151 }
 152
 153 static void
 154 echo_free_object (struct ec_object *eco)
 155 {
 156         struct obd_device      *obd = eco->eco_device;
 157         struct echo_client_obd *ec = &obd->u.echo_client;
 158
 159         LASSERT (eco->eco_refcount == 0);
 160         if (!eco->eco_lsm)
 161                 CERROR("No object %s\n", obd->obd_name);
 162         else
 163                 obd_free_memmd(ec->ec_exp, &eco->eco_lsm);
 164         OBD_FREE (eco, sizeof (*eco));
 165 }
 166
 167 static int echo_create_object(struct obd_device *obd, int on_target,
 168                               struct obdo *oa, void *ulsm, int ulsm_nob,
 169                               struct obd_trans_info *oti)
 170 {
 171         struct echo_client_obd *ec = &obd->u.echo_client;
 172         struct ec_object       *eco2;
 173         struct ec_object       *eco;
 174         struct lov_stripe_md   *lsm;
 175         int                     rc;
 176         int                     i, idx;
 177
 178         if ((oa->o_valid & OBD_MD_FLID) == 0 && /* no obj id */
 179             (on_target ||                       /* set_stripe */
 180              ec->ec_nstripes != 0)) {           /* LOV */
 181                 CERROR ("No valid oid\n");
 182                 return (-EINVAL);
 183         }
 184
 185         if (ulsm != NULL) {
 186                 eco = echo_allocate_object (obd);
 187                 if (eco == NULL)
 188                         return (-ENOMEM);
 189
 190                 lsm = eco->eco_lsm;
 191
 192                 rc = echo_copyin_lsm (obd, lsm, ulsm, ulsm_nob);
 193                 if (rc != 0)
 194                         goto failed;
 195
 196                 /* setup object ID here for !on_target and LOV hint */
 197                 if ((oa->o_valid & OBD_MD_FLID) != 0)
 198                         eco->eco_id = lsm->lsm_object_id = oa->o_id;
 199
 200                 if (lsm->lsm_stripe_count == 0)
 201                         lsm->lsm_stripe_count = ec->ec_nstripes;
 202
 203                 if (lsm->lsm_stripe_size == 0)
 204                         lsm->lsm_stripe_size = CFS_PAGE_SIZE;
 205
 206                 idx = ll_rand();
 207
 208                 /* setup stripes: indices + default ids if required */
 209                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 210                         if (lsm->lsm_oinfo[i]->loi_id == 0)
 211                                 lsm->lsm_oinfo[i]->loi_id = lsm->lsm_object_id;
 212
 213                         lsm->lsm_oinfo[i]->loi_ost_idx =
 214                                 (idx + i) % ec->ec_nstripes;
 215                 }
 216         } else {
 217                 OBD_ALLOC(eco, sizeof(*eco));
 218                 if (!eco)
 219                         return (-ENOMEM);
 220                 eco->eco_device = obd;
 221                 lsm = NULL;
 222         }
 223
 224         if (oa->o_id == 0)
 225                 oa->o_id = ++last_object_id;
 226
 227         if (on_target) {
 228                 oa->o_gr = FILTER_GROUP_ECHO;
 229                 oa->o_valid |= OBD_MD_FLGROUP;
 230
 231                 rc = obd_create(ec->ec_exp, oa, &lsm, oti);
 232                 if (rc != 0)
 233                         goto failed;
 234
 235                 /* See what object ID we were given */
 236                 eco->eco_id = oa->o_id = lsm->lsm_object_id;
 237                 oa->o_valid |= OBD_MD_FLID;
 238
 239                 LASSERT(eco->eco_lsm == NULL || eco->eco_lsm == lsm);
 240                 eco->eco_lsm = lsm;
 241         }
 242
 243         spin_lock (&ec->ec_lock);
 244
 245         eco2 = echo_find_object_locked (obd, oa->o_id);
 246         if (eco2 != NULL) {                     /* conflict */
 247                 spin_unlock (&ec->ec_lock);
 248
 249                 CERROR ("Can't create object id "LPX64": id already exists%s\n",
 250                         oa->o_id, on_target ? " (undoing create)" : "");
 251
 252                 if (on_target)
 253                         obd_destroy(ec->ec_exp, oa, lsm, oti, NULL);
 254
 255                 rc = -EEXIST;
 256                 goto failed;
 257         }
 258
 259         list_add (&eco->eco_obj_chain, &ec->ec_objects);
 260         spin_unlock (&ec->ec_lock);
 261         CDEBUG (D_INFO,
 262                 "created %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 263                 eco, eco->eco_id,
 264                 eco->eco_lsm->lsm_stripe_size,
 265                 eco->eco_lsm->lsm_stripe_count,
 266                 eco->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 267                 eco->eco_refcount, eco->eco_deleted);
 268         return (0);
 269
 270  failed:
 271         echo_free_object (eco);
 272         if (rc)
 273                 CERROR("%s: err %d on create\n", obd->obd_name, rc);
 274         return (rc);
 275 }
 276
 277 static int
 278 echo_get_object (struct ec_object **ecop, struct obd_device *obd,
 279                  struct obdo *oa)
 280 {
 281         struct echo_client_obd *ec = &obd->u.echo_client;
 282         struct ec_object       *eco;
 283         struct ec_object       *eco2;
 284         int                     rc;
 285
 286         if ((oa->o_valid & OBD_MD_FLID) == 0 ||
 287             oa->o_id == 0)                      /* disallow use of object id 0 */
 288         {
 289                 CERROR ("No valid oid\n");
 290                 return (-EINVAL);
 291         }
 292
 293         spin_lock (&ec->ec_lock);
 294         eco = echo_find_object_locked (obd, oa->o_id);
 295         if (eco != NULL) {
 296                 if (eco->eco_deleted) {           /* being deleted */
 297                         spin_unlock(&ec->ec_lock);/* (see comment in cleanup) */
 298                         return (-EAGAIN);
 299                 }
 300
 301                 eco->eco_refcount++;
 302                 spin_unlock (&ec->ec_lock);
 303                 *ecop = eco;
 304                 CDEBUG (D_INFO,
 305                         "found %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 306                         eco, eco->eco_id,
 307                         eco->eco_lsm->lsm_stripe_size,
 308                         eco->eco_lsm->lsm_stripe_count,
 309                         eco->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 310                         eco->eco_refcount, eco->eco_deleted);
 311                 return (0);
 312         }
 313         spin_unlock (&ec->ec_lock);
 314
 315         if (ec->ec_nstripes != 0)               /* striping required */
 316                 return (-ENOENT);
 317
 318         eco = echo_allocate_object (obd);
 319         if (eco == NULL)
 320                 return (-ENOMEM);
 321
 322         eco->eco_id = eco->eco_lsm->lsm_object_id = oa->o_id;
 323
 324         spin_lock (&ec->ec_lock);
 325
 326         eco2 = echo_find_object_locked (obd, oa->o_id);
 327         if (eco2 == NULL) {                     /* didn't race */
 328                 list_add (&eco->eco_obj_chain, &ec->ec_objects);
 329                 spin_unlock (&ec->ec_lock);
 330                 eco->eco_refcount = 1;
 331                 *ecop = eco;
 332                 CDEBUG (D_INFO,
 333                         "created %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 334                         eco, eco->eco_id,
 335                         eco->eco_lsm->lsm_stripe_size,
 336                         eco->eco_lsm->lsm_stripe_count,
 337                         eco->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 338                         eco->eco_refcount, eco->eco_deleted);
 339                 return (0);
 340         }
 341
 342         if (eco2->eco_deleted)
 343                 rc = -EAGAIN;                   /* lose race */
 344         else {
 345                 eco2->eco_refcount++;           /* take existing */
 346                 *ecop = eco2;
 347                 rc = 0;
 348                 LASSERT (eco2->eco_id == eco2->eco_lsm->lsm_object_id);
 349                 CDEBUG (D_INFO,
 350                         "found(2) %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 351                         eco2, eco2->eco_id,
 352                         eco2->eco_lsm->lsm_stripe_size,
 353                         eco2->eco_lsm->lsm_stripe_count,
 354                         eco2->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 355                         eco2->eco_refcount, eco2->eco_deleted);
 356         }
 357
 358         spin_unlock (&ec->ec_lock);
 359
 360         echo_free_object (eco);
 361         return (rc);
 362 }
 363
 364 static void
 365 echo_put_object (struct ec_object *eco)
 366 {
 367         struct obd_device      *obd = eco->eco_device;
 368         struct echo_client_obd *ec = &obd->u.echo_client;
 369
 370         /* Release caller's ref on the object.
 371          * delete => mark for deletion when last ref goes
 372          */
 373
 374         spin_lock (&ec->ec_lock);
 375
 376         eco->eco_refcount--;
 377         LASSERT (eco->eco_refcount >= 0);
 378
 379         CDEBUG(D_INFO, "put %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 380                eco, eco->eco_id,
 381                eco->eco_lsm->lsm_stripe_size,
 382                eco->eco_lsm->lsm_stripe_count,
 383                eco->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 384                eco->eco_refcount, eco->eco_deleted);
 385
 386         if (eco->eco_refcount != 0 || !eco->eco_deleted) {
 387                 spin_unlock (&ec->ec_lock);
 388                 return;
 389         }
 390
 391         spin_unlock (&ec->ec_lock);
 392
 393         /* NB leave obj in the object list.  We must prevent anyone from
 394          * attempting to enqueue on this object number until we can be
 395          * sure there will be no more lock callbacks.
 396          */
 397         obd_cancel_unused(ec->ec_exp, eco->eco_lsm, 0, NULL);
 398
 399         /* now we can let it go */
 400         spin_lock (&ec->ec_lock);
 401         list_del (&eco->eco_obj_chain);
 402         spin_unlock (&ec->ec_lock);
 403
 404         LASSERT (eco->eco_refcount == 0);
 405
 406         echo_free_object (eco);
 407 }
 408
 409 static void
 410 echo_get_stripe_off_id (struct lov_stripe_md *lsm, obd_off *offp, obd_id *idp)
 411 {
 412         unsigned long stripe_count;
 413         unsigned long stripe_size;
 414         unsigned long width;
 415         unsigned long woffset;
 416         int           stripe_index;
 417         obd_off       offset;
 418
 419         if (lsm->lsm_stripe_count <= 1)
 420                 return;
 421
 422         offset       = *offp;
 423         stripe_size  = lsm->lsm_stripe_size;
 424         stripe_count = lsm->lsm_stripe_count;
 425
 426         /* width = # bytes in all stripes */
 427         width = stripe_size * stripe_count;
 428
 429         /* woffset = offset within a width; offset = whole number of widths */
 430         woffset = do_div (offset, width);
 431
 432         stripe_index = woffset / stripe_size;
 433
 434         *idp = lsm->lsm_oinfo[stripe_index]->loi_id;
 435         *offp = offset * stripe_size + woffset % stripe_size;
 436 }
 437
 438 static void
 439 echo_client_page_debug_setup(struct lov_stripe_md *lsm,
 440                              cfs_page_t *page, int rw, obd_id id,
 441                              obd_off offset, obd_off count)
 442 {
 443         char    *addr;
 444         obd_off  stripe_off;
 445         obd_id   stripe_id;
 446         int      delta;
 447
 448         /* no partial pages on the client */
 449         LASSERT(count == CFS_PAGE_SIZE);
 450
 451         addr = cfs_kmap(page);
 452
 453         for (delta = 0; delta < CFS_PAGE_SIZE; delta += OBD_ECHO_BLOCK_SIZE) {
 454                 if (rw == OBD_BRW_WRITE) {
 455                         stripe_off = offset + delta;
 456                         stripe_id = id;
 457                         echo_get_stripe_off_id(lsm, &stripe_off, &stripe_id);
 458                 } else {
 459                         stripe_off = 0xdeadbeef00c0ffeeULL;
 460                         stripe_id = 0xdeadbeef00c0ffeeULL;
 461                 }
 462                 block_debug_setup(addr + delta, OBD_ECHO_BLOCK_SIZE,
 463                                   stripe_off, stripe_id);
 464         }
 465
 466         cfs_kunmap(page);
 467 }
 468
 469 static int echo_client_page_debug_check(struct lov_stripe_md *lsm,
 470                                         cfs_page_t *page, obd_id id,
 471                                         obd_off offset, obd_off count)
 472 {
 473         obd_off stripe_off;
 474         obd_id  stripe_id;
 475         char   *addr;
 476         int     delta;
 477         int     rc;
 478         int     rc2;
 479
 480         /* no partial pages on the client */
 481         LASSERT(count == CFS_PAGE_SIZE);
 482
 483         addr = cfs_kmap(page);
 484
 485         for (rc = delta = 0; delta < CFS_PAGE_SIZE; delta += OBD_ECHO_BLOCK_SIZE) {
 486                 stripe_off = offset + delta;
 487                 stripe_id = id;
 488                 echo_get_stripe_off_id (lsm, &stripe_off, &stripe_id);
 489
 490                 rc2 = block_debug_check("test_brw",
 491                                         addr + delta, OBD_ECHO_BLOCK_SIZE,
 492                                         stripe_off, stripe_id);
 493                 if (rc2 != 0) {
 494                         CERROR ("Error in echo object "LPX64"\n", id);
 495                         rc = rc2;
 496                 }
 497         }
 498
 499         cfs_kunmap(page);
 500         return rc;
 501 }
 502
 503 static int echo_client_kbrw(struct obd_device *obd, int rw, struct obdo *oa,
 504                             struct lov_stripe_md *lsm, obd_off offset,
 505                             obd_size count, struct obd_trans_info *oti)
 506 {
 507         struct echo_client_obd *ec = &obd->u.echo_client;
 508         struct obd_info         oinfo = { { { 0 } } };
 509         obd_count               npages;
 510         struct brw_page        *pga;
 511         struct brw_page        *pgp;
 512         obd_off                 off;
 513         int                     i;
 514         int                     rc;
 515         int                     verify;
 516         int                     gfp_mask;
 517
 518         verify = ((oa->o_id) != ECHO_PERSISTENT_OBJID &&
 519                   (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 520                   (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
 521
 522         gfp_mask = ((oa->o_id & 2) == 0) ? CFS_ALLOC_STD : CFS_ALLOC_HIGHUSER;
 523
 524         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 525         LASSERT(lsm != NULL);
 526         LASSERT(lsm->lsm_object_id == oa->o_id);
 527
 528         if (count <= 0 ||
 529             (count & (~CFS_PAGE_MASK)) != 0)
 530                 return (-EINVAL);
 531
 532         /* XXX think again with misaligned I/O */
 533         npages = count >> CFS_PAGE_SHIFT;
 534
 535         OBD_ALLOC(pga, npages * sizeof(*pga));
 536         if (pga == NULL)
 537                 return (-ENOMEM);
 538
 539         for (i = 0, pgp = pga, off = offset;
 540              i < npages;
 541              i++, pgp++, off += CFS_PAGE_SIZE) {
 542
 543                 LASSERT (pgp->pg == NULL);      /* for cleanup */
 544
 545                 rc = -ENOMEM;
 546                 OBD_PAGE_ALLOC(pgp->pg, gfp_mask);
 547                 if (pgp->pg == NULL)
 548                         goto out;
 549
 550                 pgp->count = CFS_PAGE_SIZE;
 551                 pgp->off = off;
 552                 pgp->flag = 0;
 553
 554                 if (verify)
 555                         echo_client_page_debug_setup(lsm, pgp->pg, rw,
 556                                                      oa->o_id, off, pgp->count);
 557         }
 558
 559         oinfo.oi_oa = oa;
 560         oinfo.oi_md = lsm;
 561         rc = obd_brw(rw, ec->ec_exp, &oinfo, npages, pga, oti);
 562
 563  out:
 564         if (rc != 0 || rw != OBD_BRW_READ)
 565                 verify = 0;
 566
 567         for (i = 0, pgp = pga; i < npages; i++, pgp++) {
 568                 if (pgp->pg == NULL)
 569                         continue;
 570
 571                 if (verify) {
 572                         int vrc;
 573                         vrc = echo_client_page_debug_check(lsm, pgp->pg, oa->o_id,
 574                                                            pgp->off, pgp->count);
 575                         if (vrc != 0 && rc == 0)
 576                                 rc = vrc;
 577                 }
 578                 OBD_PAGE_FREE(pgp->pg);
 579         }
 580         OBD_FREE(pga, npages * sizeof(*pga));
 581         return (rc);
 582 }
 583
 584 #ifdef __KERNEL__
 585 static int echo_client_ubrw(struct obd_device *obd, int rw,
 586                             struct obdo *oa, struct lov_stripe_md *lsm,
 587                             obd_off offset, obd_size count, char *buffer,
 588                             struct obd_trans_info *oti)
 589 {
 590 #warning "echo_client_ubrw() needs to be ported on 2.6 yet"
 591         LBUG();
 592         return 0;
 593 }
 594 #endif
 595
 596 struct echo_async_state;
 597
 598 #define EAP_MAGIC 79277927
 599 struct echo_async_page {
 600         int                     eap_magic;
 601         cfs_page_t             *eap_page;
 602         void                    *eap_cookie;
 603         obd_off                 eap_off;
 604         struct echo_async_state *eap_eas;
 605         struct list_head        eap_item;
 606 };
 607
 608 #define EAP_FROM_COOKIE(c)                                                      \
 609         (LASSERT(((struct echo_async_page *)(c))->eap_magic == EAP_MAGIC),      \
 610          (struct echo_async_page *)(c))
 611
 612 struct echo_async_state {
 613         spinlock_t              eas_lock;
 614         obd_off                 eas_next_offset;
 615         obd_off                 eas_end_offset;
 616         int                     eas_in_flight;
 617         int                     eas_rc;
 618         cfs_waitq_t             eas_waitq;
 619         struct list_head        eas_avail;
 620         struct obdo             eas_oa;
 621         struct lov_stripe_md    *eas_lsm;
 622 };
 623
 624 static int eas_should_wake(struct echo_async_state *eas)
 625 {
 626         int rc = 0;
 627
 628         spin_lock(&eas->eas_lock);
 629         if (eas->eas_rc == 0 && !list_empty(&eas->eas_avail))
 630             rc = 1;
 631         spin_unlock(&eas->eas_lock);
 632         return rc;
 633 };
 634
 635 static int ec_ap_make_ready(void *data, int cmd)
 636 {
 637         /* our pages are issued ready */
 638         LBUG();
 639         return 0;
 640 }
 641 static int ec_ap_refresh_count(void *data, int cmd)
 642 {
 643         /* our pages are issued with a stable count */
 644         LBUG();
 645         return CFS_PAGE_SIZE;
 646 }
 647 static void ec_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 648 {
 649         struct echo_async_page *eap = EAP_FROM_COOKIE(data);
 650
 651         memcpy(oa, &eap->eap_eas->eas_oa, sizeof(*oa));
 652 }
 653
 654 static int ec_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 655 {
 656         struct echo_async_page *eap = EAP_FROM_COOKIE(data);
 657         struct echo_async_state *eas;
 658
 659         eas = eap->eap_eas;
 660
 661         if (cmd == OBD_BRW_READ &&
 662             eas->eas_oa.o_id != ECHO_PERSISTENT_OBJID &&
 663             (eas->eas_oa.o_valid & OBD_MD_FLFLAGS) != 0 &&
 664             (eas->eas_oa.o_flags & OBD_FL_DEBUG_CHECK) != 0)
 665                 echo_client_page_debug_check(eas->eas_lsm, eap->eap_page,
 666                                              eas->eas_oa.o_id, eap->eap_off,
 667                                              CFS_PAGE_SIZE);
 668
 669         spin_lock(&eas->eas_lock);
 670         if (rc && !eas->eas_rc)
 671                 eas->eas_rc = rc;
 672         eas->eas_in_flight--;
 673         list_add(&eap->eap_item, &eas->eas_avail);
 674         cfs_waitq_signal(&eas->eas_waitq);
 675         spin_unlock(&eas->eas_lock);
 676         return 0;
 677 }
 678
 679 static struct obd_async_page_ops ec_async_page_ops = {
 680         .ap_make_ready =        ec_ap_make_ready,
 681         .ap_refresh_count =     ec_ap_refresh_count,
 682         .ap_fill_obdo =         ec_ap_fill_obdo,
 683         .ap_completion =        ec_ap_completion,
 684 };
 685
 686 static int echo_client_async_page(struct obd_export *exp, int rw,
 687                                    struct obdo *oa, struct lov_stripe_md *lsm,
 688                                    obd_off offset, obd_size count,
 689                                    obd_size batching)
 690 {
 691         obd_count npages, i;
 692         struct echo_async_page *eap;
 693         struct echo_async_state eas;
 694         int rc = 0;
 695         struct echo_async_page **aps = NULL;
 696
 697         ENTRY;
 698 #if 0
 699         int                     verify;
 700         int                     gfp_mask;
 701
 702         verify = ((oa->o_id) != ECHO_PERSISTENT_OBJID &&
 703                   (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 704                   (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
 705
 706         gfp_mask = ((oa->o_id & 2) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 707 #endif
 708
 709         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 710
 711         if (count <= 0 ||
 712             (count & (~CFS_PAGE_MASK)) != 0 ||
 713             (lsm != NULL &&
 714              lsm->lsm_object_id != oa->o_id))
 715                 return (-EINVAL);
 716
 717         /* XXX think again with misaligned I/O */
 718         npages = batching >> CFS_PAGE_SHIFT;
 719
 720         memcpy(&eas.eas_oa, oa, sizeof(*oa));
 721         eas.eas_next_offset = offset;
 722         eas.eas_end_offset = offset + count;
 723         spin_lock_init(&eas.eas_lock);
 724         cfs_waitq_init(&eas.eas_waitq);
 725         eas.eas_in_flight = 0;
 726         eas.eas_rc = 0;
 727         eas.eas_lsm = lsm;
 728         CFS_INIT_LIST_HEAD(&eas.eas_avail);
 729
 730         OBD_ALLOC(aps, npages * sizeof aps[0]);
 731         if (aps == NULL)
 732                 return (-ENOMEM);
 733
 734         /* prepare the group of pages that we're going to be keeping
 735          * in flight */
 736         for (i = 0; i < npages; i++) {
 737                 cfs_page_t *page;
 738                 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
 739                 if (page == NULL)
 740                         GOTO(out, rc = -ENOMEM);
 741
 742                 OBD_ALLOC(eap, sizeof(*eap));
 743                 if (eap == NULL) {
 744                         OBD_PAGE_FREE(page);
 745                         GOTO(out, rc = -ENOMEM);
 746                 }
 747
 748                 eap->eap_magic = EAP_MAGIC;
 749                 eap->eap_page = page;
 750                 eap->eap_eas = &eas;
 751                 list_add_tail(&eap->eap_item, &eas.eas_avail);
 752                 aps[i] = eap;
 753         }
 754
 755         /* first we spin queueing io and being woken by its completion */
 756         spin_lock(&eas.eas_lock);
 757         for(;;) {
 758                 int rc;
 759
 760                 /* sleep until we have a page to send */
 761                 spin_unlock(&eas.eas_lock);
 762                 rc = wait_event_interruptible(eas.eas_waitq,
 763                                               eas_should_wake(&eas));
 764                 spin_lock(&eas.eas_lock);
 765                 if (rc && !eas.eas_rc)
 766                         eas.eas_rc = rc;
 767                 if (eas.eas_rc)
 768                         break;
 769                 if (list_empty(&eas.eas_avail))
 770                         continue;
 771                 eap = list_entry(eas.eas_avail.next, struct echo_async_page,
 772                                  eap_item);
 773                 list_del(&eap->eap_item);
 774                 spin_unlock(&eas.eas_lock);
 775
 776                 /* unbind the eap from its old page offset */
 777                 if (eap->eap_cookie != NULL) {
 778                         obd_teardown_async_page(exp, lsm, NULL,
 779                                                 eap->eap_cookie);
 780                         eap->eap_cookie = NULL;
 781                 }
 782
 783                 eas.eas_next_offset += CFS_PAGE_SIZE;
 784                 eap->eap_off = eas.eas_next_offset;
 785
 786                 rc = obd_prep_async_page(exp, lsm, NULL, eap->eap_page,
 787                                          eap->eap_off, &ec_async_page_ops,
 788                                          eap, &eap->eap_cookie);
 789                 if (rc) {
 790                         spin_lock(&eas.eas_lock);
 791                         eas.eas_rc = rc;
 792                         break;
 793                 }
 794
 795                 if (oa->o_id != ECHO_PERSISTENT_OBJID &&
 796                     (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 797                     (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0)
 798                         echo_client_page_debug_setup(lsm, eap->eap_page, rw,
 799                                                      oa->o_id,
 800                                                      eap->eap_off, CFS_PAGE_SIZE);
 801
 802                 /* always asserts urgent, which isn't quite right */
 803                 rc = obd_queue_async_io(exp, lsm, NULL, eap->eap_cookie,
 804                                         rw, 0, CFS_PAGE_SIZE, 0,
 805                                         ASYNC_READY | ASYNC_URGENT |
 806                                         ASYNC_COUNT_STABLE);
 807                 spin_lock(&eas.eas_lock);
 808                 if (rc && !eas.eas_rc) {
 809                         eas.eas_rc = rc;
 810                         break;
 811                 }
 812                 eas.eas_in_flight++;
 813                 if (eas.eas_next_offset == eas.eas_end_offset)
 814                         break;
 815         }
 816
 817         /* still hold the eas_lock here.. */
 818
 819         /* now we just spin waiting for all the rpcs to complete */
 820         while(eas.eas_in_flight) {
 821                 spin_unlock(&eas.eas_lock);
 822                 wait_event_interruptible(eas.eas_waitq,
 823                                          eas.eas_in_flight == 0);
 824                 spin_lock(&eas.eas_lock);
 825         }
 826         spin_unlock(&eas.eas_lock);
 827
 828 out:
 829         if (aps != NULL) {
 830                 for (i = 0; i < npages; ++ i) {
 831                         cfs_page_t *page;
 832
 833                         eap = aps[i];
 834                         page = eap->eap_page;
 835                         if (eap->eap_cookie != NULL)
 836                                 obd_teardown_async_page(exp, lsm, NULL,
 837                                                         eap->eap_cookie);
 838                         OBD_FREE(eap, sizeof(*eap));
 839                         OBD_PAGE_FREE(page);
 840                 }
 841                 OBD_FREE(aps, npages * sizeof aps[0]);
 842         }
 843
 844         RETURN(rc);
 845 }
 846
 847 static int echo_client_prep_commit(struct obd_export *exp, int rw,
 848                                    struct obdo *oa, struct lov_stripe_md *lsm,
 849                                    obd_off offset, obd_size count,
 850                                    obd_size batch, struct obd_trans_info *oti)
 851 {
 852         struct obd_ioobj ioo;
 853         struct niobuf_local *lnb;
 854         struct niobuf_remote *rnb;
 855         obd_off off;
 856         obd_size npages, tot_pages;
 857         int i, ret = 0;
 858         ENTRY;
 859
 860         if (count <= 0 || (count & (~CFS_PAGE_MASK)) != 0 ||
 861             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 862                 RETURN(-EINVAL);
 863
 864         npages = batch >> CFS_PAGE_SHIFT;
 865         tot_pages = count >> CFS_PAGE_SHIFT;
 866
 867         OBD_ALLOC(lnb, npages * sizeof(struct niobuf_local));
 868         OBD_ALLOC(rnb, npages * sizeof(struct niobuf_remote));
 869
 870         if (lnb == NULL || rnb == NULL)
 871                 GOTO(out, ret = -ENOMEM);
 872
 873         obdo_to_ioobj(oa, &ioo);
 874
 875         off = offset;
 876
 877         for(; tot_pages; tot_pages -= npages) {
 878                 if (tot_pages < npages)
 879                         npages = tot_pages;
 880
 881                 for (i = 0; i < npages; i++, off += CFS_PAGE_SIZE) {
 882                         rnb[i].offset = off;
 883                         rnb[i].len = CFS_PAGE_SIZE;
 884                 }
 885
 886                 ioo.ioo_bufcnt = npages;
 887                 oti->oti_transno = 0;
 888
 889                 ret = obd_preprw(rw, exp, oa, 1, &ioo, npages, rnb, lnb, oti,
 890                                  NULL);
 891                 if (ret != 0)
 892                         GOTO(out, ret);
 893
 894                 for (i = 0; i < npages; i++) {
 895                         cfs_page_t *page = lnb[i].page;
 896
 897                         /* read past eof? */
 898                         if (page == NULL && lnb[i].rc == 0)
 899                                 continue;
 900
 901                         if (oa->o_id == ECHO_PERSISTENT_OBJID ||
 902                             (oa->o_valid & OBD_MD_FLFLAGS) == 0 ||
 903                             (oa->o_flags & OBD_FL_DEBUG_CHECK) == 0)
 904                                 continue;
 905
 906                         if (rw == OBD_BRW_WRITE)
 907                                 echo_client_page_debug_setup(lsm, page, rw,
 908                                                              oa->o_id,
 909                                                              rnb[i].offset,
 910                                                              rnb[i].len);
 911                         else
 912                                 echo_client_page_debug_check(lsm, page,
 913                                                              oa->o_id,
 914                                                              rnb[i].offset,
 915                                                              rnb[i].len);
 916                 }
 917
 918                 ret = obd_commitrw(rw, exp, oa, 1, &ioo, npages, lnb, oti, ret);
 919                 if (ret != 0)
 920                         GOTO(out, ret);
 921         }
 922
 923 out:
 924         if (lnb)
 925                 OBD_FREE(lnb, npages * sizeof(struct niobuf_local));
 926         if (rnb)
 927                 OBD_FREE(rnb, npages * sizeof(struct niobuf_remote));
 928         RETURN(ret);
 929 }
 930
 931 int echo_client_brw_ioctl(int rw, struct obd_export *exp,
 932                           struct obd_ioctl_data *data)
 933 {
 934         struct obd_device *obd = class_exp2obd(exp);
 935         struct echo_client_obd *ec = &obd->u.echo_client;
 936         struct obd_trans_info dummy_oti = { .oti_thread_id = -1 };
 937         struct ec_object *eco;
 938         int rc;
 939         ENTRY;
 940
 941         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
 942         if (rc)
 943                 RETURN(rc);
 944
 945         data->ioc_obdo1.o_valid &= ~OBD_MD_FLHANDLE;
 946         data->ioc_obdo1.o_valid |= OBD_MD_FLGROUP;
 947         data->ioc_obdo1.o_gr = FILTER_GROUP_ECHO;
 948
 949         switch((long)data->ioc_pbuf1) {
 950         case 1:
 951                 if (data->ioc_pbuf2 == NULL) { // NULL user data pointer
 952                         rc = echo_client_kbrw(obd, rw, &data->ioc_obdo1,
 953                                               eco->eco_lsm, data->ioc_offset,
 954                                               data->ioc_count, &dummy_oti);
 955                 } else {
 956 #ifdef __KERNEL__
 957                         rc = echo_client_ubrw(obd, rw, &data->ioc_obdo1,
 958                                               eco->eco_lsm, data->ioc_offset,
 959                                               data->ioc_count, data->ioc_pbuf2,
 960                                               &dummy_oti);
 961 #endif
 962                 }
 963                 break;
 964         case 2:
 965                 rc = echo_client_async_page(ec->ec_exp, rw, &data->ioc_obdo1,
 966                                            eco->eco_lsm, data->ioc_offset,
 967                                            data->ioc_count, data->ioc_plen1);
 968                 break;
 969         case 3:
 970                 rc = echo_client_prep_commit(ec->ec_exp, rw, &data->ioc_obdo1,
 971                                             eco->eco_lsm, data->ioc_offset,
 972                                             data->ioc_count, data->ioc_plen1,
 973                                             &dummy_oti);
 974                 break;
 975         default:
 976                 rc = -EINVAL;
 977         }
 978         echo_put_object(eco);
 979         RETURN(rc);
 980 }
 981
 982 static int
 983 echo_ldlm_callback (struct ldlm_lock *lock, struct ldlm_lock_desc *new,
 984                     void *data, int flag)
 985 {
 986         struct ec_object       *eco = (struct ec_object *)data;
 987         struct echo_client_obd *ec = &(eco->eco_device->u.echo_client);
 988         struct lustre_handle    lockh;
 989         struct list_head       *el;
 990         int                     found = 0;
 991         int                     rc;
 992
 993         ldlm_lock2handle (lock, &lockh);
 994
 995         /* #ifdef this out if we're not feeling paranoid */
 996         spin_lock (&ec->ec_lock);
 997         list_for_each (el, &ec->ec_objects) {
 998                 found = (eco == list_entry(el, struct ec_object,
 999                                            eco_obj_chain));
1000                 if (found)
1001                         break;
1002         }
1003         spin_unlock (&ec->ec_lock);
1004         LASSERT (found);
1005
1006         switch (flag) {
1007         case LDLM_CB_BLOCKING:
1008                 CDEBUG(D_INFO, "blocking callback on "LPX64", handle "LPX64"\n",
1009                        eco->eco_id, lockh.cookie);
1010                 rc = ldlm_cli_cancel (&lockh);
1011                 if (rc != ELDLM_OK)
1012                         CERROR ("ldlm_cli_cancel failed: %d\n", rc);
1013                 break;
1014
1015         case LDLM_CB_CANCELING:
1016                 CDEBUG(D_INFO, "cancel callback on "LPX64", handle "LPX64"\n",
1017                        eco->eco_id, lockh.cookie);
1018                 break;
1019
1020         default:
1021                 LBUG ();
1022         }
1023
1024         return (0);
1025 }
1026
1027 static int
1028 echo_client_enqueue(struct obd_export *exp, struct obdo *oa,
1029                     int mode, obd_off offset, obd_size nob)
1030 {
1031         struct obd_device      *obd = exp->exp_obd;
1032         struct echo_client_obd *ec = &obd->u.echo_client;
1033         struct lustre_handle   *ulh = obdo_handle (oa);
1034         struct ldlm_enqueue_info einfo = { 0 };
1035         struct obd_info oinfo = { { { 0 } } };
1036         struct ec_object       *eco;
1037         struct ec_lock         *ecl;
1038         int                     rc;
1039
1040         if (!(mode == LCK_PR || mode == LCK_PW))
1041                 return -EINVAL;
1042
1043         if ((offset & (~CFS_PAGE_MASK)) != 0 ||
1044             (nob & (~CFS_PAGE_MASK)) != 0)
1045                 return -EINVAL;
1046
1047         rc = echo_get_object (&eco, obd, oa);
1048         if (rc != 0)
1049                 return rc;
1050
1051         rc = -ENOMEM;
1052         OBD_ALLOC (ecl, sizeof (*ecl));
1053         if (ecl == NULL)
1054                 goto failed_0;
1055
1056         ecl->ecl_mode = mode;
1057         ecl->ecl_object = eco;
1058         ecl->ecl_policy.l_extent.start = offset;
1059         ecl->ecl_policy.l_extent.end =
1060                 (nob == 0) ? ((obd_off) -1) : (offset + nob - 1);
1061
1062         einfo.ei_type = LDLM_EXTENT;
1063         einfo.ei_mode = mode;
1064         einfo.ei_cb_bl = echo_ldlm_callback;
1065         einfo.ei_cb_cp = ldlm_completion_ast;
1066         einfo.ei_cb_gl = NULL;
1067         einfo.ei_cbdata = eco;
1068
1069         oinfo.oi_policy = ecl->ecl_policy;
1070         oinfo.oi_lockh = &ecl->ecl_lock_handle;
1071         oinfo.oi_md = eco->eco_lsm;
1072         rc = obd_enqueue(ec->ec_exp, &oinfo, &einfo, NULL);
1073         if (rc != 0)
1074                 goto failed_1;
1075
1076         CDEBUG(D_INFO, "enqueue handle "LPX64"\n", ecl->ecl_lock_handle.cookie);
1077
1078         /* NB ecl takes object ref from echo_get_object() above */
1079         spin_lock(&ec->ec_lock);
1080
1081         list_add(&ecl->ecl_exp_chain, &exp->exp_ec_data.eced_locks);
1082         ulh->cookie = ecl->ecl_cookie = ec->ec_unique++;
1083
1084         spin_unlock(&ec->ec_lock);
1085
1086         oa->o_valid |= OBD_MD_FLHANDLE;
1087         return 0;
1088
1089  failed_1:
1090         OBD_FREE (ecl, sizeof (*ecl));
1091  failed_0:
1092         echo_put_object (eco);
1093         return (rc);
1094 }
1095
1096 static int
1097 echo_client_cancel(struct obd_export *exp, struct obdo *oa)
1098 {
1099         struct obd_device      *obd = exp->exp_obd;
1100         struct echo_client_obd *ec = &obd->u.echo_client;
1101         struct lustre_handle   *ulh = obdo_handle (oa);
1102         struct ec_lock         *ecl = NULL;
1103         int                     found = 0;
1104         struct list_head       *el;
1105         int                     rc;
1106
1107         if ((oa->o_valid & OBD_MD_FLHANDLE) == 0)
1108                 return -EINVAL;
1109
1110         spin_lock (&ec->ec_lock);
1111
1112         list_for_each (el, &exp->exp_ec_data.eced_locks) {
1113                 ecl = list_entry (el, struct ec_lock, ecl_exp_chain);
1114                 found = (ecl->ecl_cookie == ulh->cookie);
1115                 if (found) {
1116                         list_del (&ecl->ecl_exp_chain);
1117                         break;
1118                 }
1119         }
1120
1121         spin_unlock (&ec->ec_lock);
1122
1123         if (!found)
1124                 return (-ENOENT);
1125
1126         rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm, ecl->ecl_mode,
1127                         &ecl->ecl_lock_handle);
1128
1129         echo_put_object (ecl->ecl_object);
1130         OBD_FREE (ecl, sizeof (*ecl));
1131
1132         return rc;
1133 }
1134
1135 static int
1136 echo_client_iocontrol(unsigned int cmd, struct obd_export *exp,
1137                       int len, void *karg, void *uarg)
1138 {
1139         struct obd_device      *obd;
1140         struct echo_client_obd *ec;
1141         struct ec_object       *eco;
1142         struct obd_ioctl_data  *data = karg;
1143         struct obd_trans_info   dummy_oti;
1144         struct oti_req_ack_lock *ack_lock;
1145         struct obdo            *oa;
1146         int                     rw = OBD_BRW_READ;
1147         int                     rc = 0;
1148         int                     i;
1149         ENTRY;
1150
1151         unlock_kernel();
1152
1153         memset(&dummy_oti, 0, sizeof(dummy_oti));
1154
1155         obd = exp->exp_obd;
1156         ec = &obd->u.echo_client;
1157
1158         switch (cmd) {
1159         case OBD_IOC_CREATE:                    /* may create echo object */
1160                 if (!capable (CAP_SYS_ADMIN))
1161                         GOTO (out, rc = -EPERM);
1162
1163                 rc = echo_create_object (obd, 1, &data->ioc_obdo1,
1164                                          data->ioc_pbuf1, data->ioc_plen1,
1165                                          &dummy_oti);
1166                 GOTO(out, rc);
1167
1168         case OBD_IOC_DESTROY:
1169                 if (!capable (CAP_SYS_ADMIN))
1170                         GOTO (out, rc = -EPERM);
1171
1172                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1173                 if (rc == 0) {
1174                         oa = &data->ioc_obdo1;
1175                         oa->o_gr = FILTER_GROUP_ECHO;
1176                         oa->o_valid |= OBD_MD_FLGROUP;
1177                         rc = obd_destroy(ec->ec_exp, oa, eco->eco_lsm,
1178                                          &dummy_oti, NULL);
1179                         if (rc == 0)
1180                                 eco->eco_deleted = 1;
1181                         echo_put_object(eco);
1182                 }
1183                 GOTO(out, rc);
1184
1185         case OBD_IOC_GETATTR:
1186                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1187                 if (rc == 0) {
1188                         struct obd_info oinfo = { { { 0 } } };
1189                         oinfo.oi_md = eco->eco_lsm;
1190                         oinfo.oi_oa = &data->ioc_obdo1;
1191                         rc = obd_getattr(ec->ec_exp, &oinfo);
1192                         echo_put_object(eco);
1193                 }
1194                 GOTO(out, rc);
1195
1196         case OBD_IOC_SETATTR:
1197                 if (!capable (CAP_SYS_ADMIN))
1198                         GOTO (out, rc = -EPERM);
1199
1200                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1201                 if (rc == 0) {
1202                         struct obd_info oinfo = { { { 0 } } };
1203                         oinfo.oi_oa = &data->ioc_obdo1;
1204                         oinfo.oi_md = eco->eco_lsm;
1205
1206                         rc = obd_setattr(ec->ec_exp, &oinfo, NULL);
1207                         echo_put_object(eco);
1208                 }
1209                 GOTO(out, rc);
1210
1211         case OBD_IOC_BRW_WRITE:
1212                 if (!capable (CAP_SYS_ADMIN))
1213                         GOTO (out, rc = -EPERM);
1214
1215                 rw = OBD_BRW_WRITE;
1216                 /* fall through */
1217         case OBD_IOC_BRW_READ:
1218                 rc = echo_client_brw_ioctl(rw, exp, data);
1219                 GOTO(out, rc);
1220
1221         case ECHO_IOC_GET_STRIPE:
1222                 rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1223                 if (rc == 0) {
1224                         rc = echo_copyout_lsm(eco->eco_lsm, data->ioc_pbuf1,
1225                                               data->ioc_plen1);
1226                         echo_put_object(eco);
1227                 }
1228                 GOTO(out, rc);
1229
1230         case ECHO_IOC_SET_STRIPE:
1231                 if (!capable (CAP_SYS_ADMIN))
1232                         GOTO (out, rc = -EPERM);
1233
1234                 if (data->ioc_pbuf1 == NULL) {  /* unset */
1235                         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1236                         if (rc == 0) {
1237                                 eco->eco_deleted = 1;
1238                                 echo_put_object(eco);
1239                         }
1240                 } else {
1241                         rc = echo_create_object(obd, 0, &data->ioc_obdo1,
1242                                                 data->ioc_pbuf1,
1243                                                 data->ioc_plen1, &dummy_oti);
1244                 }
1245                 GOTO (out, rc);
1246
1247         case ECHO_IOC_ENQUEUE:
1248                 if (!capable (CAP_SYS_ADMIN))
1249                         GOTO (out, rc = -EPERM);
1250
1251                 rc = echo_client_enqueue(exp, &data->ioc_obdo1,
1252                                          data->ioc_conn1, /* lock mode */
1253                                    data->ioc_offset, data->ioc_count);/*extent*/
1254                 GOTO (out, rc);
1255
1256         case ECHO_IOC_CANCEL:
1257                 rc = echo_client_cancel(exp, &data->ioc_obdo1);
1258                 GOTO (out, rc);
1259
1260         default:
1261                 CERROR ("echo_ioctl(): unrecognised ioctl %#x\n", cmd);
1262                 GOTO (out, rc = -ENOTTY);
1263         }
1264
1265         EXIT;
1266  out:
1267
1268         /* XXX this should be in a helper also called by target_send_reply */
1269         for (ack_lock = dummy_oti.oti_ack_locks, i = 0; i < 4;
1270              i++, ack_lock++) {
1271                 if (!ack_lock->mode)
1272                         break;
1273                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1274         }
1275
1276         lock_kernel();
1277
1278         return rc;
1279 }
1280
1281 static int echo_client_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
1282 {
1283         struct echo_client_obd *ec = &obddev->u.echo_client;
1284         struct obd_device *tgt;
1285         struct lustre_handle conn = {0, };
1286         struct obd_uuid echo_uuid = { "ECHO_UUID" };
1287         struct obd_connect_data *ocd = NULL;
1288         int rc;
1289         ENTRY;
1290
1291         if (lcfg->lcfg_bufcount < 2 || LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1292                 CERROR("requires a TARGET OBD name\n");
1293                 RETURN(-EINVAL);
1294         }
1295
1296         tgt = class_name2obd(lustre_cfg_string(lcfg, 1));
1297         if (!tgt || !tgt->obd_attached || !tgt->obd_set_up) {
1298                 CERROR("device not attached or not set up (%s)\n",
1299                        lustre_cfg_string(lcfg, 1));
1300                 RETURN(-EINVAL);
1301         }
1302
1303         spin_lock_init (&ec->ec_lock);
1304         CFS_INIT_LIST_HEAD (&ec->ec_objects);
1305         ec->ec_unique = 0;
1306
1307         OBD_ALLOC(ocd, sizeof(*ocd));
1308         if (ocd == NULL) {
1309                 CERROR("Can't alloc ocd connecting to %s\n",
1310                        lustre_cfg_string(lcfg, 1));
1311                 return -ENOMEM;
1312         }
1313
1314         ocd->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_REQPORTAL;
1315         ocd->ocd_version = LUSTRE_VERSION_CODE;
1316         ocd->ocd_group = FILTER_GROUP_ECHO;
1317
1318         rc = obd_connect(NULL, &conn, tgt, &echo_uuid, ocd);
1319
1320         OBD_FREE(ocd, sizeof(*ocd));
1321
1322         if (rc != 0) {
1323                 CERROR("fail to connect to device %s\n",
1324                        lustre_cfg_string(lcfg, 1));
1325                 return (rc);
1326         }
1327         ec->ec_exp = class_conn2export(&conn);
1328
1329         RETURN(rc);
1330 }
1331
1332 static int echo_client_cleanup(struct obd_device *obddev)
1333 {
1334         struct list_head       *el;
1335         struct ec_object       *eco;
1336         struct echo_client_obd *ec = &obddev->u.echo_client;
1337         int rc;
1338         ENTRY;
1339
1340         if (!list_empty(&obddev->obd_exports)) {
1341                 CERROR("still has clients!\n");
1342                 RETURN(-EBUSY);
1343         }
1344
1345         /* XXX assuming sole access */
1346         while (!list_empty(&ec->ec_objects)) {
1347                 el = ec->ec_objects.next;
1348                 eco = list_entry(el, struct ec_object, eco_obj_chain);
1349
1350                 LASSERT(eco->eco_refcount == 0);
1351                 eco->eco_refcount = 1;
1352                 eco->eco_deleted = 1;
1353                 echo_put_object(eco);
1354         }
1355
1356         rc = obd_disconnect(ec->ec_exp);
1357         if (rc != 0)
1358                 CERROR("fail to disconnect device: %d\n", rc);
1359
1360         RETURN(rc);
1361 }
1362
1363 static int echo_client_connect(const struct lu_env *env,
1364                                struct lustre_handle *conn,
1365                                struct obd_device *src, struct obd_uuid *cluuid,
1366                                struct obd_connect_data *data)
1367 {
1368         struct obd_export *exp;
1369         int                rc;
1370
1371         ENTRY;
1372         rc = class_connect(conn, src, cluuid);
1373         if (rc == 0) {
1374                 exp = class_conn2export(conn);
1375                 CFS_INIT_LIST_HEAD(&exp->exp_ec_data.eced_locks);
1376                 class_export_put(exp);
1377         }
1378
1379         RETURN (rc);
1380 }
1381
1382 static int echo_client_disconnect(struct obd_export *exp)
1383 {
1384         struct obd_device      *obd;
1385         struct echo_client_obd *ec;
1386         struct ec_lock         *ecl;
1387         int                     rc;
1388         ENTRY;
1389
1390         if (exp == NULL)
1391                 GOTO(out, rc = -EINVAL);
1392
1393         obd = exp->exp_obd;
1394         ec = &obd->u.echo_client;
1395
1396         /* no more contention on export's lock list */
1397         while (!list_empty (&exp->exp_ec_data.eced_locks)) {
1398                 ecl = list_entry (exp->exp_ec_data.eced_locks.next,
1399                                   struct ec_lock, ecl_exp_chain);
1400                 list_del (&ecl->ecl_exp_chain);
1401
1402                 rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm,
1403                                  ecl->ecl_mode, &ecl->ecl_lock_handle);
1404
1405                 CDEBUG (D_INFO, "Cancel lock on object "LPX64" on disconnect "
1406                         "(%d)\n", ecl->ecl_object->eco_id, rc);
1407
1408                 echo_put_object (ecl->ecl_object);
1409                 OBD_FREE (ecl, sizeof (*ecl));
1410         }
1411
1412         rc = class_disconnect(exp);
1413         GOTO(out, rc);
1414  out:
1415         return rc;
1416 }
1417
1418 static struct obd_ops echo_obd_ops = {
1419         .o_owner       = THIS_MODULE,
1420         .o_setup       = echo_client_setup,
1421         .o_cleanup     = echo_client_cleanup,
1422         .o_iocontrol   = echo_client_iocontrol,
1423         .o_connect     = echo_client_connect,
1424         .o_disconnect  = echo_client_disconnect
1425 };
1426
1427 int echo_client_init(void)
1428 {
1429         struct lprocfs_static_vars lvars = { 0 };
1430
1431         lprocfs_echo_init_vars(&lvars);
1432         return class_register_type(&echo_obd_ops, NULL, lvars.module_vars,
1433                                    LUSTRE_ECHO_CLIENT_NAME, NULL);
1434 }
1435
1436 void echo_client_exit(void)
1437 {
1438         class_unregister_type(LUSTRE_ECHO_CLIENT_NAME);
1439 }