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                 pgp->pg = cfs_alloc_page (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                 cfs_free_page(pgp->pg);
 579         }
 580         OBD_FREE(pga, npages * sizeof(*pga));
 581         return (rc);
 582 }
 583
 584 #ifdef __KERNEL__
 585 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 586 #include <linux/iobuf.h>
 587
 588 static int echo_client_ubrw(struct obd_device *obd, int rw,
 589                             struct obdo *oa, struct lov_stripe_md *lsm,
 590                             obd_off offset, obd_size count, char *buffer,
 591                             struct obd_trans_info *oti)
 592 {
 593         struct echo_client_obd *ec = &obd->u.echo_client;
 594         struct obd_info         oinfo = { { { 0 } } };
 595         obd_count               npages;
 596         struct brw_page        *pga;
 597         struct brw_page        *pgp;
 598         obd_off                 off;
 599         struct kiobuf          *kiobuf;
 600         int                     i;
 601         int                     rc;
 602
 603         LASSERT (rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 604
 605         /* NB: for now, only whole pages, page aligned */
 606
 607         if (count <= 0 ||
 608             ((long)buffer & (~CFS_PAGE_MASK)) != 0 ||
 609             (count & (~CFS_PAGE_MASK)) != 0 ||
 610             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 611                 return (-EINVAL);
 612
 613         /* XXX think again with misaligned I/O */
 614         npages = count >> CFS_PAGE_SHIFT;
 615
 616         OBD_ALLOC(pga, npages * sizeof(*pga));
 617         if (pga == NULL)
 618                 return (-ENOMEM);
 619
 620         rc = alloc_kiovec (1, &kiobuf);
 621         if (rc != 0)
 622                 goto out_1;
 623
 624         rc = map_user_kiobuf ((rw == OBD_BRW_READ) ? READ : WRITE,
 625                               kiobuf, (unsigned long)buffer, count);
 626         if (rc != 0)
 627                 goto out_2;
 628
 629         LASSERT (kiobuf->offset == 0);
 630         LASSERT (kiobuf->nr_pages == npages);
 631
 632         for (i = 0, off = offset, pgp = pga;
 633              i < npages;
 634              i++, off += CFS_PAGE_SIZE, pgp++) {
 635                 pgp->off = off;
 636                 pgp->pg = kiobuf->maplist[i];
 637                 pgp->count = CFS_PAGE_SIZE;
 638                 pgp->flag = 0;
 639         }
 640
 641         oinfo.oi_oa = oa;
 642         oinfo.oi_md = lsm;
 643         rc = obd_brw(rw, ec->ec_exp, &oinfo, npages, pga, oti);
 644
 645         //        if (rw == OBD_BRW_READ)
 646         //                mark_dirty_kiobuf (kiobuf, count);
 647
 648         unmap_kiobuf (kiobuf);
 649  out_2:
 650         free_kiovec (1, &kiobuf);
 651  out_1:
 652         OBD_FREE(pga, npages * sizeof(*pga));
 653         return (rc);
 654 }
 655 #else
 656 static int echo_client_ubrw(struct obd_device *obd, int rw,
 657                             struct obdo *oa, struct lov_stripe_md *lsm,
 658                             obd_off offset, obd_size count, char *buffer,
 659                             struct obd_trans_info *oti)
 660 {
 661 #warning "echo_client_ubrw() needs to be ported on 2.6 yet"
 662         LBUG();
 663         return 0;
 664 }
 665 #endif
 666 #endif
 667
 668 struct echo_async_state;
 669
 670 #define EAP_MAGIC 79277927
 671 struct echo_async_page {
 672         int                     eap_magic;
 673         cfs_page_t             *eap_page;
 674         void                    *eap_cookie;
 675         obd_off                 eap_off;
 676         struct echo_async_state *eap_eas;
 677         struct list_head        eap_item;
 678 };
 679
 680 #define EAP_FROM_COOKIE(c)                                                      \
 681         (LASSERT(((struct echo_async_page *)(c))->eap_magic == EAP_MAGIC),      \
 682          (struct echo_async_page *)(c))
 683
 684 struct echo_async_state {
 685         spinlock_t              eas_lock;
 686         obd_off                 eas_next_offset;
 687         obd_off                 eas_end_offset;
 688         int                     eas_in_flight;
 689         int                     eas_rc;
 690         cfs_waitq_t             eas_waitq;
 691         struct list_head        eas_avail;
 692         struct obdo             eas_oa;
 693         struct lov_stripe_md    *eas_lsm;
 694 };
 695
 696 static int eas_should_wake(struct echo_async_state *eas)
 697 {
 698         int rc = 0;
 699
 700         spin_lock(&eas->eas_lock);
 701         if (eas->eas_rc == 0 && !list_empty(&eas->eas_avail))
 702             rc = 1;
 703         spin_unlock(&eas->eas_lock);
 704         return rc;
 705 };
 706
 707 static int ec_ap_make_ready(void *data, int cmd)
 708 {
 709         /* our pages are issued ready */
 710         LBUG();
 711         return 0;
 712 }
 713 static int ec_ap_refresh_count(void *data, int cmd)
 714 {
 715         /* our pages are issued with a stable count */
 716         LBUG();
 717         return CFS_PAGE_SIZE;
 718 }
 719 static void ec_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 720 {
 721         struct echo_async_page *eap = EAP_FROM_COOKIE(data);
 722
 723         memcpy(oa, &eap->eap_eas->eas_oa, sizeof(*oa));
 724 }
 725
 726 static int ec_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 727 {
 728         struct echo_async_page *eap = EAP_FROM_COOKIE(data);
 729         struct echo_async_state *eas;
 730
 731         eas = eap->eap_eas;
 732
 733         if (cmd == OBD_BRW_READ &&
 734             eas->eas_oa.o_id != ECHO_PERSISTENT_OBJID &&
 735             (eas->eas_oa.o_valid & OBD_MD_FLFLAGS) != 0 &&
 736             (eas->eas_oa.o_flags & OBD_FL_DEBUG_CHECK) != 0)
 737                 echo_client_page_debug_check(eas->eas_lsm, eap->eap_page,
 738                                              eas->eas_oa.o_id, eap->eap_off,
 739                                              CFS_PAGE_SIZE);
 740
 741         spin_lock(&eas->eas_lock);
 742         if (rc && !eas->eas_rc)
 743                 eas->eas_rc = rc;
 744         eas->eas_in_flight--;
 745         list_add(&eap->eap_item, &eas->eas_avail);
 746         cfs_waitq_signal(&eas->eas_waitq);
 747         spin_unlock(&eas->eas_lock);
 748         return 0;
 749 }
 750
 751 static struct obd_async_page_ops ec_async_page_ops = {
 752         .ap_make_ready =        ec_ap_make_ready,
 753         .ap_refresh_count =     ec_ap_refresh_count,
 754         .ap_fill_obdo =         ec_ap_fill_obdo,
 755         .ap_completion =        ec_ap_completion,
 756 };
 757
 758 static int echo_client_async_page(struct obd_export *exp, int rw,
 759                                    struct obdo *oa, struct lov_stripe_md *lsm,
 760                                    obd_off offset, obd_size count,
 761                                    obd_size batching)
 762 {
 763         obd_count npages, i;
 764         struct echo_async_page *eap;
 765         struct echo_async_state eas;
 766         int rc = 0;
 767         struct echo_async_page **aps = NULL;
 768
 769         ENTRY;
 770 #if 0
 771         int                     verify;
 772         int                     gfp_mask;
 773
 774         verify = ((oa->o_id) != ECHO_PERSISTENT_OBJID &&
 775                   (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 776                   (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
 777
 778         gfp_mask = ((oa->o_id & 2) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 779 #endif
 780
 781         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 782
 783         if (count <= 0 ||
 784             (count & (~CFS_PAGE_MASK)) != 0 ||
 785             (lsm != NULL &&
 786              lsm->lsm_object_id != oa->o_id))
 787                 return (-EINVAL);
 788
 789         /* XXX think again with misaligned I/O */
 790         npages = batching >> CFS_PAGE_SHIFT;
 791
 792         memcpy(&eas.eas_oa, oa, sizeof(*oa));
 793         eas.eas_next_offset = offset;
 794         eas.eas_end_offset = offset + count;
 795         spin_lock_init(&eas.eas_lock);
 796         cfs_waitq_init(&eas.eas_waitq);
 797         eas.eas_in_flight = 0;
 798         eas.eas_rc = 0;
 799         eas.eas_lsm = lsm;
 800         CFS_INIT_LIST_HEAD(&eas.eas_avail);
 801
 802         OBD_ALLOC(aps, npages * sizeof aps[0]);
 803         if (aps == NULL)
 804                 return (-ENOMEM);
 805
 806         /* prepare the group of pages that we're going to be keeping
 807          * in flight */
 808         for (i = 0; i < npages; i++) {
 809                 cfs_page_t *page = cfs_alloc_page(CFS_ALLOC_STD);
 810                 if (page == NULL)
 811                         GOTO(out, rc = -ENOMEM);
 812
 813                 OBD_ALLOC(eap, sizeof(*eap));
 814                 if (eap == NULL) {
 815                         cfs_free_page(page);
 816                         GOTO(out, rc = -ENOMEM);
 817                 }
 818
 819                 eap->eap_magic = EAP_MAGIC;
 820                 eap->eap_page = page;
 821                 eap->eap_eas = &eas;
 822                 list_add_tail(&eap->eap_item, &eas.eas_avail);
 823                 aps[i] = eap;
 824         }
 825
 826         /* first we spin queueing io and being woken by its completion */
 827         spin_lock(&eas.eas_lock);
 828         for(;;) {
 829                 int rc;
 830
 831                 /* sleep until we have a page to send */
 832                 spin_unlock(&eas.eas_lock);
 833                 rc = wait_event_interruptible(eas.eas_waitq,
 834                                               eas_should_wake(&eas));
 835                 spin_lock(&eas.eas_lock);
 836                 if (rc && !eas.eas_rc)
 837                         eas.eas_rc = rc;
 838                 if (eas.eas_rc)
 839                         break;
 840                 if (list_empty(&eas.eas_avail))
 841                         continue;
 842                 eap = list_entry(eas.eas_avail.next, struct echo_async_page,
 843                                  eap_item);
 844                 list_del(&eap->eap_item);
 845                 spin_unlock(&eas.eas_lock);
 846
 847                 /* unbind the eap from its old page offset */
 848                 if (eap->eap_cookie != NULL) {
 849                         obd_teardown_async_page(exp, lsm, NULL,
 850                                                 eap->eap_cookie);
 851                         eap->eap_cookie = NULL;
 852                 }
 853
 854                 eas.eas_next_offset += CFS_PAGE_SIZE;
 855                 eap->eap_off = eas.eas_next_offset;
 856
 857                 rc = obd_prep_async_page(exp, lsm, NULL, eap->eap_page,
 858                                          eap->eap_off, &ec_async_page_ops,
 859                                          eap, &eap->eap_cookie);
 860                 if (rc) {
 861                         spin_lock(&eas.eas_lock);
 862                         eas.eas_rc = rc;
 863                         break;
 864                 }
 865
 866                 if (oa->o_id != ECHO_PERSISTENT_OBJID &&
 867                     (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 868                     (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0)
 869                         echo_client_page_debug_setup(lsm, eap->eap_page, rw,
 870                                                      oa->o_id,
 871                                                      eap->eap_off, CFS_PAGE_SIZE);
 872
 873                 /* always asserts urgent, which isn't quite right */
 874                 rc = obd_queue_async_io(exp, lsm, NULL, eap->eap_cookie,
 875                                         rw, 0, CFS_PAGE_SIZE, 0,
 876                                         ASYNC_READY | ASYNC_URGENT |
 877                                         ASYNC_COUNT_STABLE);
 878                 spin_lock(&eas.eas_lock);
 879                 if (rc && !eas.eas_rc) {
 880                         eas.eas_rc = rc;
 881                         break;
 882                 }
 883                 eas.eas_in_flight++;
 884                 if (eas.eas_next_offset == eas.eas_end_offset)
 885                         break;
 886         }
 887
 888         /* still hold the eas_lock here.. */
 889
 890         /* now we just spin waiting for all the rpcs to complete */
 891         while(eas.eas_in_flight) {
 892                 spin_unlock(&eas.eas_lock);
 893                 wait_event_interruptible(eas.eas_waitq,
 894                                          eas.eas_in_flight == 0);
 895                 spin_lock(&eas.eas_lock);
 896         }
 897         spin_unlock(&eas.eas_lock);
 898
 899 out:
 900         if (aps != NULL) {
 901                 for (i = 0; i < npages; ++ i) {
 902                         cfs_page_t *page;
 903
 904                         eap = aps[i];
 905                         page = eap->eap_page;
 906                         if (eap->eap_cookie != NULL)
 907                                 obd_teardown_async_page(exp, lsm, NULL,
 908                                                         eap->eap_cookie);
 909                         OBD_FREE(eap, sizeof(*eap));
 910                         cfs_free_page(page);
 911                 }
 912                 OBD_FREE(aps, npages * sizeof aps[0]);
 913         }
 914
 915         RETURN(rc);
 916 }
 917
 918 static int echo_client_prep_commit(struct obd_export *exp, int rw,
 919                                    struct obdo *oa, struct lov_stripe_md *lsm,
 920                                    obd_off offset, obd_size count,
 921                                    obd_size batch, struct obd_trans_info *oti)
 922 {
 923         struct obd_ioobj ioo;
 924         struct niobuf_local *lnb;
 925         struct niobuf_remote *rnb;
 926         obd_off off;
 927         obd_size npages, tot_pages;
 928         int i, ret = 0;
 929         ENTRY;
 930
 931         if (count <= 0 || (count & (~CFS_PAGE_MASK)) != 0 ||
 932             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 933                 RETURN(-EINVAL);
 934
 935         npages = batch >> CFS_PAGE_SHIFT;
 936         tot_pages = count >> CFS_PAGE_SHIFT;
 937
 938         OBD_ALLOC(lnb, npages * sizeof(struct niobuf_local));
 939         OBD_ALLOC(rnb, npages * sizeof(struct niobuf_remote));
 940
 941         if (lnb == NULL || rnb == NULL)
 942                 GOTO(out, ret = -ENOMEM);
 943
 944         obdo_to_ioobj(oa, &ioo);
 945
 946         off = offset;
 947
 948         for(; tot_pages; tot_pages -= npages) {
 949                 if (tot_pages < npages)
 950                         npages = tot_pages;
 951
 952                 for (i = 0; i < npages; i++, off += CFS_PAGE_SIZE) {
 953                         rnb[i].offset = off;
 954                         rnb[i].len = CFS_PAGE_SIZE;
 955                 }
 956
 957                 ioo.ioo_bufcnt = npages;
 958                 oti->oti_transno = 0;
 959
 960                 ret = obd_preprw(rw, exp, oa, 1, &ioo, npages, rnb, lnb, oti,
 961                                  NULL);
 962                 if (ret != 0)
 963                         GOTO(out, ret);
 964
 965                 for (i = 0; i < npages; i++) {
 966                         cfs_page_t *page = lnb[i].page;
 967
 968                         /* read past eof? */
 969                         if (page == NULL && lnb[i].rc == 0)
 970                                 continue;
 971
 972                         if (oa->o_id == ECHO_PERSISTENT_OBJID ||
 973                             (oa->o_valid & OBD_MD_FLFLAGS) == 0 ||
 974                             (oa->o_flags & OBD_FL_DEBUG_CHECK) == 0)
 975                                 continue;
 976
 977                         if (rw == OBD_BRW_WRITE)
 978                                 echo_client_page_debug_setup(lsm, page, rw,
 979                                                              oa->o_id,
 980                                                              rnb[i].offset,
 981                                                              rnb[i].len);
 982                         else
 983                                 echo_client_page_debug_check(lsm, page,
 984                                                              oa->o_id,
 985                                                              rnb[i].offset,
 986                                                              rnb[i].len);
 987                 }
 988
 989                 ret = obd_commitrw(rw, exp, oa, 1, &ioo, npages, lnb, oti, ret);
 990                 if (ret != 0)
 991                         GOTO(out, ret);
 992         }
 993
 994 out:
 995         if (lnb)
 996                 OBD_FREE(lnb, npages * sizeof(struct niobuf_local));
 997         if (rnb)
 998                 OBD_FREE(rnb, npages * sizeof(struct niobuf_remote));
 999         RETURN(ret);
1000 }
1001
1002 int echo_client_brw_ioctl(int rw, struct obd_export *exp,
1003                           struct obd_ioctl_data *data)
1004 {
1005         struct obd_device *obd = class_exp2obd(exp);
1006         struct echo_client_obd *ec = &obd->u.echo_client;
1007         struct obd_trans_info dummy_oti = { .oti_thread_id = -1 };
1008         struct ec_object *eco;
1009         int rc;
1010         ENTRY;
1011
1012         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1013         if (rc)
1014                 RETURN(rc);
1015
1016         data->ioc_obdo1.o_valid &= ~OBD_MD_FLHANDLE;
1017         data->ioc_obdo1.o_valid |= OBD_MD_FLGROUP;
1018         data->ioc_obdo1.o_gr = FILTER_GROUP_ECHO;
1019
1020         switch((long)data->ioc_pbuf1) {
1021         case 1:
1022                 if (data->ioc_pbuf2 == NULL) { // NULL user data pointer
1023                         rc = echo_client_kbrw(obd, rw, &data->ioc_obdo1,
1024                                               eco->eco_lsm, data->ioc_offset,
1025                                               data->ioc_count, &dummy_oti);
1026                 } else {
1027 #ifdef __KERNEL__
1028                         rc = echo_client_ubrw(obd, rw, &data->ioc_obdo1,
1029                                               eco->eco_lsm, data->ioc_offset,
1030                                               data->ioc_count, data->ioc_pbuf2,
1031                                               &dummy_oti);
1032 #endif
1033                 }
1034                 break;
1035         case 2:
1036                 rc = echo_client_async_page(ec->ec_exp, rw, &data->ioc_obdo1,
1037                                            eco->eco_lsm, data->ioc_offset,
1038                                            data->ioc_count, data->ioc_plen1);
1039                 break;
1040         case 3:
1041                 rc = echo_client_prep_commit(ec->ec_exp, rw, &data->ioc_obdo1,
1042                                             eco->eco_lsm, data->ioc_offset,
1043                                             data->ioc_count, data->ioc_plen1,
1044                                             &dummy_oti);
1045                 break;
1046         default:
1047                 rc = -EINVAL;
1048         }
1049         echo_put_object(eco);
1050         RETURN(rc);
1051 }
1052
1053 static int
1054 echo_ldlm_callback (struct ldlm_lock *lock, struct ldlm_lock_desc *new,
1055                     void *data, int flag)
1056 {
1057         struct ec_object       *eco = (struct ec_object *)data;
1058         struct echo_client_obd *ec = &(eco->eco_device->u.echo_client);
1059         struct lustre_handle    lockh;
1060         struct list_head       *el;
1061         int                     found = 0;
1062         int                     rc;
1063
1064         ldlm_lock2handle (lock, &lockh);
1065
1066         /* #ifdef this out if we're not feeling paranoid */
1067         spin_lock (&ec->ec_lock);
1068         list_for_each (el, &ec->ec_objects) {
1069                 found = (eco == list_entry(el, struct ec_object,
1070                                            eco_obj_chain));
1071                 if (found)
1072                         break;
1073         }
1074         spin_unlock (&ec->ec_lock);
1075         LASSERT (found);
1076
1077         switch (flag) {
1078         case LDLM_CB_BLOCKING:
1079                 CDEBUG(D_INFO, "blocking callback on "LPX64", handle "LPX64"\n",
1080                        eco->eco_id, lockh.cookie);
1081                 rc = ldlm_cli_cancel (&lockh);
1082                 if (rc != ELDLM_OK)
1083                         CERROR ("ldlm_cli_cancel failed: %d\n", rc);
1084                 break;
1085
1086         case LDLM_CB_CANCELING:
1087                 CDEBUG(D_INFO, "cancel callback on "LPX64", handle "LPX64"\n",
1088                        eco->eco_id, lockh.cookie);
1089                 break;
1090
1091         default:
1092                 LBUG ();
1093         }
1094
1095         return (0);
1096 }
1097
1098 static int
1099 echo_client_enqueue(struct obd_export *exp, struct obdo *oa,
1100                     int mode, obd_off offset, obd_size nob)
1101 {
1102         struct obd_device      *obd = exp->exp_obd;
1103         struct echo_client_obd *ec = &obd->u.echo_client;
1104         struct lustre_handle   *ulh = obdo_handle (oa);
1105         struct ldlm_enqueue_info einfo = { 0 };
1106         struct obd_info oinfo = { { { 0 } } };
1107         struct ec_object       *eco;
1108         struct ec_lock         *ecl;
1109         int                     rc;
1110
1111         if (!(mode == LCK_PR || mode == LCK_PW))
1112                 return -EINVAL;
1113
1114         if ((offset & (~CFS_PAGE_MASK)) != 0 ||
1115             (nob & (~CFS_PAGE_MASK)) != 0)
1116                 return -EINVAL;
1117
1118         rc = echo_get_object (&eco, obd, oa);
1119         if (rc != 0)
1120                 return rc;
1121
1122         rc = -ENOMEM;
1123         OBD_ALLOC (ecl, sizeof (*ecl));
1124         if (ecl == NULL)
1125                 goto failed_0;
1126
1127         ecl->ecl_mode = mode;
1128         ecl->ecl_object = eco;
1129         ecl->ecl_policy.l_extent.start = offset;
1130         ecl->ecl_policy.l_extent.end =
1131                 (nob == 0) ? ((obd_off) -1) : (offset + nob - 1);
1132
1133         einfo.ei_type = LDLM_EXTENT;
1134         einfo.ei_mode = mode;
1135         einfo.ei_cb_bl = echo_ldlm_callback;
1136         einfo.ei_cb_cp = ldlm_completion_ast;
1137         einfo.ei_cb_gl = NULL;
1138         einfo.ei_cbdata = eco;
1139
1140         oinfo.oi_policy = ecl->ecl_policy;
1141         oinfo.oi_lockh = &ecl->ecl_lock_handle;
1142         oinfo.oi_md = eco->eco_lsm;
1143         rc = obd_enqueue(ec->ec_exp, &oinfo, &einfo, NULL);
1144         if (rc != 0)
1145                 goto failed_1;
1146
1147         CDEBUG(D_INFO, "enqueue handle "LPX64"\n", ecl->ecl_lock_handle.cookie);
1148
1149         /* NB ecl takes object ref from echo_get_object() above */
1150         spin_lock(&ec->ec_lock);
1151
1152         list_add(&ecl->ecl_exp_chain, &exp->exp_ec_data.eced_locks);
1153         ulh->cookie = ecl->ecl_cookie = ec->ec_unique++;
1154
1155         spin_unlock(&ec->ec_lock);
1156
1157         oa->o_valid |= OBD_MD_FLHANDLE;
1158         return 0;
1159
1160  failed_1:
1161         OBD_FREE (ecl, sizeof (*ecl));
1162  failed_0:
1163         echo_put_object (eco);
1164         return (rc);
1165 }
1166
1167 static int
1168 echo_client_cancel(struct obd_export *exp, struct obdo *oa)
1169 {
1170         struct obd_device      *obd = exp->exp_obd;
1171         struct echo_client_obd *ec = &obd->u.echo_client;
1172         struct lustre_handle   *ulh = obdo_handle (oa);
1173         struct ec_lock         *ecl = NULL;
1174         int                     found = 0;
1175         struct list_head       *el;
1176         int                     rc;
1177
1178         if ((oa->o_valid & OBD_MD_FLHANDLE) == 0)
1179                 return -EINVAL;
1180
1181         spin_lock (&ec->ec_lock);
1182
1183         list_for_each (el, &exp->exp_ec_data.eced_locks) {
1184                 ecl = list_entry (el, struct ec_lock, ecl_exp_chain);
1185                 found = (ecl->ecl_cookie == ulh->cookie);
1186                 if (found) {
1187                         list_del (&ecl->ecl_exp_chain);
1188                         break;
1189                 }
1190         }
1191
1192         spin_unlock (&ec->ec_lock);
1193
1194         if (!found)
1195                 return (-ENOENT);
1196
1197         rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm, ecl->ecl_mode,
1198                         &ecl->ecl_lock_handle);
1199
1200         echo_put_object (ecl->ecl_object);
1201         OBD_FREE (ecl, sizeof (*ecl));
1202
1203         return rc;
1204 }
1205
1206 static int
1207 echo_client_iocontrol(unsigned int cmd, struct obd_export *exp,
1208                       int len, void *karg, void *uarg)
1209 {
1210         struct obd_device      *obd;
1211         struct echo_client_obd *ec;
1212         struct ec_object       *eco;
1213         struct obd_ioctl_data  *data = karg;
1214         struct obd_trans_info   dummy_oti;
1215         struct oti_req_ack_lock *ack_lock;
1216         struct obdo            *oa;
1217         int                     rw = OBD_BRW_READ;
1218         int                     rc = 0;
1219         int                     i;
1220         ENTRY;
1221
1222         unlock_kernel();
1223
1224         memset(&dummy_oti, 0, sizeof(dummy_oti));
1225
1226         obd = exp->exp_obd;
1227         ec = &obd->u.echo_client;
1228
1229         switch (cmd) {
1230         case OBD_IOC_CREATE:                    /* may create echo object */
1231                 if (!capable (CAP_SYS_ADMIN))
1232                         GOTO (out, rc = -EPERM);
1233
1234                 rc = echo_create_object (obd, 1, &data->ioc_obdo1,
1235                                          data->ioc_pbuf1, data->ioc_plen1,
1236                                          &dummy_oti);
1237                 GOTO(out, rc);
1238
1239         case OBD_IOC_DESTROY:
1240                 if (!capable (CAP_SYS_ADMIN))
1241                         GOTO (out, rc = -EPERM);
1242
1243                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1244                 if (rc == 0) {
1245                         oa = &data->ioc_obdo1;
1246                         oa->o_gr = FILTER_GROUP_ECHO;
1247                         oa->o_valid |= OBD_MD_FLGROUP;
1248                         rc = obd_destroy(ec->ec_exp, oa, eco->eco_lsm,
1249                                          &dummy_oti, NULL);
1250                         if (rc == 0)
1251                                 eco->eco_deleted = 1;
1252                         echo_put_object(eco);
1253                 }
1254                 GOTO(out, rc);
1255
1256         case OBD_IOC_GETATTR:
1257                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1258                 if (rc == 0) {
1259                         struct obd_info oinfo = { { { 0 } } };
1260                         oinfo.oi_md = eco->eco_lsm;
1261                         oinfo.oi_oa = &data->ioc_obdo1;
1262                         rc = obd_getattr(ec->ec_exp, &oinfo);
1263                         echo_put_object(eco);
1264                 }
1265                 GOTO(out, rc);
1266
1267         case OBD_IOC_SETATTR:
1268                 if (!capable (CAP_SYS_ADMIN))
1269                         GOTO (out, rc = -EPERM);
1270
1271                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1272                 if (rc == 0) {
1273                         struct obd_info oinfo = { { { 0 } } };
1274                         oinfo.oi_oa = &data->ioc_obdo1;
1275                         oinfo.oi_md = eco->eco_lsm;
1276
1277                         rc = obd_setattr(ec->ec_exp, &oinfo, NULL);
1278                         echo_put_object(eco);
1279                 }
1280                 GOTO(out, rc);
1281
1282         case OBD_IOC_BRW_WRITE:
1283                 if (!capable (CAP_SYS_ADMIN))
1284                         GOTO (out, rc = -EPERM);
1285
1286                 rw = OBD_BRW_WRITE;
1287                 /* fall through */
1288         case OBD_IOC_BRW_READ:
1289                 rc = echo_client_brw_ioctl(rw, exp, data);
1290                 GOTO(out, rc);
1291
1292         case ECHO_IOC_GET_STRIPE:
1293                 rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1294                 if (rc == 0) {
1295                         rc = echo_copyout_lsm(eco->eco_lsm, data->ioc_pbuf1,
1296                                               data->ioc_plen1);
1297                         echo_put_object(eco);
1298                 }
1299                 GOTO(out, rc);
1300
1301         case ECHO_IOC_SET_STRIPE:
1302                 if (!capable (CAP_SYS_ADMIN))
1303                         GOTO (out, rc = -EPERM);
1304
1305                 if (data->ioc_pbuf1 == NULL) {  /* unset */
1306                         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1307                         if (rc == 0) {
1308                                 eco->eco_deleted = 1;
1309                                 echo_put_object(eco);
1310                         }
1311                 } else {
1312                         rc = echo_create_object(obd, 0, &data->ioc_obdo1,
1313                                                 data->ioc_pbuf1,
1314                                                 data->ioc_plen1, &dummy_oti);
1315                 }
1316                 GOTO (out, rc);
1317
1318         case ECHO_IOC_ENQUEUE:
1319                 if (!capable (CAP_SYS_ADMIN))
1320                         GOTO (out, rc = -EPERM);
1321
1322                 rc = echo_client_enqueue(exp, &data->ioc_obdo1,
1323                                          data->ioc_conn1, /* lock mode */
1324                                    data->ioc_offset, data->ioc_count);/*extent*/
1325                 GOTO (out, rc);
1326
1327         case ECHO_IOC_CANCEL:
1328                 rc = echo_client_cancel(exp, &data->ioc_obdo1);
1329                 GOTO (out, rc);
1330
1331         default:
1332                 CERROR ("echo_ioctl(): unrecognised ioctl %#x\n", cmd);
1333                 GOTO (out, rc = -ENOTTY);
1334         }
1335
1336         EXIT;
1337  out:
1338
1339         /* XXX this should be in a helper also called by target_send_reply */
1340         for (ack_lock = dummy_oti.oti_ack_locks, i = 0; i < 4;
1341              i++, ack_lock++) {
1342                 if (!ack_lock->mode)
1343                         break;
1344                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1345         }
1346
1347         lock_kernel();
1348
1349         return rc;
1350 }
1351
1352 static int echo_client_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
1353 {
1354         struct echo_client_obd *ec = &obddev->u.echo_client;
1355         struct obd_device *tgt;
1356         struct lustre_handle conn = {0, };
1357         struct obd_uuid echo_uuid = { "ECHO_UUID" };
1358         struct obd_connect_data *ocd = NULL;
1359         int rc;
1360         ENTRY;
1361
1362         if (lcfg->lcfg_bufcount < 2 || LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1363                 CERROR("requires a TARGET OBD name\n");
1364                 RETURN(-EINVAL);
1365         }
1366
1367         tgt = class_name2obd(lustre_cfg_string(lcfg, 1));
1368         if (!tgt || !tgt->obd_attached || !tgt->obd_set_up) {
1369                 CERROR("device not attached or not set up (%s)\n",
1370                        lustre_cfg_string(lcfg, 1));
1371                 RETURN(-EINVAL);
1372         }
1373
1374         spin_lock_init (&ec->ec_lock);
1375         CFS_INIT_LIST_HEAD (&ec->ec_objects);
1376         ec->ec_unique = 0;
1377
1378         OBD_ALLOC(ocd, sizeof(*ocd));
1379         if (ocd == NULL) {
1380                 CERROR("Can't alloc ocd connecting to %s\n",
1381                        lustre_cfg_string(lcfg, 1));
1382                 return -ENOMEM;
1383         }
1384
1385         ocd->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_REQPORTAL;
1386         ocd->ocd_version = LUSTRE_VERSION_CODE;
1387         ocd->ocd_group = FILTER_GROUP_ECHO;
1388
1389         rc = obd_connect(NULL, &conn, tgt, &echo_uuid, ocd);
1390
1391         OBD_FREE(ocd, sizeof(*ocd));
1392
1393         if (rc != 0) {
1394                 CERROR("fail to connect to device %s\n",
1395                        lustre_cfg_string(lcfg, 1));
1396                 return (rc);
1397         }
1398         ec->ec_exp = class_conn2export(&conn);
1399
1400         RETURN(rc);
1401 }
1402
1403 static int echo_client_cleanup(struct obd_device *obddev)
1404 {
1405         struct list_head       *el;
1406         struct ec_object       *eco;
1407         struct echo_client_obd *ec = &obddev->u.echo_client;
1408         int rc;
1409         ENTRY;
1410
1411         if (!list_empty(&obddev->obd_exports)) {
1412                 CERROR("still has clients!\n");
1413                 RETURN(-EBUSY);
1414         }
1415
1416         /* XXX assuming sole access */
1417         while (!list_empty(&ec->ec_objects)) {
1418                 el = ec->ec_objects.next;
1419                 eco = list_entry(el, struct ec_object, eco_obj_chain);
1420
1421                 LASSERT(eco->eco_refcount == 0);
1422                 eco->eco_refcount = 1;
1423                 eco->eco_deleted = 1;
1424                 echo_put_object(eco);
1425         }
1426
1427         rc = obd_disconnect(ec->ec_exp);
1428         if (rc != 0)
1429                 CERROR("fail to disconnect device: %d\n", rc);
1430
1431         RETURN(rc);
1432 }
1433
1434 static int echo_client_connect(const struct lu_env *env,
1435                                struct lustre_handle *conn,
1436                                struct obd_device *src, struct obd_uuid *cluuid,
1437                                struct obd_connect_data *data)
1438 {
1439         struct obd_export *exp;
1440         int                rc;
1441
1442         ENTRY;
1443         rc = class_connect(conn, src, cluuid);
1444         if (rc == 0) {
1445                 exp = class_conn2export(conn);
1446                 CFS_INIT_LIST_HEAD(&exp->exp_ec_data.eced_locks);
1447                 class_export_put(exp);
1448         }
1449
1450         RETURN (rc);
1451 }
1452
1453 static int echo_client_disconnect(struct obd_export *exp)
1454 {
1455         struct obd_device      *obd;
1456         struct echo_client_obd *ec;
1457         struct ec_lock         *ecl;
1458         int                     rc;
1459         ENTRY;
1460
1461         if (exp == NULL)
1462                 GOTO(out, rc = -EINVAL);
1463
1464         obd = exp->exp_obd;
1465         ec = &obd->u.echo_client;
1466
1467         /* no more contention on export's lock list */
1468         while (!list_empty (&exp->exp_ec_data.eced_locks)) {
1469                 ecl = list_entry (exp->exp_ec_data.eced_locks.next,
1470                                   struct ec_lock, ecl_exp_chain);
1471                 list_del (&ecl->ecl_exp_chain);
1472
1473                 rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm,
1474                                  ecl->ecl_mode, &ecl->ecl_lock_handle);
1475
1476                 CDEBUG (D_INFO, "Cancel lock on object "LPX64" on disconnect "
1477                         "(%d)\n", ecl->ecl_object->eco_id, rc);
1478
1479                 echo_put_object (ecl->ecl_object);
1480                 OBD_FREE (ecl, sizeof (*ecl));
1481         }
1482
1483         rc = class_disconnect(exp);
1484         GOTO(out, rc);
1485  out:
1486         return rc;
1487 }
1488
1489 static struct obd_ops echo_obd_ops = {
1490         .o_owner       = THIS_MODULE,
1491         .o_setup       = echo_client_setup,
1492         .o_cleanup     = echo_client_cleanup,
1493         .o_iocontrol   = echo_client_iocontrol,
1494         .o_connect     = echo_client_connect,
1495         .o_disconnect  = echo_client_disconnect
1496 };
1497
1498 int echo_client_init(void)
1499 {
1500         struct lprocfs_static_vars lvars;
1501
1502         lprocfs_init_vars(echo, &lvars);
1503         return class_register_type(&echo_obd_ops, NULL, lvars.module_vars,
1504                                    LUSTRE_ECHO_CLIENT_NAME, NULL);
1505 }
1506
1507 void echo_client_exit(void)
1508 {
1509         class_unregister_type(LUSTRE_ECHO_CLIENT_NAME);
1510 }