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 Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #define DEBUG_SUBSYSTEM S_ECHO
  23 #ifdef __KERNEL__
  24 #include <linux/version.h>
  25 #include <linux/module.h>
  26 #include <linux/fs.h>
  27 #include <linux/completion.h>
  28 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  29 #include <linux/iobuf.h>
  30 #endif
  31 #include <asm/div64.h>
  32 #else
  33 #include <liblustre.h>
  34 #endif
  35
  36 #include <linux/obd.h>
  37 #include <linux/obd_support.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/obd_echo.h>
  40 #include <linux/lustre_debug.h>
  41 #include <linux/lprocfs_status.h>
  42
  43 static obd_id last_object_id;
  44
  45 #if 0
  46 static void
  47 echo_printk_object (char *msg, struct ec_object *eco)
  48 {
  49         struct lov_stripe_md *lsm = eco->eco_lsm;
  50         int                   i;
  51
  52         printk (KERN_INFO "Lustre: %s: object %p: "LPX64", refs %d%s: "LPX64
  53                 "=%u!%u\n", msg, eco, eco->eco_id, eco->eco_refcount,
  54                 eco->eco_deleted ? "(deleted) " : "",
  55                 lsm->lsm_object_id, lsm->lsm_stripe_size,
  56                 lsm->lsm_stripe_count);
  57
  58         for (i = 0; i < lsm->lsm_stripe_count; i++)
  59                 printk (KERN_INFO "Lustre:   @%2u:"LPX64"\n",
  60                         lsm->lsm_oinfo[i].loi_ost_idx,
  61                         lsm->lsm_oinfo[i].loi_id);
  62 }
  63 #endif
  64
  65 static struct ec_object *
  66 echo_find_object_locked (struct obd_device *obd, obd_id id)
  67 {
  68         struct echo_client_obd *ec = &obd->u.echo_client;
  69         struct ec_object       *eco = NULL;
  70         struct list_head       *el;
  71
  72         list_for_each (el, &ec->ec_objects) {
  73                 eco = list_entry (el, struct ec_object, eco_obj_chain);
  74
  75                 if (eco->eco_id == id)
  76                         return (eco);
  77         }
  78         return (NULL);
  79 }
  80
  81 static int
  82 echo_copyout_lsm (struct lov_stripe_md *lsm, void *ulsm, int ulsm_nob)
  83 {
  84         int nob;
  85
  86         nob = offsetof (struct lov_stripe_md, lsm_oinfo[lsm->lsm_stripe_count]);
  87         if (nob > ulsm_nob)
  88                 return (-EINVAL);
  89
  90         if (copy_to_user (ulsm, lsm, nob))
  91                 return (-EFAULT);
  92
  93         return (0);
  94 }
  95
  96 static int
  97 echo_copyin_lsm (struct obd_device *obd, struct lov_stripe_md *lsm,
  98                  void *ulsm, int ulsm_nob)
  99 {
 100         struct echo_client_obd *ec = &obd->u.echo_client;
 101         int                     nob;
 102
 103         if (ulsm_nob < sizeof (*lsm))
 104                 return (-EINVAL);
 105
 106         if (copy_from_user (lsm, ulsm, sizeof (*lsm)))
 107                 return (-EFAULT);
 108
 109         nob = lsm->lsm_stripe_count * sizeof (lsm->lsm_oinfo[0]);
 110
 111         if (ulsm_nob < nob ||
 112             lsm->lsm_stripe_count > ec->ec_nstripes ||
 113             lsm->lsm_magic != LOV_MAGIC ||
 114             (lsm->lsm_stripe_size & (PAGE_SIZE - 1)) != 0 ||
 115             ((__u64)lsm->lsm_stripe_size * lsm->lsm_stripe_count > ~0UL))
 116                 return (-EINVAL);
 117
 118         if (copy_from_user(lsm->lsm_oinfo,
 119                            ((struct lov_stripe_md *)ulsm)->lsm_oinfo, nob))
 120                 return (-EFAULT);
 121
 122         return (0);
 123 }
 124
 125 static struct ec_object *
 126 echo_allocate_object (struct obd_device *obd)
 127 {
 128         struct echo_client_obd *ec = &obd->u.echo_client;
 129         struct ec_object       *eco;
 130         int rc;
 131
 132         OBD_ALLOC(eco, sizeof (*eco));
 133         if (eco == NULL)
 134                 return NULL;
 135
 136         rc = obd_alloc_memmd(ec->ec_exp, &eco->eco_lsm);
 137         if (rc < 0) {
 138                 OBD_FREE(eco, sizeof (*eco));
 139                 return NULL;
 140         }
 141
 142         eco->eco_device = obd;
 143         eco->eco_deleted = 0;
 144         eco->eco_refcount = 0;
 145         eco->eco_lsm->lsm_magic = LOV_MAGIC;
 146         /* leave stripe count 0 by default */
 147
 148         return (eco);
 149 }
 150
 151 static void
 152 echo_free_object (struct ec_object *eco)
 153 {
 154         struct obd_device      *obd = eco->eco_device;
 155         struct echo_client_obd *ec = &obd->u.echo_client;
 156
 157         LASSERT (eco->eco_refcount == 0);
 158         obd_free_memmd(ec->ec_exp, &eco->eco_lsm);
 159         OBD_FREE (eco, sizeof (*eco));
 160 }
 161
 162 static int echo_create_object(struct obd_device *obd, int on_target,
 163                               struct obdo *oa, void *ulsm, int ulsm_nob,
 164                               struct obd_trans_info *oti)
 165 {
 166         struct echo_client_obd *ec = &obd->u.echo_client;
 167         struct ec_object       *eco2;
 168         struct ec_object       *eco;
 169         struct lov_stripe_md   *lsm;
 170         int                     rc;
 171         int                     i, idx;
 172
 173         if ((oa->o_valid & OBD_MD_FLID) == 0 && /* no obj id */
 174             (on_target ||                       /* set_stripe */
 175              ec->ec_nstripes != 0)) {           /* LOV */
 176                 CERROR ("No valid oid\n");
 177                 return (-EINVAL);
 178         }
 179
 180         if (ulsm != NULL) {
 181                 eco = echo_allocate_object (obd);
 182                 if (eco == NULL)
 183                         return (-ENOMEM);
 184
 185                 lsm = eco->eco_lsm;
 186
 187                 rc = echo_copyin_lsm (obd, lsm, ulsm, ulsm_nob);
 188                 if (rc != 0)
 189                         goto failed;
 190
 191                 /* setup object ID here for !on_target and LOV hint */
 192                 if ((oa->o_valid & OBD_MD_FLID) != 0)
 193                         eco->eco_id = lsm->lsm_object_id = oa->o_id;
 194
 195                 if (lsm->lsm_stripe_count == 0)
 196                         lsm->lsm_stripe_count = ec->ec_nstripes;
 197
 198                 if (lsm->lsm_stripe_size == 0)
 199                         lsm->lsm_stripe_size = PAGE_SIZE;
 200
 201                 idx = ll_insecure_random_int();
 202
 203                 /* setup stripes: indices + default ids if required */
 204                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 205                         if (lsm->lsm_oinfo[i].loi_id == 0)
 206                                 lsm->lsm_oinfo[i].loi_id = lsm->lsm_object_id;
 207
 208                         lsm->lsm_oinfo[i].loi_ost_idx =
 209                                 (idx + i) % ec->ec_nstripes;
 210                 }
 211         } else {
 212                 OBD_ALLOC(eco, sizeof(*eco));
 213                 eco->eco_device = obd;
 214                 lsm = NULL;
 215         }
 216
 217         if (oa->o_id == 0)
 218                 oa->o_id = ++last_object_id;
 219
 220         if (on_target) {
 221                 /* XXX get some filter group constants */
 222                 oa->o_gr = 2;
 223                 oa->o_valid |= OBD_MD_FLGROUP;
 224                 rc = obd_create(ec->ec_exp, oa, &lsm, oti);
 225                 if (rc != 0)
 226                         goto failed;
 227
 228                 /* See what object ID we were given */
 229                 eco->eco_id = oa->o_id = lsm->lsm_object_id;
 230                 oa->o_valid |= OBD_MD_FLID;
 231
 232                 LASSERT(eco->eco_lsm == NULL || eco->eco_lsm == lsm);
 233                 eco->eco_lsm = lsm;
 234         }
 235
 236         spin_lock (&ec->ec_lock);
 237
 238         eco2 = echo_find_object_locked (obd, oa->o_id);
 239         if (eco2 != NULL) {                     /* conflict */
 240                 spin_unlock (&ec->ec_lock);
 241
 242                 CERROR ("Can't create object id "LPX64": id already exists%s\n",
 243                         oa->o_id, on_target ? " (undoing create)" : "");
 244
 245                 if (on_target)
 246                         obd_destroy(ec->ec_exp, oa, lsm, oti);
 247
 248                 rc = -EEXIST;
 249                 goto failed;
 250         }
 251
 252         list_add (&eco->eco_obj_chain, &ec->ec_objects);
 253         spin_unlock (&ec->ec_lock);
 254         CDEBUG (D_INFO,
 255                 "created %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 256                 eco, eco->eco_id,
 257                 eco->eco_lsm->lsm_stripe_size,
 258                 eco->eco_lsm->lsm_stripe_count,
 259                 eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 260                 eco->eco_refcount, eco->eco_deleted);
 261         return (0);
 262
 263  failed:
 264         echo_free_object (eco);
 265         return (rc);
 266 }
 267
 268 static int
 269 echo_get_object (struct ec_object **ecop, struct obd_device *obd,
 270                  struct obdo *oa)
 271 {
 272         struct echo_client_obd *ec = &obd->u.echo_client;
 273         struct ec_object       *eco;
 274         struct ec_object       *eco2;
 275         int                     rc;
 276
 277         if ((oa->o_valid & OBD_MD_FLID) == 0)
 278         {
 279                 CERROR ("No valid oid\n");
 280                 return (-EINVAL);
 281         }
 282
 283         spin_lock (&ec->ec_lock);
 284         eco = echo_find_object_locked (obd, oa->o_id);
 285         if (eco != NULL) {
 286                 if (eco->eco_deleted)           /* being deleted */
 287                         return (-EAGAIN);       /* (see comment in cleanup) */
 288
 289                 eco->eco_refcount++;
 290                 spin_unlock (&ec->ec_lock);
 291                 *ecop = eco;
 292                 CDEBUG (D_INFO,
 293                         "found %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 294                         eco, eco->eco_id,
 295                         eco->eco_lsm->lsm_stripe_size,
 296                         eco->eco_lsm->lsm_stripe_count,
 297                         eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 298                         eco->eco_refcount, eco->eco_deleted);
 299                 return (0);
 300         }
 301         spin_unlock (&ec->ec_lock);
 302
 303         if (ec->ec_nstripes != 0)               /* striping required */
 304                 return (-ENOENT);
 305
 306         eco = echo_allocate_object (obd);
 307         if (eco == NULL)
 308                 return (-ENOMEM);
 309
 310         eco->eco_id = eco->eco_lsm->lsm_object_id = oa->o_id;
 311
 312         spin_lock (&ec->ec_lock);
 313
 314         eco2 = echo_find_object_locked (obd, oa->o_id);
 315         if (eco2 == NULL) {                     /* didn't race */
 316                 list_add (&eco->eco_obj_chain, &ec->ec_objects);
 317                 spin_unlock (&ec->ec_lock);
 318                 eco->eco_refcount = 1;
 319                 *ecop = eco;
 320                 CDEBUG (D_INFO,
 321                         "created %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 322                         eco, eco->eco_id,
 323                         eco->eco_lsm->lsm_stripe_size,
 324                         eco->eco_lsm->lsm_stripe_count,
 325                         eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 326                         eco->eco_refcount, eco->eco_deleted);
 327                 return (0);
 328         }
 329
 330         if (eco2->eco_deleted)
 331                 rc = -EAGAIN;                   /* lose race */
 332         else {
 333                 eco2->eco_refcount++;           /* take existing */
 334                 *ecop = eco2;
 335                 rc = 0;
 336                 LASSERT (eco2->eco_id == eco2->eco_lsm->lsm_object_id);
 337                 CDEBUG (D_INFO,
 338                         "found(2) %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 339                         eco2, eco2->eco_id,
 340                         eco2->eco_lsm->lsm_stripe_size,
 341                         eco2->eco_lsm->lsm_stripe_count,
 342                         eco2->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 343                         eco2->eco_refcount, eco2->eco_deleted);
 344         }
 345
 346         spin_unlock (&ec->ec_lock);
 347
 348         echo_free_object (eco);
 349         return (rc);
 350 }
 351
 352 static void
 353 echo_put_object (struct ec_object *eco)
 354 {
 355         struct obd_device      *obd = eco->eco_device;
 356         struct echo_client_obd *ec = &obd->u.echo_client;
 357
 358         /* Release caller's ref on the object.
 359          * delete => mark for deletion when last ref goes
 360          */
 361
 362         spin_lock (&ec->ec_lock);
 363
 364         eco->eco_refcount--;
 365         LASSERT (eco->eco_refcount >= 0);
 366
 367         CDEBUG(D_INFO, "put %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 368                eco, eco->eco_id,
 369                eco->eco_lsm->lsm_stripe_size,
 370                eco->eco_lsm->lsm_stripe_count,
 371                eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 372                eco->eco_refcount, eco->eco_deleted);
 373
 374         if (eco->eco_refcount != 0 || !eco->eco_deleted) {
 375                 spin_unlock (&ec->ec_lock);
 376                 return;
 377         }
 378
 379         spin_unlock (&ec->ec_lock);
 380
 381         /* NB leave obj in the object list.  We must prevent anyone from
 382          * attempting to enqueue on this object number until we can be
 383          * sure there will be no more lock callbacks.
 384          */
 385         obd_cancel_unused(ec->ec_exp, eco->eco_lsm, 0, NULL);
 386
 387         /* now we can let it go */
 388         spin_lock (&ec->ec_lock);
 389         list_del (&eco->eco_obj_chain);
 390         spin_unlock (&ec->ec_lock);
 391
 392         LASSERT (eco->eco_refcount == 0);
 393
 394         echo_free_object (eco);
 395 }
 396
 397 static void
 398 echo_get_stripe_off_id (struct lov_stripe_md *lsm, obd_off *offp, obd_id *idp)
 399 {
 400         unsigned long stripe_count;
 401         unsigned long stripe_size;
 402         unsigned long width;
 403         unsigned long woffset;
 404         int           stripe_index;
 405         obd_off       offset;
 406
 407         if (lsm->lsm_stripe_count <= 1)
 408                 return;
 409
 410         offset       = *offp;
 411         stripe_size  = lsm->lsm_stripe_size;
 412         stripe_count = lsm->lsm_stripe_count;
 413
 414         /* width = # bytes in all stripes */
 415         width = stripe_size * stripe_count;
 416
 417         /* woffset = offset within a width; offset = whole number of widths */
 418         woffset = do_div (offset, width);
 419
 420         stripe_index = woffset / stripe_size;
 421
 422         *idp = lsm->lsm_oinfo[stripe_index].loi_id;
 423         *offp = offset * stripe_size + woffset % stripe_size;
 424 }
 425
 426 static void echo_page_debug_setup(struct lov_stripe_md *lsm,
 427                                   struct page *page, int rw, obd_id id,
 428                                   obd_off offset, obd_off count)
 429 {
 430         void *addr;
 431         obd_off stripe_off;
 432         obd_id stripe_id;
 433
 434         if (id == 0)
 435                 return;
 436
 437         addr = kmap(page);
 438
 439         if (rw == OBD_BRW_WRITE) {
 440                 stripe_off = offset;
 441                 stripe_id = id;
 442                 echo_get_stripe_off_id(lsm, &stripe_off, &stripe_id);
 443         } else {
 444                 stripe_off = 0xdeadbeef00c0ffeeULL;
 445                 stripe_id = 0xdeadbeef00c0ffeeULL;
 446         }
 447         page_debug_setup(addr, count, stripe_off, stripe_id);
 448
 449         kunmap(page);
 450 }
 451
 452 static int echo_page_debug_check(struct lov_stripe_md *lsm,
 453                                   struct page *page, obd_id id,
 454                                   obd_off offset, obd_off count)
 455 {
 456         obd_off stripe_off = offset;
 457         obd_id stripe_id = id;
 458         void *addr;
 459         int rc;
 460
 461         if (id == 0)
 462                 return 0;
 463
 464         addr = kmap(page);
 465         echo_get_stripe_off_id (lsm, &stripe_off, &stripe_id);
 466         rc = page_debug_check("test_brw", addr, count, stripe_off, stripe_id);
 467         kunmap(page);
 468         return rc;
 469 }
 470
 471 static int echo_client_kbrw(struct obd_device *obd, int rw, struct obdo *oa,
 472                             struct lov_stripe_md *lsm, obd_off offset,
 473                             obd_size count, struct obd_trans_info *oti)
 474 {
 475         struct echo_client_obd *ec = &obd->u.echo_client;
 476         obd_count               npages;
 477         struct brw_page        *pga;
 478         struct brw_page        *pgp;
 479         obd_off                 off;
 480         int                     i;
 481         int                     rc;
 482         int                     verify = 0;
 483         int                     gfp_mask;
 484
 485         /* oa_id  == 0    => speed test (no verification) else...
 486          * oa & 1         => use HIGHMEM
 487          */
 488         gfp_mask = ((oa->o_id & 1) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 489
 490         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 491
 492         if (count <= 0 ||
 493             (count & (PAGE_SIZE - 1)) != 0 ||
 494             (lsm != NULL &&
 495              lsm->lsm_object_id != oa->o_id))
 496                 return (-EINVAL);
 497
 498         /* XXX think again with misaligned I/O */
 499         npages = count >> PAGE_SHIFT;
 500
 501         OBD_ALLOC(pga, npages * sizeof(*pga));
 502         if (pga == NULL)
 503                 return (-ENOMEM);
 504
 505         for (i = 0, pgp = pga, off = offset;
 506              i < npages;
 507              i++, pgp++, off += PAGE_SIZE) {
 508
 509                 LASSERT (pgp->pg == NULL);      /* for cleanup */
 510
 511                 rc = -ENOMEM;
 512                 pgp->pg = alloc_pages (gfp_mask, 0);
 513                 if (pgp->pg == NULL)
 514                         goto out;
 515
 516                 pgp->count = PAGE_SIZE;
 517                 pgp->off = off;
 518                 pgp->flag = 0;
 519
 520                 echo_page_debug_setup(lsm, pgp->pg, rw, oa->o_id, off,
 521                                       pgp->count);
 522         }
 523
 524         rc = obd_brw(rw, ec->ec_exp, oa, lsm, npages, pga, oti);
 525
 526  out:
 527         if (rc == 0 && rw == OBD_BRW_READ)
 528                 verify = 1;
 529
 530         for (i = 0, pgp = pga; i < npages; i++, pgp++) {
 531                 if (pgp->pg == NULL)
 532                         continue;
 533
 534                 if (verify) {
 535                         int vrc;
 536                         vrc = echo_page_debug_check(lsm, pgp->pg, oa->o_id,
 537                                                     pgp->off, pgp->count);
 538                         if (vrc != 0 && rc == 0)
 539                                 rc = vrc;
 540                 }
 541                 __free_pages(pgp->pg, 0);
 542         }
 543         OBD_FREE(pga, npages * sizeof(*pga));
 544         return (rc);
 545 }
 546
 547 #ifdef __KERNEL__
 548 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 549 static int echo_client_ubrw(struct obd_device *obd, int rw,
 550                             struct obdo *oa, struct lov_stripe_md *lsm,
 551                             obd_off offset, obd_size count, char *buffer,
 552                             struct obd_trans_info *oti)
 553 {
 554         struct echo_client_obd *ec = &obd->u.echo_client;
 555         obd_count               npages;
 556         struct brw_page        *pga;
 557         struct brw_page        *pgp;
 558         obd_off                 off;
 559         struct kiobuf          *kiobuf;
 560         int                     i;
 561         int                     rc;
 562
 563         LASSERT (rw == OBD_BRW_WRITE ||
 564                  rw == OBD_BRW_READ);
 565
 566         /* NB: for now, only whole pages, page aligned */
 567
 568         if (count <= 0 ||
 569             ((long)buffer & (PAGE_SIZE - 1)) != 0 ||
 570             (count & (PAGE_SIZE - 1)) != 0 ||
 571             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 572                 return (-EINVAL);
 573
 574         /* XXX think again with misaligned I/O */
 575         npages = count >> PAGE_SHIFT;
 576
 577         OBD_ALLOC(pga, npages * sizeof(*pga));
 578         if (pga == NULL)
 579                 return (-ENOMEM);
 580
 581         rc = alloc_kiovec (1, &kiobuf);
 582         if (rc != 0)
 583                 goto out_1;
 584
 585         rc = map_user_kiobuf ((rw == OBD_BRW_READ) ? READ : WRITE,
 586                               kiobuf, (unsigned long)buffer, count);
 587         if (rc != 0)
 588                 goto out_2;
 589
 590         LASSERT (kiobuf->offset == 0);
 591         LASSERT (kiobuf->nr_pages == npages);
 592
 593         for (i = 0, off = offset, pgp = pga;
 594              i < npages;
 595              i++, off += PAGE_SIZE, pgp++) {
 596                 pgp->off = off;
 597                 pgp->pg = kiobuf->maplist[i];
 598                 pgp->count = PAGE_SIZE;
 599                 pgp->flag = 0;
 600         }
 601
 602         rc = obd_brw(rw, ec->ec_exp, oa, lsm, npages, pga, oti);
 603
 604         //        if (rw == OBD_BRW_READ)
 605         //                mark_dirty_kiobuf (kiobuf, count);
 606
 607         unmap_kiobuf (kiobuf);
 608  out_2:
 609         free_kiovec (1, &kiobuf);
 610  out_1:
 611         OBD_FREE(pga, npages * sizeof(*pga));
 612         return (rc);
 613 }
 614 #else
 615 static int echo_client_ubrw(struct obd_device *obd, int rw,
 616                             struct obdo *oa, struct lov_stripe_md *lsm,
 617                             obd_off offset, obd_size count, char *buffer,
 618                             struct obd_trans_info *oti)
 619 {
 620 #warning "echo_client_ubrw() needs to be ported on 2.6 yet"
 621         LBUG();
 622         return 0;
 623 }
 624 #endif
 625 #endif
 626
 627 struct echo_async_state;
 628
 629 #define EAP_MAGIC 79277927
 630 struct echo_async_page {
 631         int                     eap_magic;
 632         struct page             *eap_page;
 633         void                    *eap_cookie;
 634         obd_off                 eap_off;
 635         struct echo_async_state *eap_eas;
 636         struct list_head        eap_item;
 637 };
 638
 639 struct echo_async_state {
 640         spinlock_t              eas_lock;
 641         obd_off                 eas_next_offset;
 642         obd_off                 eas_end_offset;
 643         int                     eas_in_flight;
 644         int                     eas_rc;
 645         wait_queue_head_t       eas_waitq;
 646         struct list_head        eas_avail;
 647         struct obdo             eas_oa;
 648         struct lov_stripe_md    *eas_lsm;
 649 };
 650
 651 static int eas_should_wake(struct echo_async_state *eas)
 652 {
 653         unsigned long flags;
 654         int rc = 0;
 655         spin_lock_irqsave(&eas->eas_lock, flags);
 656         if (eas->eas_rc == 0 && !list_empty(&eas->eas_avail))
 657             rc = 1;
 658         spin_unlock_irqrestore(&eas->eas_lock, flags);
 659         return rc;
 660 };
 661
 662 struct echo_async_page *eap_from_cookie(void *cookie)
 663 {
 664         struct echo_async_page *eap = cookie;
 665         if (eap->eap_magic != EAP_MAGIC)
 666                 return ERR_PTR(-EINVAL);
 667         return eap;
 668 };
 669
 670 static int ec_ap_make_ready(void *data, int cmd)
 671 {
 672         /* our pages are issued ready */
 673         LBUG();
 674         return 0;
 675 }
 676 static int ec_ap_refresh_count(void *data, int cmd)
 677 {
 678         /* our pages are issued with a stable count */
 679         LBUG();
 680         return PAGE_SIZE;
 681 }
 682 static void ec_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 683 {
 684         struct echo_async_page *eap;
 685         eap = eap_from_cookie(data);
 686         if (IS_ERR(eap))
 687                 return;
 688
 689         memcpy(oa, &eap->eap_eas->eas_oa, sizeof(*oa));
 690 }
 691 static void ec_ap_completion(void *data, int cmd, int rc)
 692 {
 693         struct echo_async_page *eap = eap_from_cookie(data);
 694         struct echo_async_state *eas;
 695         unsigned long flags;
 696
 697         if (IS_ERR(eap))
 698                 return;
 699         eas = eap->eap_eas;
 700
 701         if (cmd == OBD_BRW_READ)
 702                 echo_page_debug_check(eas->eas_lsm, eap->eap_page,
 703                                       eas->eas_oa.o_id, eap->eap_off,
 704                                       PAGE_SIZE);
 705
 706         spin_lock_irqsave(&eas->eas_lock, flags);
 707         if (rc && !eas->eas_rc)
 708                 eas->eas_rc = rc;
 709         eas->eas_in_flight--;
 710         list_add(&eap->eap_item, &eas->eas_avail);
 711         wake_up(&eas->eas_waitq);
 712         spin_unlock_irqrestore(&eas->eas_lock, flags);
 713 }
 714
 715 static struct obd_async_page_ops ec_async_page_ops = {
 716         .ap_make_ready =        ec_ap_make_ready,
 717         .ap_refresh_count =     ec_ap_refresh_count,
 718         .ap_fill_obdo =         ec_ap_fill_obdo,
 719         .ap_completion =        ec_ap_completion,
 720 };
 721
 722 static int echo_client_async_page(struct obd_export *exp, int rw,
 723                                    struct obdo *oa, struct lov_stripe_md *lsm,
 724                                    obd_off offset, obd_size count,
 725                                    obd_size batching)
 726 {
 727         obd_count npages, i;
 728         struct echo_async_page *eap;
 729         struct echo_async_state eas;
 730         struct list_head *pos, *n;
 731         int rc = 0;
 732         unsigned long flags;
 733         LIST_HEAD(pages);
 734 #if 0
 735         int                     verify;
 736         int                     gfp_mask;
 737         /* oa_id  == 0    => speed test (no verification) else...
 738          * oa & 1         => use HIGHMEM
 739          */
 740         verify = (oa->o_id != 0);
 741         gfp_mask = ((oa->o_id & 1) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 742 #endif
 743
 744         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 745
 746         if (count <= 0 ||
 747             (count & (PAGE_SIZE - 1)) != 0 ||
 748             (lsm != NULL &&
 749              lsm->lsm_object_id != oa->o_id))
 750                 return (-EINVAL);
 751
 752         /* XXX think again with misaligned I/O */
 753         npages = batching >> PAGE_SHIFT;
 754
 755         memcpy(&eas.eas_oa, oa, sizeof(*oa));
 756         eas.eas_next_offset = offset;
 757         eas.eas_end_offset = offset + count;
 758         spin_lock_init(&eas.eas_lock);
 759         init_waitqueue_head(&eas.eas_waitq);
 760         eas.eas_in_flight = 0;
 761         eas.eas_rc = 0;
 762         eas.eas_lsm = lsm;
 763         INIT_LIST_HEAD(&eas.eas_avail);
 764
 765         /* prepare the group of pages that we're going to be keeping
 766          * in flight */
 767         for (i = 0; i < npages; i++) {
 768                 struct page *page = alloc_page(GFP_KERNEL);
 769                 if (page == NULL)
 770                         GOTO(out, rc = -ENOMEM);
 771
 772                 page->private = 0;
 773                 list_add_tail(&page->list, &pages);
 774
 775                 OBD_ALLOC(eap, sizeof(*eap));
 776                 if (eap == NULL)
 777                         GOTO(out, rc = -ENOMEM);
 778
 779                 eap->eap_magic = EAP_MAGIC;
 780                 eap->eap_page = page;
 781                 eap->eap_eas = &eas;
 782                 page->private = (unsigned long)eap;
 783                 list_add_tail(&eap->eap_item, &eas.eas_avail);
 784         }
 785
 786         /* first we spin queueing io and being woken by its completion */
 787         spin_lock_irqsave(&eas.eas_lock, flags);
 788         for(;;) {
 789                 int rc;
 790
 791                 /* sleep until we have a page to send */
 792                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 793                 rc = wait_event_interruptible(eas.eas_waitq,
 794                                               eas_should_wake(&eas));
 795                 spin_lock_irqsave(&eas.eas_lock, flags);
 796                 if (rc && !eas.eas_rc)
 797                         eas.eas_rc = rc;
 798                 if (eas.eas_rc)
 799                         break;
 800                 if (list_empty(&eas.eas_avail))
 801                         continue;
 802                 eap = list_entry(eas.eas_avail.next, struct echo_async_page,
 803                                  eap_item);
 804                 list_del(&eap->eap_item);
 805                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 806
 807                 /* unbind the eap from its old page offset */
 808                 if (eap->eap_cookie != NULL) {
 809                         obd_teardown_async_page(exp, lsm, NULL,
 810                                                 eap->eap_cookie);
 811                         eap->eap_cookie = NULL;
 812                 }
 813
 814                 eas.eas_next_offset += PAGE_SIZE;
 815                 eap->eap_off = eas.eas_next_offset;
 816
 817                 rc = obd_prep_async_page(exp, lsm, NULL, eap->eap_page,
 818                                          eap->eap_off, &ec_async_page_ops,
 819                                          eap, &eap->eap_cookie);
 820                 if (rc) {
 821                         spin_lock_irqsave(&eas.eas_lock, flags);
 822                         eas.eas_rc = rc;
 823                         break;
 824                 }
 825
 826                 if (rw == OBD_BRW_WRITE)
 827                         echo_page_debug_setup(lsm, eap->eap_page, rw, oa->o_id,
 828                                               eap->eap_off, PAGE_SIZE);
 829
 830                 /* always asserts urgent, which isn't quite right */
 831                 rc = obd_queue_async_io(exp, lsm, NULL, eap->eap_cookie,
 832                                         rw, 0, PAGE_SIZE, 0,
 833                                         ASYNC_READY | ASYNC_URGENT |
 834                                         ASYNC_COUNT_STABLE);
 835                 spin_lock_irqsave(&eas.eas_lock, flags);
 836                 if (rc && !eas.eas_rc) {
 837                         eas.eas_rc = rc;
 838                         break;
 839                 }
 840                 eas.eas_in_flight++;
 841                 if (eas.eas_next_offset == eas.eas_end_offset)
 842                         break;
 843         }
 844
 845         /* still hold the eas_lock here.. */
 846
 847         /* now we just spin waiting for all the rpcs to complete */
 848         while(eas.eas_in_flight) {
 849                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 850                 wait_event_interruptible(eas.eas_waitq,
 851                                          eas.eas_in_flight == 0);
 852                 spin_lock_irqsave(&eas.eas_lock, flags);
 853         }
 854         spin_unlock_irqrestore(&eas.eas_lock, flags);
 855
 856 out:
 857         list_for_each_safe(pos, n, &pages) {
 858                 struct page *page = list_entry(pos, struct page, list);
 859
 860                 list_del(&page->list);
 861                 if (page->private != 0) {
 862                         eap = (struct echo_async_page *)page->private;
 863                         if (eap->eap_cookie != NULL)
 864                                 obd_teardown_async_page(exp, lsm, NULL,
 865                                                         eap->eap_cookie);
 866                         OBD_FREE(eap, sizeof(*eap));
 867                 }
 868                 __free_page(page);
 869         }
 870
 871         RETURN(rc);
 872 }
 873
 874 static int echo_client_prep_commit(struct obd_export *exp, int rw,
 875                                    struct obdo *oa, struct lov_stripe_md *lsm,
 876                                    obd_off offset, obd_size count,
 877                                    obd_size batch, struct obd_trans_info *oti)
 878 {
 879         struct obd_ioobj ioo;
 880         struct niobuf_local *lnb;
 881         struct niobuf_remote *rnb;
 882         obd_off off;
 883         obd_size npages, tot_pages;
 884         int i, ret = 0, err = 0;
 885         ENTRY;
 886
 887         if (count <= 0 || (count & (PAGE_SIZE - 1)) != 0 ||
 888             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 889                 RETURN(-EINVAL);
 890
 891         npages = batch >> PAGE_SHIFT;
 892         tot_pages = count >> PAGE_SHIFT;
 893
 894         OBD_ALLOC(lnb, npages * sizeof(struct niobuf_local));
 895         OBD_ALLOC(rnb, npages * sizeof(struct niobuf_remote));
 896
 897         if (lnb == NULL || rnb == NULL)
 898                 GOTO(out, ret = -ENOMEM);
 899
 900         obdo_to_ioobj(oa, &ioo);
 901
 902         off = offset;
 903
 904         for(; tot_pages; tot_pages -= npages) {
 905                 if (tot_pages < npages)
 906                         npages = tot_pages;
 907
 908                 for (i = 0; i < npages; i++, off += PAGE_SIZE) {
 909                         rnb[i].offset = off;
 910                         rnb[i].len = PAGE_SIZE;
 911                 }
 912
 913                 /* XXX this can't be the best.. */
 914                 memset(oti, 0, sizeof(*oti));
 915                 ioo.ioo_bufcnt = npages;
 916
 917                 ret = obd_preprw(rw, exp, oa, 1, &ioo, npages, rnb, lnb, oti);
 918                 if (ret != 0)
 919                         GOTO(out, ret);
 920
 921                 for (i = 0; i < npages; i++) {
 922                         struct page *page = lnb[i].page;
 923
 924                         /* read past eof? */
 925                         if (page == NULL && lnb[i].rc == 0)
 926                                 continue;
 927
 928                         if (rw == OBD_BRW_WRITE)
 929                                 echo_page_debug_setup(lsm, page, rw, oa->o_id,
 930                                                       rnb[i].offset,
 931                                                       rnb[i].len);
 932                         else
 933                                 echo_page_debug_check(lsm, page, oa->o_id,
 934                                                       rnb[i].offset,
 935                                                       rnb[i].len);
 936                 }
 937
 938                 ret = obd_commitrw(rw, exp, oa, 1, &ioo, npages, lnb, oti);
 939                 if (ret != 0)
 940                         GOTO(out, ret);
 941                 if (err)
 942                         GOTO(out, ret = err);
 943         }
 944
 945 out:
 946         if (lnb)
 947                 OBD_FREE(lnb, npages * sizeof(struct niobuf_local));
 948         if (rnb)
 949                 OBD_FREE(rnb, npages * sizeof(struct niobuf_remote));
 950         RETURN(ret);
 951 }
 952
 953 int echo_client_brw_ioctl(int rw, struct obd_export *exp,
 954                           struct obd_ioctl_data *data)
 955 {
 956         struct obd_device *obd = class_exp2obd(exp);
 957         struct echo_client_obd *ec = &obd->u.echo_client;
 958         struct obd_trans_info dummy_oti;
 959         struct ec_object *eco;
 960         int rc;
 961         ENTRY;
 962
 963         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
 964         if (rc)
 965                 RETURN(rc);
 966
 967         memset(&dummy_oti, 0, sizeof(dummy_oti));
 968
 969         data->ioc_obdo1.o_valid &= ~OBD_MD_FLHANDLE;
 970         data->ioc_obdo1.o_valid |= OBD_MD_FLGROUP;
 971         data->ioc_obdo1.o_gr = 2;
 972
 973         switch((long)data->ioc_pbuf1) {
 974         case 1:
 975                 if (data->ioc_pbuf2 == NULL) { // NULL user data pointer
 976                         rc = echo_client_kbrw(obd, rw, &data->ioc_obdo1,
 977                                               eco->eco_lsm, data->ioc_offset,
 978                                               data->ioc_count, &dummy_oti);
 979                 } else {
 980 #ifdef __KERNEL__
 981                         rc = echo_client_ubrw(obd, rw, &data->ioc_obdo1,
 982                                               eco->eco_lsm, data->ioc_offset,
 983                                               data->ioc_count, data->ioc_pbuf2,
 984                                               &dummy_oti);
 985 #endif
 986                 }
 987                 break;
 988         case 2:
 989                 rc = echo_client_async_page(ec->ec_exp, rw, &data->ioc_obdo1,
 990                                            eco->eco_lsm, data->ioc_offset,
 991                                            data->ioc_count, data->ioc_plen1);
 992                 break;
 993         case 3:
 994                 rc = echo_client_prep_commit(ec->ec_exp, rw, &data->ioc_obdo1,
 995                                             eco->eco_lsm, data->ioc_offset,
 996                                             data->ioc_count, data->ioc_plen1,
 997                                             &dummy_oti);
 998                 break;
 999         default:
1000                 rc = -EINVAL;
1001         }
1002         echo_put_object(eco);
1003         RETURN(rc);
1004 }
1005
1006 static int
1007 echo_ldlm_callback (struct ldlm_lock *lock, struct ldlm_lock_desc *new,
1008                     void *data, int flag)
1009 {
1010         struct ec_object       *eco = (struct ec_object *)data;
1011         struct echo_client_obd *ec = &(eco->eco_device->u.echo_client);
1012         struct lustre_handle    lockh;
1013         struct list_head       *el;
1014         int                     found = 0;
1015         int                     rc;
1016
1017         ldlm_lock2handle (lock, &lockh);
1018
1019         /* #ifdef this out if we're not feeling paranoid */
1020         spin_lock (&ec->ec_lock);
1021         list_for_each (el, &ec->ec_objects) {
1022                 found = (eco == list_entry(el, struct ec_object,
1023                                            eco_obj_chain));
1024                 if (found)
1025                         break;
1026         }
1027         spin_unlock (&ec->ec_lock);
1028         LASSERT (found);
1029
1030         switch (flag) {
1031         case LDLM_CB_BLOCKING:
1032                 CDEBUG(D_INFO, "blocking callback on "LPX64", handle "LPX64"\n",
1033                        eco->eco_id, lockh.cookie);
1034                 rc = ldlm_cli_cancel (&lockh);
1035                 if (rc != ELDLM_OK)
1036                         CERROR ("ldlm_cli_cancel failed: %d\n", rc);
1037                 break;
1038
1039         case LDLM_CB_CANCELING:
1040                 CDEBUG(D_INFO, "cancel callback on "LPX64", handle "LPX64"\n",
1041                        eco->eco_id, lockh.cookie);
1042                 break;
1043
1044         default:
1045                 LBUG ();
1046         }
1047
1048         return (0);
1049 }
1050
1051 static int
1052 echo_client_enqueue(struct obd_export *exp, struct obdo *oa,
1053                     int mode, obd_off offset, obd_size nob)
1054 {
1055         struct obd_device      *obd = exp->exp_obd;
1056         struct echo_client_obd *ec = &obd->u.echo_client;
1057         struct lustre_handle   *ulh = obdo_handle (oa);
1058         struct ec_object       *eco;
1059         struct ec_lock         *ecl;
1060         int                     flags;
1061         int                     rc;
1062
1063         if (!(mode == LCK_PR || mode == LCK_PW))
1064                 return -EINVAL;
1065
1066         if ((offset & (PAGE_SIZE - 1)) != 0 ||
1067             (nob & (PAGE_SIZE - 1)) != 0)
1068                 return -EINVAL;
1069
1070         rc = echo_get_object (&eco, obd, oa);
1071         if (rc != 0)
1072                 return rc;
1073
1074         rc = -ENOMEM;
1075         OBD_ALLOC (ecl, sizeof (*ecl));
1076         if (ecl == NULL)
1077                 goto failed_0;
1078
1079         ecl->ecl_mode = mode;
1080         ecl->ecl_object = eco;
1081         ecl->ecl_policy.l_extent.start = offset;
1082         ecl->ecl_policy.l_extent.end =
1083                 (nob == 0) ? ((obd_off) -1) : (offset + nob - 1);
1084
1085         flags = 0;
1086         rc = obd_enqueue(ec->ec_exp, eco->eco_lsm, LDLM_EXTENT,
1087                          &ecl->ecl_policy, mode, &flags, echo_ldlm_callback,
1088                          ldlm_completion_ast, NULL, eco, sizeof(struct ost_lvb),
1089                          lustre_swab_ost_lvb, &ecl->ecl_lock_handle);
1090         if (rc != 0)
1091                 goto failed_1;
1092
1093         CDEBUG(D_INFO, "enqueue handle "LPX64"\n", ecl->ecl_lock_handle.cookie);
1094
1095         /* NB ecl takes object ref from echo_get_object() above */
1096         spin_lock(&ec->ec_lock);
1097
1098         list_add(&ecl->ecl_exp_chain, &exp->exp_ec_data.eced_locks);
1099         ulh->cookie = ecl->ecl_cookie = ec->ec_unique++;
1100
1101         spin_unlock(&ec->ec_lock);
1102
1103         oa->o_valid |= OBD_MD_FLHANDLE;
1104         return 0;
1105
1106  failed_1:
1107         OBD_FREE (ecl, sizeof (*ecl));
1108  failed_0:
1109         echo_put_object (eco);
1110         return (rc);
1111 }
1112
1113 static int
1114 echo_client_cancel(struct obd_export *exp, struct obdo *oa)
1115 {
1116         struct obd_device      *obd = exp->exp_obd;
1117         struct echo_client_obd *ec = &obd->u.echo_client;
1118         struct lustre_handle   *ulh = obdo_handle (oa);
1119         struct ec_lock         *ecl = NULL;
1120         int                     found = 0;
1121         struct list_head       *el;
1122         int                     rc;
1123
1124         if ((oa->o_valid & OBD_MD_FLHANDLE) == 0)
1125                 return -EINVAL;
1126
1127         spin_lock (&ec->ec_lock);
1128
1129         list_for_each (el, &exp->exp_ec_data.eced_locks) {
1130                 ecl = list_entry (el, struct ec_lock, ecl_exp_chain);
1131                 found = (ecl->ecl_cookie == ulh->cookie);
1132                 if (found) {
1133                         list_del (&ecl->ecl_exp_chain);
1134                         break;
1135                 }
1136         }
1137
1138         spin_unlock (&ec->ec_lock);
1139
1140         if (!found)
1141                 return (-ENOENT);
1142
1143         rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm, ecl->ecl_mode,
1144                         &ecl->ecl_lock_handle);
1145
1146         echo_put_object (ecl->ecl_object);
1147         OBD_FREE (ecl, sizeof (*ecl));
1148
1149         return rc;
1150 }
1151
1152 static int
1153 echo_client_iocontrol(unsigned int cmd, struct obd_export *exp,
1154                       int len, void *karg, void *uarg)
1155 {
1156         struct obd_device      *obd;
1157         struct echo_client_obd *ec;
1158         struct ec_object       *eco;
1159         struct obd_ioctl_data  *data = karg;
1160         struct obd_trans_info   dummy_oti;
1161         struct oti_req_ack_lock *ack_lock;
1162         struct obdo            *oa;
1163         int                     rw = OBD_BRW_READ;
1164         int                     rc = 0;
1165         int                     i;
1166         ENTRY;
1167
1168         memset(&dummy_oti, 0, sizeof(dummy_oti));
1169
1170         obd = exp->exp_obd;
1171         ec = &obd->u.echo_client;
1172
1173         switch (cmd) {
1174         case OBD_IOC_CREATE:                    /* may create echo object */
1175                 if (!capable (CAP_SYS_ADMIN))
1176                         GOTO (out, rc = -EPERM);
1177
1178                 rc = echo_create_object (obd, 1, &data->ioc_obdo1,
1179                                          data->ioc_pbuf1, data->ioc_plen1,
1180                                          &dummy_oti);
1181                 GOTO(out, rc);
1182
1183         case OBD_IOC_DESTROY:
1184                 if (!capable (CAP_SYS_ADMIN))
1185                         GOTO (out, rc = -EPERM);
1186
1187                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1188                 if (rc == 0) {
1189                         oa = &data->ioc_obdo1;
1190                         oa->o_gr = 2;
1191                         oa->o_valid |= OBD_MD_FLGROUP;
1192                         rc = obd_destroy(ec->ec_exp, oa, eco->eco_lsm,
1193                                          &dummy_oti);
1194                         if (rc == 0)
1195                                 eco->eco_deleted = 1;
1196                         echo_put_object(eco);
1197                 }
1198                 GOTO(out, rc);
1199
1200         case OBD_IOC_GETATTR:
1201                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1202                 if (rc == 0) {
1203                         rc = obd_getattr(ec->ec_exp, &data->ioc_obdo1,
1204                                          eco->eco_lsm);
1205                         echo_put_object(eco);
1206                 }
1207                 GOTO(out, rc);
1208
1209         case OBD_IOC_SETATTR:
1210                 if (!capable (CAP_SYS_ADMIN))
1211                         GOTO (out, rc = -EPERM);
1212
1213                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1214                 if (rc == 0) {
1215                         rc = obd_setattr(ec->ec_exp, &data->ioc_obdo1,
1216                                          eco->eco_lsm, NULL);
1217                         echo_put_object(eco);
1218                 }
1219                 GOTO(out, rc);
1220
1221         case OBD_IOC_BRW_WRITE:
1222                 if (!capable (CAP_SYS_ADMIN))
1223                         GOTO (out, rc = -EPERM);
1224
1225                 rw = OBD_BRW_WRITE;
1226                 /* fall through */
1227         case OBD_IOC_BRW_READ:
1228                 rc = echo_client_brw_ioctl(rw, exp, data);
1229                 GOTO(out, rc);
1230
1231         case ECHO_IOC_GET_STRIPE:
1232                 rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1233                 if (rc == 0) {
1234                         rc = echo_copyout_lsm(eco->eco_lsm, data->ioc_pbuf1,
1235                                               data->ioc_plen1);
1236                         echo_put_object(eco);
1237                 }
1238                 GOTO(out, rc);
1239
1240         case ECHO_IOC_SET_STRIPE:
1241                 if (!capable (CAP_SYS_ADMIN))
1242                         GOTO (out, rc = -EPERM);
1243
1244                 if (data->ioc_pbuf1 == NULL) {  /* unset */
1245                         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1246                         if (rc == 0) {
1247                                 eco->eco_deleted = 1;
1248                                 echo_put_object(eco);
1249                         }
1250                 } else {
1251                         rc = echo_create_object(obd, 0, &data->ioc_obdo1,
1252                                                 data->ioc_pbuf1,
1253                                                 data->ioc_plen1, &dummy_oti);
1254                 }
1255                 GOTO (out, rc);
1256
1257         case ECHO_IOC_ENQUEUE:
1258                 if (!capable (CAP_SYS_ADMIN))
1259                         GOTO (out, rc = -EPERM);
1260
1261                 rc = echo_client_enqueue(exp, &data->ioc_obdo1,
1262                                          data->ioc_conn1, /* lock mode */
1263                                    data->ioc_offset, data->ioc_count);/*extent*/
1264                 GOTO (out, rc);
1265
1266         case ECHO_IOC_CANCEL:
1267                 rc = echo_client_cancel(exp, &data->ioc_obdo1);
1268                 GOTO (out, rc);
1269
1270         default:
1271                 CERROR ("echo_ioctl(): unrecognised ioctl %#x\n", cmd);
1272                 GOTO (out, rc = -ENOTTY);
1273         }
1274
1275         EXIT;
1276  out:
1277
1278         /* XXX this should be in a helper also called by target_send_reply */
1279         for (ack_lock = dummy_oti.oti_ack_locks, i = 0; i < 4;
1280              i++, ack_lock++) {
1281                 if (!ack_lock->mode)
1282                         break;
1283                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1284         }
1285
1286         return rc;
1287 }
1288
1289 static int
1290 echo_client_setup(struct obd_device *obddev, obd_count len, void *buf)
1291 {
1292         struct lustre_cfg* lcfg = buf;
1293         struct echo_client_obd *ec = &obddev->u.echo_client;
1294         struct obd_device *tgt;
1295         struct lustre_handle conn = {0, };
1296         struct obd_uuid echo_uuid = { "ECHO_UUID" };
1297         int rc;
1298         ENTRY;
1299
1300         if (lcfg->lcfg_inllen1 < 1) {
1301                 CERROR("requires a TARGET OBD name\n");
1302                 RETURN(-EINVAL);
1303         }
1304
1305         tgt = class_name2obd(lcfg->lcfg_inlbuf1);
1306         if (!tgt || !tgt->obd_attached || !tgt->obd_set_up) {
1307                 CERROR("device not attached or not set up (%s)\n",
1308                        lcfg->lcfg_inlbuf1);
1309                 RETURN(-EINVAL);
1310         }
1311
1312         spin_lock_init (&ec->ec_lock);
1313         INIT_LIST_HEAD (&ec->ec_objects);
1314         ec->ec_unique = 0;
1315
1316         rc = obd_connect(&conn, tgt, &echo_uuid);
1317         if (rc) {
1318                 CERROR("fail to connect to device %s\n", lcfg->lcfg_inlbuf1);
1319                 return (rc);
1320         }
1321         ec->ec_exp = class_conn2export(&conn);
1322
1323         RETURN(rc);
1324 }
1325
1326 static int echo_client_cleanup(struct obd_device *obddev, int flags)
1327 {
1328         struct list_head       *el;
1329         struct ec_object       *eco;
1330         struct echo_client_obd *ec = &obddev->u.echo_client;
1331         int rc;
1332         ENTRY;
1333
1334         if (!list_empty(&obddev->obd_exports)) {
1335                 CERROR("still has clients!\n");
1336                 RETURN(-EBUSY);
1337         }
1338
1339         /* XXX assuming sole access */
1340         while (!list_empty(&ec->ec_objects)) {
1341                 el = ec->ec_objects.next;
1342                 eco = list_entry(el, struct ec_object, eco_obj_chain);
1343
1344                 LASSERT(eco->eco_refcount == 0);
1345                 eco->eco_refcount = 1;
1346                 eco->eco_deleted = 1;
1347                 echo_put_object(eco);
1348         }
1349
1350         rc = obd_disconnect(ec->ec_exp, 0);
1351         if (rc != 0)
1352                 CERROR("fail to disconnect device: %d\n", rc);
1353
1354         RETURN(rc);
1355 }
1356
1357 static int echo_client_connect(struct lustre_handle *conn,
1358                                struct obd_device *src, struct obd_uuid *cluuid)
1359 {
1360         struct obd_export *exp;
1361         int                rc;
1362
1363         rc = class_connect(conn, src, cluuid);
1364         if (rc == 0) {
1365                 exp = class_conn2export(conn);
1366                 INIT_LIST_HEAD(&exp->exp_ec_data.eced_locks);
1367                 class_export_put(exp);
1368         }
1369
1370         RETURN (rc);
1371 }
1372
1373 static int echo_client_disconnect(struct obd_export *exp, int flags)
1374 {
1375         struct obd_device      *obd;
1376         struct echo_client_obd *ec;
1377         struct ec_lock         *ecl;
1378         int                     rc;
1379         ENTRY;
1380
1381         if (exp == NULL)
1382                 GOTO(out, rc = -EINVAL);
1383
1384         obd = exp->exp_obd;
1385         ec = &obd->u.echo_client;
1386
1387         /* no more contention on export's lock list */
1388         while (!list_empty (&exp->exp_ec_data.eced_locks)) {
1389                 ecl = list_entry (exp->exp_ec_data.eced_locks.next,
1390                                   struct ec_lock, ecl_exp_chain);
1391                 list_del (&ecl->ecl_exp_chain);
1392
1393                 rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm,
1394                                  ecl->ecl_mode, &ecl->ecl_lock_handle);
1395
1396                 CDEBUG (D_INFO, "Cancel lock on object "LPX64" on disconnect "
1397                         "(%d)\n", ecl->ecl_object->eco_id, rc);
1398
1399                 echo_put_object (ecl->ecl_object);
1400                 OBD_FREE (ecl, sizeof (*ecl));
1401         }
1402
1403         rc = class_disconnect(exp, 0);
1404         GOTO(out, rc);
1405  out:
1406         return rc;
1407 }
1408
1409 static struct obd_ops echo_obd_ops = {
1410         o_owner:       THIS_MODULE,
1411         o_setup:       echo_client_setup,
1412         o_cleanup:     echo_client_cleanup,
1413         o_iocontrol:   echo_client_iocontrol,
1414         o_connect:     echo_client_connect,
1415         o_disconnect:  echo_client_disconnect
1416 };
1417
1418 int echo_client_init(void)
1419 {
1420         struct lprocfs_static_vars lvars;
1421
1422         lprocfs_init_vars(echo, &lvars);
1423         return class_register_type(&echo_obd_ops, lvars.module_vars,
1424                                    OBD_ECHO_CLIENT_DEVICENAME);
1425 }
1426
1427 void echo_client_exit(void)
1428 {
1429         class_unregister_type(OBD_ECHO_CLIENT_DEVICENAME);
1430 }