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 #include <linux/smp_lock.h>
  33 #else
  34 #include <liblustre.h>
  35 #endif
  36
  37 #include <linux/obd.h>
  38 #include <linux/obd_support.h>
  39 #include <linux/obd_class.h>
  40 #include <linux/obd_echo.h>
  41 #include <linux/lustre_debug.h>
  42 #include <linux/lprocfs_status.h>
  43
  44 static obd_id last_object_id;
  45
  46 #if 0
  47 static void
  48 echo_printk_object (char *msg, struct ec_object *eco)
  49 {
  50         struct lov_stripe_md *lsm = eco->eco_lsm;
  51         int                   i;
  52
  53         printk (KERN_INFO "Lustre: %s: object %p: "LPX64", refs %d%s: "LPX64
  54                 "=%u!%u\n", msg, eco, eco->eco_id, eco->eco_refcount,
  55                 eco->eco_deleted ? "(deleted) " : "",
  56                 lsm->lsm_object_id, lsm->lsm_stripe_size,
  57                 lsm->lsm_stripe_count);
  58
  59         for (i = 0; i < lsm->lsm_stripe_count; i++)
  60                 printk (KERN_INFO "Lustre:   @%2u:"LPX64"\n",
  61                         lsm->lsm_oinfo[i].loi_ost_idx,
  62                         lsm->lsm_oinfo[i].loi_id);
  63 }
  64 #endif
  65
  66 static struct ec_object *
  67 echo_find_object_locked (struct obd_device *obd, obd_id id)
  68 {
  69         struct echo_client_obd *ec = &obd->u.echocli;
  70         struct ec_object       *eco = NULL;
  71         struct list_head       *el;
  72
  73         list_for_each (el, &ec->ec_objects) {
  74                 eco = list_entry (el, struct ec_object, eco_obj_chain);
  75
  76                 if (eco->eco_id == id)
  77                         return (eco);
  78         }
  79         return (NULL);
  80 }
  81
  82 static int
  83 echo_copyout_lsm (struct lov_stripe_md *lsm, void *ulsm, int ulsm_nob)
  84 {
  85         int nob;
  86
  87         nob = offsetof (struct lov_stripe_md, lsm_oinfo[lsm->lsm_stripe_count]);
  88         if (nob > ulsm_nob)
  89                 return (-EINVAL);
  90
  91         if (copy_to_user (ulsm, lsm, nob))
  92                 return (-EFAULT);
  93
  94         return (0);
  95 }
  96
  97 static int
  98 echo_copyin_lsm (struct obd_device *obd, struct lov_stripe_md *lsm,
  99                  void *ulsm, int ulsm_nob)
 100 {
 101         struct echo_client_obd *ec = &obd->u.echocli;
 102         int                     nob;
 103
 104         if (ulsm_nob < sizeof (*lsm))
 105                 return (-EINVAL);
 106
 107         if (copy_from_user (lsm, ulsm, sizeof (*lsm)))
 108                 return (-EFAULT);
 109
 110         nob = lsm->lsm_stripe_count * sizeof (lsm->lsm_oinfo[0]);
 111
 112         if (ulsm_nob < nob ||
 113             lsm->lsm_stripe_count > ec->ec_nstripes ||
 114             lsm->lsm_magic != LOV_MAGIC ||
 115             (lsm->lsm_stripe_size & (PAGE_SIZE - 1)) != 0 ||
 116             ((__u64)lsm->lsm_stripe_size * lsm->lsm_stripe_count > ~0UL))
 117                 return (-EINVAL);
 118
 119         if (copy_from_user(lsm->lsm_oinfo,
 120                            ((struct lov_stripe_md *)ulsm)->lsm_oinfo, nob))
 121                 return (-EFAULT);
 122
 123         return (0);
 124 }
 125
 126 static struct ec_object *
 127 echo_allocate_object (struct obd_device *obd)
 128 {
 129         struct echo_client_obd *ec = &obd->u.echocli;
 130         struct ec_object       *eco;
 131         int rc;
 132
 133         OBD_ALLOC(eco, sizeof (*eco));
 134         if (eco == NULL)
 135                 return NULL;
 136
 137         rc = obd_alloc_memmd(ec->ec_exp, &eco->eco_lsm);
 138         if (rc < 0) {
 139                 OBD_FREE(eco, sizeof (*eco));
 140                 return NULL;
 141         }
 142
 143         eco->eco_device = obd;
 144         eco->eco_deleted = 0;
 145         eco->eco_refcount = 0;
 146         eco->eco_lsm->lsm_magic = LOV_MAGIC;
 147         /* leave stripe count 0 by default */
 148
 149         return (eco);
 150 }
 151
 152 static void
 153 echo_free_object (struct ec_object *eco)
 154 {
 155         struct obd_device      *obd = eco->eco_device;
 156         struct echo_client_obd *ec = &obd->u.echocli;
 157
 158         LASSERT (eco->eco_refcount == 0);
 159         obd_free_memmd(ec->ec_exp, &eco->eco_lsm);
 160         OBD_FREE (eco, sizeof (*eco));
 161 }
 162
 163 static int echo_create_object(struct obd_device *obd, int on_target,
 164                               struct obdo *oa, void *ulsm, int ulsm_nob,
 165                               struct obd_trans_info *oti)
 166 {
 167         struct echo_client_obd *ec = &obd->u.echocli;
 168         struct ec_object       *eco2;
 169         struct ec_object       *eco;
 170         struct lov_stripe_md   *lsm;
 171         int                     rc;
 172         int                     i, idx;
 173
 174         if ((oa->o_valid & OBD_MD_FLID) == 0 && /* no obj id */
 175             (on_target ||                       /* set_stripe */
 176              ec->ec_nstripes != 0)) {           /* LOV */
 177                 CERROR ("No valid oid\n");
 178                 return (-EINVAL);
 179         }
 180
 181         if (ulsm != NULL) {
 182                 eco = echo_allocate_object (obd);
 183                 if (eco == NULL)
 184                         return (-ENOMEM);
 185
 186                 lsm = eco->eco_lsm;
 187
 188                 rc = echo_copyin_lsm (obd, lsm, ulsm, ulsm_nob);
 189                 if (rc != 0)
 190                         goto failed;
 191
 192                 /* setup object ID here for !on_target and LOV hint */
 193                 if ((oa->o_valid & OBD_MD_FLID) != 0)
 194                         eco->eco_id = lsm->lsm_object_id = oa->o_id;
 195
 196                 if (lsm->lsm_stripe_count == 0)
 197                         lsm->lsm_stripe_count = ec->ec_nstripes;
 198
 199                 if (lsm->lsm_stripe_size == 0)
 200                         lsm->lsm_stripe_size = PAGE_SIZE;
 201
 202                 idx = ll_insecure_random_int();
 203
 204                 /* setup stripes: indices + default ids if required */
 205                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 206                         if (lsm->lsm_oinfo[i].loi_id == 0)
 207                                 lsm->lsm_oinfo[i].loi_id = lsm->lsm_object_id;
 208
 209                         lsm->lsm_oinfo[i].loi_ost_idx =
 210                                 (idx + i) % ec->ec_nstripes;
 211                 }
 212         } else {
 213                 OBD_ALLOC(eco, sizeof(*eco));
 214                 eco->eco_device = obd;
 215                 lsm = NULL;
 216         }
 217
 218         if (oa->o_id == 0)
 219                 oa->o_id = ++last_object_id;
 220
 221         if (on_target) {
 222                 oa->o_gr = FILTER_GROUP_ECHO;
 223                 oa->o_valid |= OBD_MD_FLGROUP;
 224                 rc = obd_create(ec->ec_exp, oa, NULL, 0, &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.echocli;
 273         struct ec_object       *eco;
 274         struct ec_object       *eco2;
 275         int                     rc;
 276
 277         if ((oa->o_valid & OBD_MD_FLID) == 0 ||
 278             oa->o_id == 0)                      /* disallow use of object id 0 */
 279         {
 280                 CERROR ("No valid oid\n");
 281                 return (-EINVAL);
 282         }
 283
 284         spin_lock (&ec->ec_lock);
 285         eco = echo_find_object_locked (obd, oa->o_id);
 286         if (eco != NULL) {
 287                 if (eco->eco_deleted)           /* being deleted */
 288                         return (-EAGAIN);       /* (see comment in cleanup) */
 289
 290                 eco->eco_refcount++;
 291                 spin_unlock (&ec->ec_lock);
 292                 *ecop = eco;
 293                 CDEBUG (D_INFO,
 294                         "found %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 295                         eco, eco->eco_id,
 296                         eco->eco_lsm->lsm_stripe_size,
 297                         eco->eco_lsm->lsm_stripe_count,
 298                         eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 299                         eco->eco_refcount, eco->eco_deleted);
 300                 return (0);
 301         }
 302         spin_unlock (&ec->ec_lock);
 303
 304         if (ec->ec_nstripes != 0)               /* striping required */
 305                 return (-ENOENT);
 306
 307         eco = echo_allocate_object (obd);
 308         if (eco == NULL)
 309                 return (-ENOMEM);
 310
 311         eco->eco_id = eco->eco_lsm->lsm_object_id = oa->o_id;
 312
 313         spin_lock (&ec->ec_lock);
 314
 315         eco2 = echo_find_object_locked (obd, oa->o_id);
 316         if (eco2 == NULL) {                     /* didn't race */
 317                 list_add (&eco->eco_obj_chain, &ec->ec_objects);
 318                 spin_unlock (&ec->ec_lock);
 319                 eco->eco_refcount = 1;
 320                 *ecop = eco;
 321                 CDEBUG (D_INFO,
 322                         "created %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 323                         eco, eco->eco_id,
 324                         eco->eco_lsm->lsm_stripe_size,
 325                         eco->eco_lsm->lsm_stripe_count,
 326                         eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 327                         eco->eco_refcount, eco->eco_deleted);
 328                 return (0);
 329         }
 330
 331         if (eco2->eco_deleted)
 332                 rc = -EAGAIN;                   /* lose race */
 333         else {
 334                 eco2->eco_refcount++;           /* take existing */
 335                 *ecop = eco2;
 336                 rc = 0;
 337                 LASSERT (eco2->eco_id == eco2->eco_lsm->lsm_object_id);
 338                 CDEBUG (D_INFO,
 339                         "found(2) %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 340                         eco2, eco2->eco_id,
 341                         eco2->eco_lsm->lsm_stripe_size,
 342                         eco2->eco_lsm->lsm_stripe_count,
 343                         eco2->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 344                         eco2->eco_refcount, eco2->eco_deleted);
 345         }
 346
 347         spin_unlock (&ec->ec_lock);
 348
 349         echo_free_object (eco);
 350         return (rc);
 351 }
 352
 353 static void
 354 echo_put_object (struct ec_object *eco)
 355 {
 356         struct obd_device      *obd = eco->eco_device;
 357         struct echo_client_obd *ec = &obd->u.echocli;
 358
 359         /* Release caller's ref on the object.
 360          * delete => mark for deletion when last ref goes
 361          */
 362
 363         spin_lock (&ec->ec_lock);
 364
 365         eco->eco_refcount--;
 366         LASSERT (eco->eco_refcount >= 0);
 367
 368         CDEBUG(D_INFO, "put %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 369                eco, eco->eco_id,
 370                eco->eco_lsm->lsm_stripe_size,
 371                eco->eco_lsm->lsm_stripe_count,
 372                eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 373                eco->eco_refcount, eco->eco_deleted);
 374
 375         if (eco->eco_refcount != 0 || !eco->eco_deleted) {
 376                 spin_unlock (&ec->ec_lock);
 377                 return;
 378         }
 379
 380         spin_unlock (&ec->ec_lock);
 381
 382         /* NB leave obj in the object list.  We must prevent anyone from
 383          * attempting to enqueue on this object number until we can be
 384          * sure there will be no more lock callbacks.
 385          */
 386         obd_cancel_unused(ec->ec_exp, eco->eco_lsm, 0, NULL);
 387
 388         /* now we can let it go */
 389         spin_lock (&ec->ec_lock);
 390         list_del (&eco->eco_obj_chain);
 391         spin_unlock (&ec->ec_lock);
 392
 393         LASSERT (eco->eco_refcount == 0);
 394
 395         echo_free_object (eco);
 396 }
 397
 398 static void
 399 echo_get_stripe_off_id (struct lov_stripe_md *lsm, obd_off *offp, obd_id *idp)
 400 {
 401         unsigned long stripe_count;
 402         unsigned long stripe_size;
 403         unsigned long width;
 404         unsigned long woffset;
 405         int           stripe_index;
 406         obd_off       offset;
 407
 408         if (lsm->lsm_stripe_count <= 1)
 409                 return;
 410
 411         offset       = *offp;
 412         stripe_size  = lsm->lsm_stripe_size;
 413         stripe_count = lsm->lsm_stripe_count;
 414
 415         /* width = # bytes in all stripes */
 416         width = stripe_size * stripe_count;
 417
 418         /* woffset = offset within a width; offset = whole number of widths */
 419         woffset = do_div (offset, width);
 420
 421         stripe_index = woffset / stripe_size;
 422
 423         *idp = lsm->lsm_oinfo[stripe_index].loi_id;
 424         *offp = offset * stripe_size + woffset % stripe_size;
 425 }
 426
 427 static void
 428 echo_client_page_debug_setup(struct lov_stripe_md *lsm,
 429                              struct page *page, int rw, obd_id id,
 430                              obd_off offset, obd_off count)
 431 {
 432         char    *addr;
 433         obd_off  stripe_off;
 434         obd_id   stripe_id;
 435         int      delta;
 436
 437         /* no partial pages on the client */
 438         LASSERT(count == PAGE_SIZE);
 439
 440         addr = kmap(page);
 441
 442         for (delta = 0; delta < PAGE_SIZE; delta += OBD_ECHO_BLOCK_SIZE) {
 443                 if (rw == OBD_BRW_WRITE) {
 444                         stripe_off = offset + delta;
 445                         stripe_id = id;
 446                         echo_get_stripe_off_id(lsm, &stripe_off, &stripe_id);
 447                 } else {
 448                         stripe_off = 0xdeadbeef00c0ffeeULL;
 449                         stripe_id = 0xdeadbeef00c0ffeeULL;
 450                 }
 451                 block_debug_setup(addr + delta, OBD_ECHO_BLOCK_SIZE,
 452                                   stripe_off, stripe_id);
 453         }
 454
 455         kunmap(page);
 456 }
 457
 458 static int
 459 echo_client_page_debug_check(struct lov_stripe_md *lsm,
 460                              struct page *page, obd_id id,
 461                              obd_off offset, obd_off count)
 462 {
 463         obd_off stripe_off;
 464         obd_id  stripe_id;
 465         char   *addr;
 466         int     delta;
 467         int     rc;
 468         int     rc2;
 469
 470         /* no partial pages on the client */
 471         LASSERT(count == PAGE_SIZE);
 472
 473         addr = kmap(page);
 474
 475         for (rc = delta = 0; delta < PAGE_SIZE; delta += OBD_ECHO_BLOCK_SIZE) {
 476                 stripe_off = offset + delta;
 477                 stripe_id = id;
 478                 echo_get_stripe_off_id (lsm, &stripe_off, &stripe_id);
 479
 480                 rc2 = block_debug_check("test_brw",
 481                                         addr + delta, OBD_ECHO_BLOCK_SIZE,
 482                                         stripe_off, stripe_id);
 483                 if (rc2 != 0) {
 484                         CERROR ("Error in echo object "LPX64"\n", id);
 485                         rc = rc2;
 486                 }
 487         }
 488
 489         kunmap(page);
 490         return rc;
 491 }
 492
 493 static int echo_client_kbrw(struct obd_device *obd, int rw, struct obdo *oa,
 494                             struct lov_stripe_md *lsm, obd_off offset,
 495                             obd_size count, struct obd_trans_info *oti)
 496 {
 497         struct echo_client_obd *ec = &obd->u.echocli;
 498         obd_count               npages;
 499         struct brw_page        *pga;
 500         struct brw_page        *pgp;
 501         obd_off                 off;
 502         int                     i;
 503         int                     rc;
 504         int                     verify = 0;
 505         int                     gfp_mask;
 506
 507         verify = ((oa->o_id) != ECHO_PERSISTENT_OBJID &&
 508                   (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 509                   (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
 510
 511         gfp_mask = ((oa->o_id & 2) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 512
 513         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 514
 515         if (count <= 0 ||
 516             (count & (PAGE_SIZE - 1)) != 0 ||
 517             (lsm != NULL &&
 518              lsm->lsm_object_id != oa->o_id))
 519                 return (-EINVAL);
 520
 521         /* XXX think again with misaligned I/O */
 522         npages = count >> PAGE_SHIFT;
 523
 524         OBD_ALLOC(pga, npages * sizeof(*pga));
 525         if (pga == NULL)
 526                 return (-ENOMEM);
 527
 528         for (i = 0, pgp = pga, off = offset;
 529              i < npages;
 530              i++, pgp++, off += PAGE_SIZE) {
 531
 532                 LASSERT (pgp->pg == NULL);      /* for cleanup */
 533
 534                 rc = -ENOMEM;
 535                 pgp->pg = alloc_pages (gfp_mask, 0);
 536                 if (pgp->pg == NULL)
 537                         goto out;
 538
 539                 pgp->count = PAGE_SIZE;
 540                 pgp->disk_offset = pgp->page_offset = off;
 541                 pgp->flag = 0;
 542
 543                 if (verify)
 544                         echo_client_page_debug_setup(lsm, pgp->pg, rw,
 545                                                      oa->o_id, off, pgp->count);
 546         }
 547
 548         rc = obd_brw(rw, ec->ec_exp, oa, lsm, npages, pga, oti);
 549
 550  out:
 551         if (rc != 0 || rw != OBD_BRW_READ)
 552                 verify = 0;
 553
 554         for (i = 0, pgp = pga; i < npages; i++, pgp++) {
 555                 if (pgp->pg == NULL)
 556                         continue;
 557
 558                 if (verify) {
 559                         int vrc;
 560                         vrc = echo_client_page_debug_check(lsm, pgp->pg, oa->o_id,
 561                                                            pgp->page_offset,
 562                                                            pgp->count);
 563                         if (vrc != 0 && rc == 0)
 564                                 rc = vrc;
 565                 }
 566                 __free_pages(pgp->pg, 0);
 567         }
 568         OBD_FREE(pga, npages * sizeof(*pga));
 569         return (rc);
 570 }
 571
 572 #ifdef __KERNEL__
 573 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 574 static int echo_client_ubrw(struct obd_device *obd, int rw,
 575                             struct obdo *oa, struct lov_stripe_md *lsm,
 576                             obd_off offset, obd_size count, char *buffer,
 577                             struct obd_trans_info *oti)
 578 {
 579         struct echo_client_obd *ec = &obd->u.echocli;
 580         obd_count               npages;
 581         struct brw_page        *pga;
 582         struct brw_page        *pgp;
 583         obd_off                 off;
 584         struct kiobuf          *kiobuf;
 585         int                     i;
 586         int                     rc;
 587
 588         LASSERT (rw == OBD_BRW_WRITE ||
 589                  rw == OBD_BRW_READ);
 590
 591         /* NB: for now, only whole pages, page aligned */
 592
 593         if (count <= 0 ||
 594             ((long)buffer & (PAGE_SIZE - 1)) != 0 ||
 595             (count & (PAGE_SIZE - 1)) != 0 ||
 596             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 597                 return (-EINVAL);
 598
 599         /* XXX think again with misaligned I/O */
 600         npages = count >> PAGE_SHIFT;
 601
 602         OBD_ALLOC(pga, npages * sizeof(*pga));
 603         if (pga == NULL)
 604                 return (-ENOMEM);
 605
 606         rc = alloc_kiovec (1, &kiobuf);
 607         if (rc != 0)
 608                 goto out_1;
 609
 610         rc = map_user_kiobuf ((rw == OBD_BRW_READ) ? READ : WRITE,
 611                               kiobuf, (unsigned long)buffer, count);
 612         if (rc != 0)
 613                 goto out_2;
 614
 615         LASSERT (kiobuf->offset == 0);
 616         LASSERT (kiobuf->nr_pages == npages);
 617
 618         for (i = 0, off = offset, pgp = pga;
 619              i < npages;
 620              i++, off += PAGE_SIZE, pgp++) {
 621                 pgp->disk_offset = pgp->page_offset = off;
 622                 pgp->pg = kiobuf->maplist[i];
 623                 pgp->count = PAGE_SIZE;
 624                 pgp->flag = 0;
 625         }
 626
 627         rc = obd_brw(rw, ec->ec_exp, oa, lsm, npages, pga, oti);
 628
 629         //        if (rw == OBD_BRW_READ)
 630         //                mark_dirty_kiobuf (kiobuf, count);
 631
 632         unmap_kiobuf (kiobuf);
 633  out_2:
 634         free_kiovec (1, &kiobuf);
 635  out_1:
 636         OBD_FREE(pga, npages * sizeof(*pga));
 637         return (rc);
 638 }
 639 #else
 640 static int echo_client_ubrw(struct obd_device *obd, int rw,
 641                             struct obdo *oa, struct lov_stripe_md *lsm,
 642                             obd_off offset, obd_size count, char *buffer,
 643                             struct obd_trans_info *oti)
 644 {
 645 #warning "echo_client_ubrw() needs to be ported on 2.6 yet"
 646         LBUG();
 647         return 0;
 648 }
 649 #endif
 650 #endif
 651
 652 struct echo_async_state;
 653
 654 #define EAP_MAGIC 79277927
 655 struct echo_async_page {
 656         int                     eap_magic;
 657         struct page             *eap_page;
 658         void                    *eap_cookie;
 659         obd_off                 eap_off;
 660         struct echo_async_state *eap_eas;
 661         struct list_head        eap_item;
 662 };
 663
 664 #define EAP_FROM_COOKIE(c)                                                      \
 665         (LASSERT(((struct echo_async_page *)(c))->eap_magic == EAP_MAGIC),      \
 666          (struct echo_async_page *)(c))
 667
 668 struct echo_async_state {
 669         spinlock_t              eas_lock;
 670         obd_off                 eas_next_offset;
 671         obd_off                 eas_end_offset;
 672         int                     eas_in_flight;
 673         int                     eas_rc;
 674         wait_queue_head_t       eas_waitq;
 675         struct list_head        eas_avail;
 676         struct obdo             eas_oa;
 677         struct lov_stripe_md    *eas_lsm;
 678 };
 679
 680 static int eas_should_wake(struct echo_async_state *eas)
 681 {
 682         unsigned long flags;
 683         int rc = 0;
 684         spin_lock_irqsave(&eas->eas_lock, flags);
 685         if (eas->eas_rc == 0 && !list_empty(&eas->eas_avail))
 686             rc = 1;
 687         spin_unlock_irqrestore(&eas->eas_lock, flags);
 688         return rc;
 689 };
 690
 691 static int ec_ap_make_ready(void *data, int cmd)
 692 {
 693         /* our pages are issued ready */
 694         LBUG();
 695         return 0;
 696 }
 697 static int ec_ap_refresh_count(void *data, int cmd)
 698 {
 699         /* our pages are issued with a stable count */
 700         LBUG();
 701         return PAGE_SIZE;
 702 }
 703 static void ec_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 704 {
 705         struct echo_async_page *eap = EAP_FROM_COOKIE(data);
 706
 707         memcpy(oa, &eap->eap_eas->eas_oa, sizeof(*oa));
 708 }
 709
 710 static void ec_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 711 {
 712         struct echo_async_page *eap = EAP_FROM_COOKIE(data);
 713         struct echo_async_state *eas;
 714         unsigned long flags;
 715
 716         eas = eap->eap_eas;
 717
 718         if (cmd == OBD_BRW_READ &&
 719             eas->eas_oa.o_id != ECHO_PERSISTENT_OBJID &&
 720             (eas->eas_oa.o_valid & OBD_MD_FLFLAGS) != 0 &&
 721             (eas->eas_oa.o_flags & OBD_FL_DEBUG_CHECK) != 0)
 722                 echo_client_page_debug_check(eas->eas_lsm, eap->eap_page,
 723                                              eas->eas_oa.o_id, eap->eap_off,
 724                                              PAGE_SIZE);
 725
 726         spin_lock_irqsave(&eas->eas_lock, flags);
 727         if (rc && !eas->eas_rc)
 728                 eas->eas_rc = rc;
 729         eas->eas_in_flight--;
 730         list_add(&eap->eap_item, &eas->eas_avail);
 731         wake_up(&eas->eas_waitq);
 732         spin_unlock_irqrestore(&eas->eas_lock, flags);
 733 }
 734
 735 static struct obd_async_page_ops ec_async_page_ops = {
 736         .ap_make_ready =        ec_ap_make_ready,
 737         .ap_refresh_count =     ec_ap_refresh_count,
 738         .ap_fill_obdo =         ec_ap_fill_obdo,
 739         .ap_completion =        ec_ap_completion,
 740 };
 741
 742 static int echo_client_async_page(struct obd_export *exp, int rw,
 743                                    struct obdo *oa, struct lov_stripe_md *lsm,
 744                                    obd_off offset, obd_size count,
 745                                    obd_size batching)
 746 {
 747         obd_count npages, i;
 748         struct echo_async_page *eap;
 749         struct echo_async_state eas;
 750         struct list_head *pos, *n;
 751         int rc = 0;
 752         unsigned long flags;
 753         LIST_HEAD(pages);
 754 #if 0
 755         int                     verify;
 756         int                     gfp_mask;
 757         /* oa_id  == 0    => speed test (no verification) else...
 758          * oa & 1         => use HIGHMEM
 759          */
 760         verify = (oa->o_id != 0);
 761         gfp_mask = ((oa->o_id & 1) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 762 #endif
 763
 764         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 765
 766         if (count <= 0 ||
 767             (count & (PAGE_SIZE - 1)) != 0 ||
 768             (lsm != NULL &&
 769              lsm->lsm_object_id != oa->o_id))
 770                 return (-EINVAL);
 771
 772         /* XXX think again with misaligned I/O */
 773         npages = batching >> PAGE_SHIFT;
 774
 775         memcpy(&eas.eas_oa, oa, sizeof(*oa));
 776         eas.eas_next_offset = offset;
 777         eas.eas_end_offset = offset + count;
 778         spin_lock_init(&eas.eas_lock);
 779         init_waitqueue_head(&eas.eas_waitq);
 780         eas.eas_in_flight = 0;
 781         eas.eas_rc = 0;
 782         eas.eas_lsm = lsm;
 783         INIT_LIST_HEAD(&eas.eas_avail);
 784
 785         /* prepare the group of pages that we're going to be keeping
 786          * in flight */
 787         for (i = 0; i < npages; i++) {
 788                 struct page *page = alloc_page(GFP_KERNEL);
 789                 if (page == NULL)
 790                         GOTO(out, rc = -ENOMEM);
 791
 792                 page->private = 0;
 793                 list_add_tail(&PAGE_LIST(page), &pages);
 794
 795                 OBD_ALLOC(eap, sizeof(*eap));
 796                 if (eap == NULL)
 797                         GOTO(out, rc = -ENOMEM);
 798
 799                 eap->eap_magic = EAP_MAGIC;
 800                 eap->eap_page = page;
 801                 eap->eap_eas = &eas;
 802                 page->private = (unsigned long)eap;
 803                 list_add_tail(&eap->eap_item, &eas.eas_avail);
 804         }
 805
 806         /* first we spin queueing io and being woken by its completion */
 807         spin_lock_irqsave(&eas.eas_lock, flags);
 808         for(;;) {
 809                 int rc;
 810
 811                 /* sleep until we have a page to send */
 812                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 813                 rc = wait_event_interruptible(eas.eas_waitq,
 814                                               eas_should_wake(&eas));
 815                 spin_lock_irqsave(&eas.eas_lock, flags);
 816                 if (rc && !eas.eas_rc)
 817                         eas.eas_rc = rc;
 818                 if (eas.eas_rc)
 819                         break;
 820                 if (list_empty(&eas.eas_avail))
 821                         continue;
 822                 eap = list_entry(eas.eas_avail.next, struct echo_async_page,
 823                                  eap_item);
 824                 list_del(&eap->eap_item);
 825                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 826
 827                 /* unbind the eap from its old page offset */
 828                 if (eap->eap_cookie != NULL) {
 829                         obd_teardown_async_page(exp, lsm, NULL,
 830                                                 eap->eap_cookie);
 831                         eap->eap_cookie = NULL;
 832                 }
 833
 834                 eas.eas_next_offset += PAGE_SIZE;
 835                 eap->eap_off = eas.eas_next_offset;
 836
 837                 rc = obd_prep_async_page(exp, lsm, NULL, eap->eap_page,
 838                                          eap->eap_off, &ec_async_page_ops,
 839                                          eap, &eap->eap_cookie);
 840                 if (rc) {
 841                         spin_lock_irqsave(&eas.eas_lock, flags);
 842                         eas.eas_rc = rc;
 843                         break;
 844                 }
 845
 846                 if (oa->o_id != ECHO_PERSISTENT_OBJID &&
 847                     (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 848                     (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0)
 849                         echo_client_page_debug_setup(lsm, eap->eap_page, rw,
 850                                                      oa->o_id,
 851                                                      eap->eap_off, PAGE_SIZE);
 852
 853                 /* always asserts urgent, which isn't quite right */
 854                 rc = obd_queue_async_io(exp, lsm, NULL, eap->eap_cookie,
 855                                         rw, 0, PAGE_SIZE, 0,
 856                                         ASYNC_READY | ASYNC_URGENT |
 857                                         ASYNC_COUNT_STABLE);
 858                 spin_lock_irqsave(&eas.eas_lock, flags);
 859                 if (rc && !eas.eas_rc) {
 860                         eas.eas_rc = rc;
 861                         break;
 862                 }
 863                 eas.eas_in_flight++;
 864                 if (eas.eas_next_offset == eas.eas_end_offset)
 865                         break;
 866         }
 867
 868         /* still hold the eas_lock here.. */
 869
 870         /* now we just spin waiting for all the rpcs to complete */
 871         while(eas.eas_in_flight) {
 872                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 873                 wait_event_interruptible(eas.eas_waitq,
 874                                          eas.eas_in_flight == 0);
 875                 spin_lock_irqsave(&eas.eas_lock, flags);
 876         }
 877         spin_unlock_irqrestore(&eas.eas_lock, flags);
 878
 879 out:
 880         list_for_each_safe(pos, n, &pages) {
 881                 struct page *page = list_entry(pos, struct page,
 882                                                PAGE_LIST_ENTRY);
 883
 884                 list_del(&PAGE_LIST(page));
 885                 if (page->private != 0) {
 886                         eap = (struct echo_async_page *)page->private;
 887                         if (eap->eap_cookie != NULL)
 888                                 obd_teardown_async_page(exp, lsm, NULL,
 889                                                         eap->eap_cookie);
 890                         OBD_FREE(eap, sizeof(*eap));
 891                 }
 892                 __free_page(page);
 893         }
 894
 895         RETURN(rc);
 896 }
 897
 898 static int echo_client_prep_commit(struct obd_export *exp, int rw,
 899                                    struct obdo *oa, struct lov_stripe_md *lsm,
 900                                    obd_off offset, obd_size count,
 901                                    obd_size batch, struct obd_trans_info *oti)
 902 {
 903         struct obd_ioobj ioo;
 904         struct niobuf_local *lnb;
 905         struct niobuf_remote *rnb;
 906         obd_off off;
 907         obd_size npages, tot_pages;
 908         int i, ret = 0;
 909         ENTRY;
 910
 911         if (count <= 0 || (count & (PAGE_SIZE - 1)) != 0 ||
 912             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 913                 RETURN(-EINVAL);
 914
 915         npages = batch >> PAGE_SHIFT;
 916         tot_pages = count >> PAGE_SHIFT;
 917
 918         OBD_ALLOC(lnb, npages * sizeof(struct niobuf_local));
 919         OBD_ALLOC(rnb, npages * sizeof(struct niobuf_remote));
 920
 921         if (lnb == NULL || rnb == NULL)
 922                 GOTO(out, ret = -ENOMEM);
 923
 924         obdo_to_ioobj(oa, &ioo);
 925
 926         off = offset;
 927
 928         for(; tot_pages; tot_pages -= npages) {
 929                 if (tot_pages < npages)
 930                         npages = tot_pages;
 931
 932                 for (i = 0; i < npages; i++, off += PAGE_SIZE) {
 933                         rnb[i].offset = off;
 934                         rnb[i].len = PAGE_SIZE;
 935                 }
 936
 937                 /* XXX this can't be the best.. */
 938                 memset(oti, 0, sizeof(*oti));
 939                 ioo.ioo_bufcnt = npages;
 940
 941                 ret = obd_preprw(rw, exp, oa, 1, &ioo, npages, rnb, lnb,
 942                                  oti, NULL);
 943                 if (ret != 0)
 944                         GOTO(out, ret);
 945
 946                 for (i = 0; i < npages; i++) {
 947                         struct page *page = lnb[i].page;
 948
 949                         /* read past eof? */
 950                         if (page == NULL && lnb[i].rc == 0)
 951                                 continue;
 952
 953                         if (oa->o_id == ECHO_PERSISTENT_OBJID ||
 954                             (oa->o_valid & OBD_MD_FLFLAGS) == 0 ||
 955                             (oa->o_flags & OBD_FL_DEBUG_CHECK) == 0)
 956                                 continue;
 957
 958
 959                         if (rw == OBD_BRW_WRITE)
 960                                 echo_client_page_debug_setup(lsm, page, rw,
 961                                                              oa->o_id,
 962                                                              rnb[i].offset,
 963                                                              rnb[i].len);
 964                         else
 965                                 echo_client_page_debug_check(lsm, page,
 966                                                              oa->o_id,
 967                                                              rnb[i].offset,
 968                                                              rnb[i].len);
 969                 }
 970
 971                 ret = obd_commitrw(rw, exp, oa, 1, &ioo, npages, lnb, oti, ret);
 972                 if (ret != 0)
 973                         GOTO(out, ret);
 974         }
 975
 976 out:
 977         if (lnb)
 978                 OBD_FREE(lnb, npages * sizeof(struct niobuf_local));
 979         if (rnb)
 980                 OBD_FREE(rnb, npages * sizeof(struct niobuf_remote));
 981         RETURN(ret);
 982 }
 983
 984 int echo_client_brw_ioctl(int rw, struct obd_export *exp,
 985                           struct obd_ioctl_data *data)
 986 {
 987         struct obd_device *obd = class_exp2obd(exp);
 988         struct echo_client_obd *ec = &obd->u.echocli;
 989         struct obd_trans_info dummy_oti;
 990         struct ec_object *eco;
 991         int rc;
 992         ENTRY;
 993
 994         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
 995         if (rc)
 996                 RETURN(rc);
 997
 998         memset(&dummy_oti, 0, sizeof(dummy_oti));
 999
1000         data->ioc_obdo1.o_valid &= ~OBD_MD_FLHANDLE;
1001         data->ioc_obdo1.o_valid |= OBD_MD_FLGROUP;
1002         data->ioc_obdo1.o_gr = FILTER_GROUP_ECHO;
1003
1004         switch((long)data->ioc_pbuf1) {
1005         case 1:
1006                 if (data->ioc_pbuf2 == NULL) { // NULL user data pointer
1007                         rc = echo_client_kbrw(obd, rw, &data->ioc_obdo1,
1008                                               eco->eco_lsm, data->ioc_offset,
1009                                               data->ioc_count, &dummy_oti);
1010                 } else {
1011 #ifdef __KERNEL__
1012                         rc = echo_client_ubrw(obd, rw, &data->ioc_obdo1,
1013                                               eco->eco_lsm, data->ioc_offset,
1014                                               data->ioc_count, data->ioc_pbuf2,
1015                                               &dummy_oti);
1016 #endif
1017                 }
1018                 break;
1019         case 2:
1020                 rc = echo_client_async_page(ec->ec_exp, rw, &data->ioc_obdo1,
1021                                            eco->eco_lsm, data->ioc_offset,
1022                                            data->ioc_count, data->ioc_plen1);
1023                 break;
1024         case 3:
1025                 rc = echo_client_prep_commit(ec->ec_exp, rw, &data->ioc_obdo1,
1026                                             eco->eco_lsm, data->ioc_offset,
1027                                             data->ioc_count, data->ioc_plen1,
1028                                             &dummy_oti);
1029                 break;
1030         default:
1031                 rc = -EINVAL;
1032         }
1033         echo_put_object(eco);
1034         RETURN(rc);
1035 }
1036
1037 static int
1038 echo_ldlm_callback (struct ldlm_lock *lock, struct ldlm_lock_desc *new,
1039                     void *data, int flag)
1040 {
1041         struct ec_object       *eco = (struct ec_object *)data;
1042         struct echo_client_obd *ec = &(eco->eco_device->u.echocli);
1043         struct lustre_handle    lockh;
1044         struct list_head       *el;
1045         int                     found = 0;
1046         int                     rc;
1047
1048         ldlm_lock2handle (lock, &lockh);
1049
1050         /* #ifdef this out if we're not feeling paranoid */
1051         spin_lock (&ec->ec_lock);
1052         list_for_each (el, &ec->ec_objects) {
1053                 found = (eco == list_entry(el, struct ec_object,
1054                                            eco_obj_chain));
1055                 if (found)
1056                         break;
1057         }
1058         spin_unlock (&ec->ec_lock);
1059         LASSERT (found);
1060
1061         switch (flag) {
1062         case LDLM_CB_BLOCKING:
1063                 CDEBUG(D_INFO, "blocking callback on "LPX64", handle "LPX64"\n",
1064                        eco->eco_id, lockh.cookie);
1065                 rc = ldlm_cli_cancel (&lockh);
1066                 if (rc != ELDLM_OK)
1067                         CERROR ("ldlm_cli_cancel failed: %d\n", rc);
1068                 break;
1069
1070         case LDLM_CB_CANCELING:
1071                 CDEBUG(D_INFO, "cancel callback on "LPX64", handle "LPX64"\n",
1072                        eco->eco_id, lockh.cookie);
1073                 break;
1074
1075         default:
1076                 LBUG ();
1077         }
1078
1079         return (0);
1080 }
1081
1082 static int
1083 echo_client_enqueue(struct obd_export *exp, struct obdo *oa,
1084                     int mode, obd_off offset, obd_size nob)
1085 {
1086         struct obd_device      *obd = exp->exp_obd;
1087         struct echo_client_obd *ec = &obd->u.echocli;
1088         struct lustre_handle   *ulh = obdo_handle (oa);
1089         struct ec_object       *eco;
1090         struct ec_lock         *ecl;
1091         int                     flags;
1092         int                     rc;
1093
1094         if (!(mode == LCK_PR || mode == LCK_PW))
1095                 return -EINVAL;
1096
1097         if ((offset & (PAGE_SIZE - 1)) != 0 ||
1098             (nob & (PAGE_SIZE - 1)) != 0)
1099                 return -EINVAL;
1100
1101         rc = echo_get_object (&eco, obd, oa);
1102         if (rc != 0)
1103                 return rc;
1104
1105         rc = -ENOMEM;
1106         OBD_ALLOC (ecl, sizeof (*ecl));
1107         if (ecl == NULL)
1108                 goto failed_0;
1109
1110         ecl->ecl_mode = mode;
1111         ecl->ecl_object = eco;
1112         ecl->ecl_policy.l_extent.start = offset;
1113         ecl->ecl_policy.l_extent.end =
1114                 (nob == 0) ? ((obd_off) -1) : (offset + nob - 1);
1115
1116         flags = 0;
1117         rc = obd_enqueue(ec->ec_exp, eco->eco_lsm, LDLM_EXTENT,
1118                          &ecl->ecl_policy, mode, &flags, echo_ldlm_callback,
1119                          ldlm_completion_ast, NULL, eco, sizeof(struct ost_lvb),
1120                          lustre_swab_ost_lvb, &ecl->ecl_lock_handle);
1121         if (rc != 0)
1122                 goto failed_1;
1123
1124         CDEBUG(D_INFO, "enqueue handle "LPX64"\n", ecl->ecl_lock_handle.cookie);
1125
1126         /* NB ecl takes object ref from echo_get_object() above */
1127         spin_lock(&ec->ec_lock);
1128
1129         list_add(&ecl->ecl_exp_chain, &exp->exp_ec_data.eced_locks);
1130         ulh->cookie = ecl->ecl_cookie = ec->ec_unique++;
1131
1132         spin_unlock(&ec->ec_lock);
1133
1134         oa->o_valid |= OBD_MD_FLHANDLE;
1135         return 0;
1136
1137  failed_1:
1138         OBD_FREE (ecl, sizeof (*ecl));
1139  failed_0:
1140         echo_put_object (eco);
1141         return (rc);
1142 }
1143
1144 static int
1145 echo_client_cancel(struct obd_export *exp, struct obdo *oa)
1146 {
1147         struct obd_device      *obd = exp->exp_obd;
1148         struct echo_client_obd *ec = &obd->u.echocli;
1149         struct lustre_handle   *ulh = obdo_handle (oa);
1150         struct ec_lock         *ecl = NULL;
1151         int                     found = 0;
1152         struct list_head       *el;
1153         int                     rc;
1154
1155         if ((oa->o_valid & OBD_MD_FLHANDLE) == 0)
1156                 return -EINVAL;
1157
1158         spin_lock (&ec->ec_lock);
1159
1160         list_for_each (el, &exp->exp_ec_data.eced_locks) {
1161                 ecl = list_entry (el, struct ec_lock, ecl_exp_chain);
1162                 found = (ecl->ecl_cookie == ulh->cookie);
1163                 if (found) {
1164                         list_del (&ecl->ecl_exp_chain);
1165                         break;
1166                 }
1167         }
1168
1169         spin_unlock (&ec->ec_lock);
1170
1171         if (!found)
1172                 return (-ENOENT);
1173
1174         rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm, ecl->ecl_mode,
1175                         &ecl->ecl_lock_handle);
1176
1177         echo_put_object (ecl->ecl_object);
1178         OBD_FREE (ecl, sizeof (*ecl));
1179
1180         return rc;
1181 }
1182
1183 static int
1184 echo_client_iocontrol(unsigned int cmd, struct obd_export *exp,
1185                       int len, void *karg, void *uarg)
1186 {
1187         struct obd_device      *obd;
1188         struct echo_client_obd *ec;
1189         struct ec_object       *eco;
1190         struct obd_ioctl_data  *data = karg;
1191         struct obd_trans_info   dummy_oti;
1192         struct oti_req_ack_lock *ack_lock;
1193         struct obdo            *oa;
1194         int                     rw = OBD_BRW_READ;
1195         int                     rc = 0;
1196         int                     i;
1197         ENTRY;
1198
1199         unlock_kernel();
1200
1201         memset(&dummy_oti, 0, sizeof(dummy_oti));
1202
1203         obd = exp->exp_obd;
1204         ec = &obd->u.echocli;
1205
1206         switch (cmd) {
1207         case OBD_IOC_CREATE:                    /* may create echo object */
1208                 if (!capable (CAP_SYS_ADMIN))
1209                         GOTO (out, rc = -EPERM);
1210
1211                 rc = echo_create_object (obd, 1, &data->ioc_obdo1,
1212                                          data->ioc_pbuf1, data->ioc_plen1,
1213                                          &dummy_oti);
1214                 GOTO(out, rc);
1215
1216         case OBD_IOC_DESTROY:
1217                 if (!capable (CAP_SYS_ADMIN))
1218                         GOTO (out, rc = -EPERM);
1219
1220                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1221                 if (rc == 0) {
1222                         oa = &data->ioc_obdo1;
1223                         oa->o_gr = FILTER_GROUP_ECHO;
1224                         oa->o_valid |= OBD_MD_FLGROUP;
1225                         rc = obd_destroy(ec->ec_exp, oa, eco->eco_lsm,
1226                                          &dummy_oti);
1227                         if (rc == 0)
1228                                 eco->eco_deleted = 1;
1229                         echo_put_object(eco);
1230                 }
1231                 GOTO(out, rc);
1232
1233         case OBD_IOC_GETATTR:
1234                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1235                 if (rc == 0) {
1236                         rc = obd_getattr(ec->ec_exp, &data->ioc_obdo1,
1237                                          eco->eco_lsm);
1238                         echo_put_object(eco);
1239                 }
1240                 GOTO(out, rc);
1241
1242         case OBD_IOC_SETATTR:
1243                 if (!capable (CAP_SYS_ADMIN))
1244                         GOTO (out, rc = -EPERM);
1245
1246                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1247                 if (rc == 0) {
1248                         rc = obd_setattr(ec->ec_exp, &data->ioc_obdo1,
1249                                          eco->eco_lsm, NULL);
1250                         echo_put_object(eco);
1251                 }
1252                 GOTO(out, rc);
1253
1254         case OBD_IOC_BRW_WRITE:
1255                 if (!capable (CAP_SYS_ADMIN))
1256                         GOTO (out, rc = -EPERM);
1257
1258                 rw = OBD_BRW_WRITE;
1259                 /* fall through */
1260         case OBD_IOC_BRW_READ:
1261                 rc = echo_client_brw_ioctl(rw, exp, data);
1262                 GOTO(out, rc);
1263
1264         case ECHO_IOC_GET_STRIPE:
1265                 rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1266                 if (rc == 0) {
1267                         rc = echo_copyout_lsm(eco->eco_lsm, data->ioc_pbuf1,
1268                                               data->ioc_plen1);
1269                         echo_put_object(eco);
1270                 }
1271                 GOTO(out, rc);
1272
1273         case ECHO_IOC_SET_STRIPE:
1274                 if (!capable (CAP_SYS_ADMIN))
1275                         GOTO (out, rc = -EPERM);
1276
1277                 if (data->ioc_pbuf1 == NULL) {  /* unset */
1278                         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1279                         if (rc == 0) {
1280                                 eco->eco_deleted = 1;
1281                                 echo_put_object(eco);
1282                         }
1283                 } else {
1284                         rc = echo_create_object(obd, 0, &data->ioc_obdo1,
1285                                                 data->ioc_pbuf1,
1286                                                 data->ioc_plen1, &dummy_oti);
1287                 }
1288                 GOTO (out, rc);
1289
1290         case ECHO_IOC_ENQUEUE:
1291                 if (!capable (CAP_SYS_ADMIN))
1292                         GOTO (out, rc = -EPERM);
1293
1294                 rc = echo_client_enqueue(exp, &data->ioc_obdo1,
1295                                          data->ioc_conn1, /* lock mode */
1296                                    data->ioc_offset, data->ioc_count);/*extent*/
1297                 GOTO (out, rc);
1298
1299         case ECHO_IOC_CANCEL:
1300                 rc = echo_client_cancel(exp, &data->ioc_obdo1);
1301                 GOTO (out, rc);
1302
1303         default:
1304                 CERROR ("echo_ioctl(): unrecognised ioctl %#x\n", cmd);
1305                 GOTO (out, rc = -ENOTTY);
1306         }
1307
1308         EXIT;
1309  out:
1310
1311         /* XXX this should be in a helper also called by target_send_reply */
1312         for (ack_lock = dummy_oti.oti_ack_locks, i = 0; i < 4;
1313              i++, ack_lock++) {
1314                 if (!ack_lock->mode)
1315                         break;
1316                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1317         }
1318
1319         lock_kernel();
1320
1321         return rc;
1322 }
1323
1324 static int
1325 echo_client_setup(struct obd_device *obddev, obd_count len, void *buf)
1326 {
1327         struct lustre_cfg* lcfg = buf;
1328         struct echo_client_obd *ec = &obddev->u.echocli;
1329         struct obd_device *tgt;
1330         struct lustre_handle conn = {0, };
1331         struct obd_uuid echo_uuid = { "ECHO_UUID" };
1332         int rc;
1333         ENTRY;
1334
1335         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1336                 CERROR("requires a TARGET OBD name\n");
1337                 RETURN(-EINVAL);
1338         }
1339
1340         tgt = class_name2obd(lustre_cfg_string(lcfg, 1));
1341         if (!tgt || !tgt->obd_attached || !tgt->obd_set_up) {
1342                 CERROR("device not attached or not set up (%s)\n",
1343                        lustre_cfg_string(lcfg, 1));
1344                 RETURN(-EINVAL);
1345         }
1346
1347         spin_lock_init (&ec->ec_lock);
1348         INIT_LIST_HEAD (&ec->ec_objects);
1349         ec->ec_unique = 0;
1350
1351         rc = obd_connect(&conn, tgt, &echo_uuid, NULL, FILTER_GROUP_ECHO);
1352         if (rc) {
1353                 CERROR("fail to connect to device %s\n", lustre_cfg_string(lcfg, 1));
1354                 return (rc);
1355         }
1356         ec->ec_exp = class_conn2export(&conn);
1357
1358         RETURN(rc);
1359 }
1360
1361 static int echo_client_cleanup(struct obd_device *obddev, int flags)
1362 {
1363         struct list_head       *el;
1364         struct ec_object       *eco;
1365         struct echo_client_obd *ec = &obddev->u.echocli;
1366         int rc;
1367         ENTRY;
1368
1369         if (!list_empty(&obddev->obd_exports)) {
1370                 CERROR("still has clients!\n");
1371                 RETURN(-EBUSY);
1372         }
1373
1374         /* XXX assuming sole access */
1375         while (!list_empty(&ec->ec_objects)) {
1376                 el = ec->ec_objects.next;
1377                 eco = list_entry(el, struct ec_object, eco_obj_chain);
1378
1379                 LASSERT(eco->eco_refcount == 0);
1380                 eco->eco_refcount = 1;
1381                 eco->eco_deleted = 1;
1382                 echo_put_object(eco);
1383         }
1384
1385         rc = obd_disconnect(ec->ec_exp, 0);
1386         if (rc != 0)
1387                 CERROR("fail to disconnect device: %d\n", rc);
1388
1389         RETURN(rc);
1390 }
1391
1392 static int echo_client_connect(struct lustre_handle *conn,
1393                                struct obd_device *src,
1394                                struct obd_uuid *cluuid,
1395                                struct obd_connect_data *data,
1396                                unsigned long flags)
1397 {
1398         struct obd_export *exp;
1399         int                rc;
1400
1401         rc = class_connect(conn, src, cluuid);
1402         if (rc == 0) {
1403                 exp = class_conn2export(conn);
1404                 INIT_LIST_HEAD(&exp->exp_ec_data.eced_locks);
1405                 class_export_put(exp);
1406         }
1407
1408         RETURN (rc);
1409 }
1410
1411 static int echo_client_disconnect(struct obd_export *exp,
1412                                   unsigned long flags)
1413 {
1414         struct obd_device      *obd;
1415         struct echo_client_obd *ec;
1416         struct ec_lock         *ecl;
1417         int                     rc;
1418         ENTRY;
1419
1420         if (exp == NULL)
1421                 GOTO(out, rc = -EINVAL);
1422
1423         obd = exp->exp_obd;
1424         ec = &obd->u.echocli;
1425
1426         /* no more contention on export's lock list */
1427         while (!list_empty (&exp->exp_ec_data.eced_locks)) {
1428                 ecl = list_entry (exp->exp_ec_data.eced_locks.next,
1429                                   struct ec_lock, ecl_exp_chain);
1430                 list_del (&ecl->ecl_exp_chain);
1431
1432                 rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm,
1433                                  ecl->ecl_mode, &ecl->ecl_lock_handle);
1434
1435                 CDEBUG (D_INFO, "Cancel lock on object "LPX64" on disconnect "
1436                         "(%d)\n", ecl->ecl_object->eco_id, rc);
1437
1438                 echo_put_object (ecl->ecl_object);
1439                 OBD_FREE (ecl, sizeof (*ecl));
1440         }
1441
1442         rc = class_disconnect(exp, 0);
1443         GOTO(out, rc);
1444  out:
1445         return rc;
1446 }
1447
1448 static struct obd_ops echo_obd_ops = {
1449         .o_owner       = THIS_MODULE,
1450         .o_setup       = echo_client_setup,
1451         .o_cleanup     = echo_client_cleanup,
1452         .o_iocontrol   = echo_client_iocontrol,
1453         .o_connect     = echo_client_connect,
1454         .o_disconnect  = echo_client_disconnect
1455 };
1456
1457 int echo_client_init(void)
1458 {
1459         struct lprocfs_static_vars lvars;
1460
1461         lprocfs_init_vars(echo, &lvars);
1462         return class_register_type(&echo_obd_ops, NULL, lvars.module_vars,
1463                                    OBD_ECHO_CLIENT_DEVICENAME);
1464 }
1465
1466 void echo_client_exit(void)
1467 {
1468         class_unregister_type(OBD_ECHO_CLIENT_DEVICENAME);
1469 }