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                 oa->o_gr = FILTER_GROUP_ECHO;
 222                 oa->o_valid |= OBD_MD_FLGROUP;
 223                 rc = obd_create(ec->ec_exp, oa, &lsm, oti);
 224                 if (rc != 0)
 225                         goto failed;
 226
 227                 /* See what object ID we were given */
 228                 eco->eco_id = oa->o_id = lsm->lsm_object_id;
 229                 oa->o_valid |= OBD_MD_FLID;
 230
 231                 LASSERT(eco->eco_lsm == NULL || eco->eco_lsm == lsm);
 232                 eco->eco_lsm = lsm;
 233         }
 234
 235         spin_lock (&ec->ec_lock);
 236
 237         eco2 = echo_find_object_locked (obd, oa->o_id);
 238         if (eco2 != NULL) {                     /* conflict */
 239                 spin_unlock (&ec->ec_lock);
 240
 241                 CERROR ("Can't create object id "LPX64": id already exists%s\n",
 242                         oa->o_id, on_target ? " (undoing create)" : "");
 243
 244                 if (on_target)
 245                         obd_destroy(ec->ec_exp, oa, lsm, oti);
 246
 247                 rc = -EEXIST;
 248                 goto failed;
 249         }
 250
 251         list_add (&eco->eco_obj_chain, &ec->ec_objects);
 252         spin_unlock (&ec->ec_lock);
 253         CDEBUG (D_INFO,
 254                 "created %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 255                 eco, eco->eco_id,
 256                 eco->eco_lsm->lsm_stripe_size,
 257                 eco->eco_lsm->lsm_stripe_count,
 258                 eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 259                 eco->eco_refcount, eco->eco_deleted);
 260         return (0);
 261
 262  failed:
 263         echo_free_object (eco);
 264         return (rc);
 265 }
 266
 267 static int
 268 echo_get_object (struct ec_object **ecop, struct obd_device *obd,
 269                  struct obdo *oa)
 270 {
 271         struct echo_client_obd *ec = &obd->u.echo_client;
 272         struct ec_object       *eco;
 273         struct ec_object       *eco2;
 274         int                     rc;
 275
 276         if ((oa->o_valid & OBD_MD_FLID) == 0 ||
 277             oa->o_id == 0)                      /* disallow use of object id 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
 427 echo_client_page_debug_setup(struct lov_stripe_md *lsm,
 428                              struct page *page, int rw, obd_id id,
 429                              obd_off offset, obd_off count)
 430 {
 431         char    *addr;
 432         obd_off  stripe_off;
 433         obd_id   stripe_id;
 434         int      delta;
 435
 436         /* no partial pages on the client */
 437         LASSERT(count == PAGE_SIZE);
 438
 439         addr = kmap(page);
 440
 441         for (delta = 0; delta < PAGE_SIZE; delta += OBD_ECHO_BLOCK_SIZE) {
 442                 if (rw == OBD_BRW_WRITE) {
 443                         stripe_off = offset + delta;
 444                         stripe_id = id;
 445                         echo_get_stripe_off_id(lsm, &stripe_off, &stripe_id);
 446                 } else {
 447                         stripe_off = 0xdeadbeef00c0ffeeULL;
 448                         stripe_id = 0xdeadbeef00c0ffeeULL;
 449                 }
 450                 block_debug_setup(addr + delta, OBD_ECHO_BLOCK_SIZE,
 451                                   stripe_off, stripe_id);
 452         }
 453
 454         kunmap(page);
 455 }
 456
 457 static int
 458 echo_client_page_debug_check(struct lov_stripe_md *lsm,
 459                              struct page *page, obd_id id,
 460                              obd_off offset, obd_off count)
 461 {
 462         obd_off stripe_off;
 463         obd_id  stripe_id;
 464         char   *addr;
 465         int     delta;
 466         int     rc;
 467         int     rc2;
 468
 469         /* no partial pages on the client */
 470         LASSERT(count == PAGE_SIZE);
 471
 472         addr = kmap(page);
 473
 474         for (rc = delta = 0; delta < PAGE_SIZE; delta += OBD_ECHO_BLOCK_SIZE) {
 475                 stripe_off = offset + delta;
 476                 stripe_id = id;
 477                 echo_get_stripe_off_id (lsm, &stripe_off, &stripe_id);
 478
 479                 rc2 = block_debug_check("test_brw",
 480                                         addr + delta, OBD_ECHO_BLOCK_SIZE,
 481                                         stripe_off, stripe_id);
 482                 if (rc2 != 0)
 483                         rc = rc2;
 484         }
 485
 486         kunmap(page);
 487         return rc;
 488 }
 489
 490 static int echo_client_kbrw(struct obd_device *obd, int rw, struct obdo *oa,
 491                             struct lov_stripe_md *lsm, obd_off offset,
 492                             obd_size count, struct obd_trans_info *oti)
 493 {
 494         struct echo_client_obd *ec = &obd->u.echo_client;
 495         obd_count               npages;
 496         struct brw_page        *pga;
 497         struct brw_page        *pgp;
 498         obd_off                 off;
 499         int                     i;
 500         int                     rc;
 501         int                     verify = 0;
 502         int                     gfp_mask;
 503
 504         /* oa_id == ECHO_PERSISTENT_OBJID => speed test (no verification).
 505          * oa & 1                         => use HIGHMEM */
 506
 507         verify = (oa->o_id) != ECHO_PERSISTENT_OBJID;
 508         gfp_mask = ((oa->o_id & 2) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 509
 510         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 511
 512         if (count <= 0 ||
 513             (count & (PAGE_SIZE - 1)) != 0 ||
 514             (lsm != NULL &&
 515              lsm->lsm_object_id != oa->o_id))
 516                 return (-EINVAL);
 517
 518         /* XXX think again with misaligned I/O */
 519         npages = count >> PAGE_SHIFT;
 520
 521         OBD_ALLOC(pga, npages * sizeof(*pga));
 522         if (pga == NULL)
 523                 return (-ENOMEM);
 524
 525         for (i = 0, pgp = pga, off = offset;
 526              i < npages;
 527              i++, pgp++, off += PAGE_SIZE) {
 528
 529                 LASSERT (pgp->pg == NULL);      /* for cleanup */
 530
 531                 rc = -ENOMEM;
 532                 pgp->pg = alloc_pages (gfp_mask, 0);
 533                 if (pgp->pg == NULL)
 534                         goto out;
 535
 536                 pgp->count = PAGE_SIZE;
 537                 pgp->off = off;
 538                 pgp->flag = 0;
 539
 540                 if (verify)
 541                         echo_client_page_debug_setup(lsm, pgp->pg, rw,
 542                                                      oa->o_id, off, pgp->count);
 543         }
 544
 545         rc = obd_brw(rw, ec->ec_exp, oa, lsm, npages, pga, oti);
 546
 547  out:
 548         if (rc != 0 || rw != OBD_BRW_READ)
 549                 verify = 0;
 550
 551         for (i = 0, pgp = pga; i < npages; i++, pgp++) {
 552                 if (pgp->pg == NULL)
 553                         continue;
 554
 555                 if (verify) {
 556                         int vrc;
 557                         vrc = echo_client_page_debug_check(lsm, pgp->pg, oa->o_id,
 558                                                            pgp->off, pgp->count);
 559                         if (vrc != 0 && rc == 0)
 560                                 rc = vrc;
 561                 }
 562                 __free_pages(pgp->pg, 0);
 563         }
 564         OBD_FREE(pga, npages * sizeof(*pga));
 565         return (rc);
 566 }
 567
 568 #ifdef __KERNEL__
 569 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 570 static int echo_client_ubrw(struct obd_device *obd, int rw,
 571                             struct obdo *oa, struct lov_stripe_md *lsm,
 572                             obd_off offset, obd_size count, char *buffer,
 573                             struct obd_trans_info *oti)
 574 {
 575         struct echo_client_obd *ec = &obd->u.echo_client;
 576         obd_count               npages;
 577         struct brw_page        *pga;
 578         struct brw_page        *pgp;
 579         obd_off                 off;
 580         struct kiobuf          *kiobuf;
 581         int                     i;
 582         int                     rc;
 583
 584         LASSERT (rw == OBD_BRW_WRITE ||
 585                  rw == OBD_BRW_READ);
 586
 587         /* NB: for now, only whole pages, page aligned */
 588
 589         if (count <= 0 ||
 590             ((long)buffer & (PAGE_SIZE - 1)) != 0 ||
 591             (count & (PAGE_SIZE - 1)) != 0 ||
 592             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 593                 return (-EINVAL);
 594
 595         /* XXX think again with misaligned I/O */
 596         npages = count >> PAGE_SHIFT;
 597
 598         OBD_ALLOC(pga, npages * sizeof(*pga));
 599         if (pga == NULL)
 600                 return (-ENOMEM);
 601
 602         rc = alloc_kiovec (1, &kiobuf);
 603         if (rc != 0)
 604                 goto out_1;
 605
 606         rc = map_user_kiobuf ((rw == OBD_BRW_READ) ? READ : WRITE,
 607                               kiobuf, (unsigned long)buffer, count);
 608         if (rc != 0)
 609                 goto out_2;
 610
 611         LASSERT (kiobuf->offset == 0);
 612         LASSERT (kiobuf->nr_pages == npages);
 613
 614         for (i = 0, off = offset, pgp = pga;
 615              i < npages;
 616              i++, off += PAGE_SIZE, pgp++) {
 617                 pgp->off = off;
 618                 pgp->pg = kiobuf->maplist[i];
 619                 pgp->count = PAGE_SIZE;
 620                 pgp->flag = 0;
 621         }
 622
 623         rc = obd_brw(rw, ec->ec_exp, oa, lsm, npages, pga, oti);
 624
 625         //        if (rw == OBD_BRW_READ)
 626         //                mark_dirty_kiobuf (kiobuf, count);
 627
 628         unmap_kiobuf (kiobuf);
 629  out_2:
 630         free_kiovec (1, &kiobuf);
 631  out_1:
 632         OBD_FREE(pga, npages * sizeof(*pga));
 633         return (rc);
 634 }
 635 #else
 636 static int echo_client_ubrw(struct obd_device *obd, int rw,
 637                             struct obdo *oa, struct lov_stripe_md *lsm,
 638                             obd_off offset, obd_size count, char *buffer,
 639                             struct obd_trans_info *oti)
 640 {
 641 #warning "echo_client_ubrw() needs to be ported on 2.6 yet"
 642         LBUG();
 643         return 0;
 644 }
 645 #endif
 646 #endif
 647
 648 struct echo_async_state;
 649
 650 #define EAP_MAGIC 79277927
 651 struct echo_async_page {
 652         int                     eap_magic;
 653         struct page             *eap_page;
 654         void                    *eap_cookie;
 655         obd_off                 eap_off;
 656         struct echo_async_state *eap_eas;
 657         struct list_head        eap_item;
 658 };
 659
 660 struct echo_async_state {
 661         spinlock_t              eas_lock;
 662         obd_off                 eas_next_offset;
 663         obd_off                 eas_end_offset;
 664         int                     eas_in_flight;
 665         int                     eas_rc;
 666         wait_queue_head_t       eas_waitq;
 667         struct list_head        eas_avail;
 668         struct obdo             eas_oa;
 669         struct lov_stripe_md    *eas_lsm;
 670 };
 671
 672 static int eas_should_wake(struct echo_async_state *eas)
 673 {
 674         unsigned long flags;
 675         int rc = 0;
 676         spin_lock_irqsave(&eas->eas_lock, flags);
 677         if (eas->eas_rc == 0 && !list_empty(&eas->eas_avail))
 678             rc = 1;
 679         spin_unlock_irqrestore(&eas->eas_lock, flags);
 680         return rc;
 681 };
 682
 683 struct echo_async_page *eap_from_cookie(void *cookie)
 684 {
 685         struct echo_async_page *eap = cookie;
 686         if (eap->eap_magic != EAP_MAGIC)
 687                 return ERR_PTR(-EINVAL);
 688         return eap;
 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;
 706         eap = eap_from_cookie(data);
 707         if (IS_ERR(eap))
 708                 return;
 709
 710         memcpy(oa, &eap->eap_eas->eas_oa, sizeof(*oa));
 711 }
 712
 713 static void ec_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 714 {
 715         struct echo_async_page *eap = eap_from_cookie(data);
 716         struct echo_async_state *eas;
 717         unsigned long flags;
 718
 719         if (IS_ERR(eap))
 720                 return;
 721         eas = eap->eap_eas;
 722
 723         if (cmd == OBD_BRW_READ &&
 724             eas->eas_oa.o_id != ECHO_PERSISTENT_OBJID)
 725                 echo_client_page_debug_check(eas->eas_lsm, eap->eap_page,
 726                                              eas->eas_oa.o_id, eap->eap_off,
 727                                              PAGE_SIZE);
 728
 729         spin_lock_irqsave(&eas->eas_lock, flags);
 730         if (rc && !eas->eas_rc)
 731                 eas->eas_rc = rc;
 732         eas->eas_in_flight--;
 733         list_add(&eap->eap_item, &eas->eas_avail);
 734         wake_up(&eas->eas_waitq);
 735         spin_unlock_irqrestore(&eas->eas_lock, flags);
 736 }
 737
 738 static struct obd_async_page_ops ec_async_page_ops = {
 739         .ap_make_ready =        ec_ap_make_ready,
 740         .ap_refresh_count =     ec_ap_refresh_count,
 741         .ap_fill_obdo =         ec_ap_fill_obdo,
 742         .ap_completion =        ec_ap_completion,
 743 };
 744
 745 static int echo_client_async_page(struct obd_export *exp, int rw,
 746                                    struct obdo *oa, struct lov_stripe_md *lsm,
 747                                    obd_off offset, obd_size count,
 748                                    obd_size batching)
 749 {
 750         obd_count npages, i;
 751         struct echo_async_page *eap;
 752         struct echo_async_state eas;
 753         struct list_head *pos, *n;
 754         int rc = 0;
 755         unsigned long flags;
 756         LIST_HEAD(pages);
 757 #if 0
 758         int                     verify;
 759         int                     gfp_mask;
 760         /* oa_id  == 0    => speed test (no verification) else...
 761          * oa & 1         => use HIGHMEM
 762          */
 763         verify = (oa->o_id != 0);
 764         gfp_mask = ((oa->o_id & 1) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 765 #endif
 766
 767         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 768
 769         if (count <= 0 ||
 770             (count & (PAGE_SIZE - 1)) != 0 ||
 771             (lsm != NULL &&
 772              lsm->lsm_object_id != oa->o_id))
 773                 return (-EINVAL);
 774
 775         /* XXX think again with misaligned I/O */
 776         npages = batching >> PAGE_SHIFT;
 777
 778         memcpy(&eas.eas_oa, oa, sizeof(*oa));
 779         eas.eas_next_offset = offset;
 780         eas.eas_end_offset = offset + count;
 781         spin_lock_init(&eas.eas_lock);
 782         init_waitqueue_head(&eas.eas_waitq);
 783         eas.eas_in_flight = 0;
 784         eas.eas_rc = 0;
 785         eas.eas_lsm = lsm;
 786         INIT_LIST_HEAD(&eas.eas_avail);
 787
 788         /* prepare the group of pages that we're going to be keeping
 789          * in flight */
 790         for (i = 0; i < npages; i++) {
 791                 struct page *page = alloc_page(GFP_KERNEL);
 792                 if (page == NULL)
 793                         GOTO(out, rc = -ENOMEM);
 794
 795                 page->private = 0;
 796                 list_add_tail(&PAGE_LIST(page), &pages);
 797
 798                 OBD_ALLOC(eap, sizeof(*eap));
 799                 if (eap == NULL)
 800                         GOTO(out, rc = -ENOMEM);
 801
 802                 eap->eap_magic = EAP_MAGIC;
 803                 eap->eap_page = page;
 804                 eap->eap_eas = &eas;
 805                 page->private = (unsigned long)eap;
 806                 list_add_tail(&eap->eap_item, &eas.eas_avail);
 807         }
 808
 809         /* first we spin queueing io and being woken by its completion */
 810         spin_lock_irqsave(&eas.eas_lock, flags);
 811         for(;;) {
 812                 int rc;
 813
 814                 /* sleep until we have a page to send */
 815                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 816                 rc = wait_event_interruptible(eas.eas_waitq,
 817                                               eas_should_wake(&eas));
 818                 spin_lock_irqsave(&eas.eas_lock, flags);
 819                 if (rc && !eas.eas_rc)
 820                         eas.eas_rc = rc;
 821                 if (eas.eas_rc)
 822                         break;
 823                 if (list_empty(&eas.eas_avail))
 824                         continue;
 825                 eap = list_entry(eas.eas_avail.next, struct echo_async_page,
 826                                  eap_item);
 827                 list_del(&eap->eap_item);
 828                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 829
 830                 /* unbind the eap from its old page offset */
 831                 if (eap->eap_cookie != NULL) {
 832                         obd_teardown_async_page(exp, lsm, NULL,
 833                                                 eap->eap_cookie);
 834                         eap->eap_cookie = NULL;
 835                 }
 836
 837                 eas.eas_next_offset += PAGE_SIZE;
 838                 eap->eap_off = eas.eas_next_offset;
 839
 840                 rc = obd_prep_async_page(exp, lsm, NULL, eap->eap_page,
 841                                          eap->eap_off, &ec_async_page_ops,
 842                                          eap, &eap->eap_cookie);
 843                 if (rc) {
 844                         spin_lock_irqsave(&eas.eas_lock, flags);
 845                         eas.eas_rc = rc;
 846                         break;
 847                 }
 848
 849                 if (oa->o_id != ECHO_PERSISTENT_OBJID)
 850                         echo_client_page_debug_setup(lsm, eap->eap_page, rw,
 851                                                      oa->o_id,
 852                                                      eap->eap_off, PAGE_SIZE);
 853
 854                 /* always asserts urgent, which isn't quite right */
 855                 rc = obd_queue_async_io(exp, lsm, NULL, eap->eap_cookie,
 856                                         rw, 0, PAGE_SIZE, 0,
 857                                         ASYNC_READY | ASYNC_URGENT |
 858                                         ASYNC_COUNT_STABLE);
 859                 spin_lock_irqsave(&eas.eas_lock, flags);
 860                 if (rc && !eas.eas_rc) {
 861                         eas.eas_rc = rc;
 862                         break;
 863                 }
 864                 eas.eas_in_flight++;
 865                 if (eas.eas_next_offset == eas.eas_end_offset)
 866                         break;
 867         }
 868
 869         /* still hold the eas_lock here.. */
 870
 871         /* now we just spin waiting for all the rpcs to complete */
 872         while(eas.eas_in_flight) {
 873                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 874                 wait_event_interruptible(eas.eas_waitq,
 875                                          eas.eas_in_flight == 0);
 876                 spin_lock_irqsave(&eas.eas_lock, flags);
 877         }
 878         spin_unlock_irqrestore(&eas.eas_lock, flags);
 879
 880 out:
 881         list_for_each_safe(pos, n, &pages) {
 882                 struct page *page = list_entry(pos, struct page,
 883                                                PAGE_LIST_ENTRY);
 884
 885                 list_del(&PAGE_LIST(page));
 886                 if (page->private != 0) {
 887                         eap = (struct echo_async_page *)page->private;
 888                         if (eap->eap_cookie != NULL)
 889                                 obd_teardown_async_page(exp, lsm, NULL,
 890                                                         eap->eap_cookie);
 891                         OBD_FREE(eap, sizeof(*eap));
 892                 }
 893                 __free_page(page);
 894         }
 895
 896         RETURN(rc);
 897 }
 898
 899 static int echo_client_prep_commit(struct obd_export *exp, int rw,
 900                                    struct obdo *oa, struct lov_stripe_md *lsm,
 901                                    obd_off offset, obd_size count,
 902                                    obd_size batch, struct obd_trans_info *oti)
 903 {
 904         struct obd_ioobj ioo;
 905         struct niobuf_local *lnb;
 906         struct niobuf_remote *rnb;
 907         obd_off off;
 908         obd_size npages, tot_pages;
 909         int i, ret = 0;
 910         ENTRY;
 911
 912         if (count <= 0 || (count & (PAGE_SIZE - 1)) != 0 ||
 913             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 914                 RETURN(-EINVAL);
 915
 916         npages = batch >> PAGE_SHIFT;
 917         tot_pages = count >> PAGE_SHIFT;
 918
 919         OBD_ALLOC(lnb, npages * sizeof(struct niobuf_local));
 920         OBD_ALLOC(rnb, npages * sizeof(struct niobuf_remote));
 921
 922         if (lnb == NULL || rnb == NULL)
 923                 GOTO(out, ret = -ENOMEM);
 924
 925         obdo_to_ioobj(oa, &ioo);
 926
 927         off = offset;
 928
 929         for(; tot_pages; tot_pages -= npages) {
 930                 if (tot_pages < npages)
 931                         npages = tot_pages;
 932
 933                 for (i = 0; i < npages; i++, off += PAGE_SIZE) {
 934                         rnb[i].offset = off;
 935                         rnb[i].len = PAGE_SIZE;
 936                 }
 937
 938                 /* XXX this can't be the best.. */
 939                 memset(oti, 0, sizeof(*oti));
 940                 ioo.ioo_bufcnt = npages;
 941
 942                 ret = obd_preprw(rw, exp, oa, 1, &ioo, npages, rnb, lnb, oti);
 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                                 continue;
 955
 956                         if (rw == OBD_BRW_WRITE)
 957                                 echo_client_page_debug_setup(lsm, page, rw,
 958                                                              oa->o_id,
 959                                                              rnb[i].offset,
 960                                                              rnb[i].len);
 961                         else
 962                                 echo_client_page_debug_check(lsm, page,
 963                                                              oa->o_id,
 964                                                              rnb[i].offset,
 965                                                              rnb[i].len);
 966                 }
 967
 968                 ret = obd_commitrw(rw, exp, oa, 1, &ioo, npages, lnb, oti, ret);
 969                 if (ret != 0)
 970                         GOTO(out, ret);
 971         }
 972
 973 out:
 974         if (lnb)
 975                 OBD_FREE(lnb, npages * sizeof(struct niobuf_local));
 976         if (rnb)
 977                 OBD_FREE(rnb, npages * sizeof(struct niobuf_remote));
 978         RETURN(ret);
 979 }
 980
 981 int echo_client_brw_ioctl(int rw, struct obd_export *exp,
 982                           struct obd_ioctl_data *data)
 983 {
 984         struct obd_device *obd = class_exp2obd(exp);
 985         struct echo_client_obd *ec = &obd->u.echo_client;
 986         struct obd_trans_info dummy_oti;
 987         struct ec_object *eco;
 988         int rc;
 989         ENTRY;
 990
 991         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
 992         if (rc)
 993                 RETURN(rc);
 994
 995         memset(&dummy_oti, 0, sizeof(dummy_oti));
 996
 997         data->ioc_obdo1.o_valid &= ~OBD_MD_FLHANDLE;
 998         data->ioc_obdo1.o_valid |= OBD_MD_FLGROUP;
 999         data->ioc_obdo1.o_gr = FILTER_GROUP_ECHO;
1000
1001         switch((long)data->ioc_pbuf1) {
1002         case 1:
1003                 if (data->ioc_pbuf2 == NULL) { // NULL user data pointer
1004                         rc = echo_client_kbrw(obd, rw, &data->ioc_obdo1,
1005                                               eco->eco_lsm, data->ioc_offset,
1006                                               data->ioc_count, &dummy_oti);
1007                 } else {
1008 #ifdef __KERNEL__
1009                         rc = echo_client_ubrw(obd, rw, &data->ioc_obdo1,
1010                                               eco->eco_lsm, data->ioc_offset,
1011                                               data->ioc_count, data->ioc_pbuf2,
1012                                               &dummy_oti);
1013 #endif
1014                 }
1015                 break;
1016         case 2:
1017                 rc = echo_client_async_page(ec->ec_exp, rw, &data->ioc_obdo1,
1018                                            eco->eco_lsm, data->ioc_offset,
1019                                            data->ioc_count, data->ioc_plen1);
1020                 break;
1021         case 3:
1022                 rc = echo_client_prep_commit(ec->ec_exp, rw, &data->ioc_obdo1,
1023                                             eco->eco_lsm, data->ioc_offset,
1024                                             data->ioc_count, data->ioc_plen1,
1025                                             &dummy_oti);
1026                 break;
1027         default:
1028                 rc = -EINVAL;
1029         }
1030         echo_put_object(eco);
1031         RETURN(rc);
1032 }
1033
1034 static int
1035 echo_ldlm_callback (struct ldlm_lock *lock, struct ldlm_lock_desc *new,
1036                     void *data, int flag)
1037 {
1038         struct ec_object       *eco = (struct ec_object *)data;
1039         struct echo_client_obd *ec = &(eco->eco_device->u.echo_client);
1040         struct lustre_handle    lockh;
1041         struct list_head       *el;
1042         int                     found = 0;
1043         int                     rc;
1044
1045         ldlm_lock2handle (lock, &lockh);
1046
1047         /* #ifdef this out if we're not feeling paranoid */
1048         spin_lock (&ec->ec_lock);
1049         list_for_each (el, &ec->ec_objects) {
1050                 found = (eco == list_entry(el, struct ec_object,
1051                                            eco_obj_chain));
1052                 if (found)
1053                         break;
1054         }
1055         spin_unlock (&ec->ec_lock);
1056         LASSERT (found);
1057
1058         switch (flag) {
1059         case LDLM_CB_BLOCKING:
1060                 CDEBUG(D_INFO, "blocking callback on "LPX64", handle "LPX64"\n",
1061                        eco->eco_id, lockh.cookie);
1062                 rc = ldlm_cli_cancel (&lockh);
1063                 if (rc != ELDLM_OK)
1064                         CERROR ("ldlm_cli_cancel failed: %d\n", rc);
1065                 break;
1066
1067         case LDLM_CB_CANCELING:
1068                 CDEBUG(D_INFO, "cancel callback on "LPX64", handle "LPX64"\n",
1069                        eco->eco_id, lockh.cookie);
1070                 break;
1071
1072         default:
1073                 LBUG ();
1074         }
1075
1076         return (0);
1077 }
1078
1079 static int
1080 echo_client_enqueue(struct obd_export *exp, struct obdo *oa,
1081                     int mode, obd_off offset, obd_size nob)
1082 {
1083         struct obd_device      *obd = exp->exp_obd;
1084         struct echo_client_obd *ec = &obd->u.echo_client;
1085         struct lustre_handle   *ulh = obdo_handle (oa);
1086         struct ec_object       *eco;
1087         struct ec_lock         *ecl;
1088         int                     flags;
1089         int                     rc;
1090
1091         if (!(mode == LCK_PR || mode == LCK_PW))
1092                 return -EINVAL;
1093
1094         if ((offset & (PAGE_SIZE - 1)) != 0 ||
1095             (nob & (PAGE_SIZE - 1)) != 0)
1096                 return -EINVAL;
1097
1098         rc = echo_get_object (&eco, obd, oa);
1099         if (rc != 0)
1100                 return rc;
1101
1102         rc = -ENOMEM;
1103         OBD_ALLOC (ecl, sizeof (*ecl));
1104         if (ecl == NULL)
1105                 goto failed_0;
1106
1107         ecl->ecl_mode = mode;
1108         ecl->ecl_object = eco;
1109         ecl->ecl_policy.l_extent.start = offset;
1110         ecl->ecl_policy.l_extent.end =
1111                 (nob == 0) ? ((obd_off) -1) : (offset + nob - 1);
1112
1113         flags = 0;
1114         rc = obd_enqueue(ec->ec_exp, eco->eco_lsm, LDLM_EXTENT,
1115                          &ecl->ecl_policy, mode, &flags, echo_ldlm_callback,
1116                          ldlm_completion_ast, NULL, eco, sizeof(struct ost_lvb),
1117                          lustre_swab_ost_lvb, &ecl->ecl_lock_handle);
1118         if (rc != 0)
1119                 goto failed_1;
1120
1121         CDEBUG(D_INFO, "enqueue handle "LPX64"\n", ecl->ecl_lock_handle.cookie);
1122
1123         /* NB ecl takes object ref from echo_get_object() above */
1124         spin_lock(&ec->ec_lock);
1125
1126         list_add(&ecl->ecl_exp_chain, &exp->exp_ec_data.eced_locks);
1127         ulh->cookie = ecl->ecl_cookie = ec->ec_unique++;
1128
1129         spin_unlock(&ec->ec_lock);
1130
1131         oa->o_valid |= OBD_MD_FLHANDLE;
1132         return 0;
1133
1134  failed_1:
1135         OBD_FREE (ecl, sizeof (*ecl));
1136  failed_0:
1137         echo_put_object (eco);
1138         return (rc);
1139 }
1140
1141 static int
1142 echo_client_cancel(struct obd_export *exp, struct obdo *oa)
1143 {
1144         struct obd_device      *obd = exp->exp_obd;
1145         struct echo_client_obd *ec = &obd->u.echo_client;
1146         struct lustre_handle   *ulh = obdo_handle (oa);
1147         struct ec_lock         *ecl = NULL;
1148         int                     found = 0;
1149         struct list_head       *el;
1150         int                     rc;
1151
1152         if ((oa->o_valid & OBD_MD_FLHANDLE) == 0)
1153                 return -EINVAL;
1154
1155         spin_lock (&ec->ec_lock);
1156
1157         list_for_each (el, &exp->exp_ec_data.eced_locks) {
1158                 ecl = list_entry (el, struct ec_lock, ecl_exp_chain);
1159                 found = (ecl->ecl_cookie == ulh->cookie);
1160                 if (found) {
1161                         list_del (&ecl->ecl_exp_chain);
1162                         break;
1163                 }
1164         }
1165
1166         spin_unlock (&ec->ec_lock);
1167
1168         if (!found)
1169                 return (-ENOENT);
1170
1171         rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm, ecl->ecl_mode,
1172                         &ecl->ecl_lock_handle);
1173
1174         echo_put_object (ecl->ecl_object);
1175         OBD_FREE (ecl, sizeof (*ecl));
1176
1177         return rc;
1178 }
1179
1180 static int
1181 echo_client_iocontrol(unsigned int cmd, struct obd_export *exp,
1182                       int len, void *karg, void *uarg)
1183 {
1184         struct obd_device      *obd;
1185         struct echo_client_obd *ec;
1186         struct ec_object       *eco;
1187         struct obd_ioctl_data  *data = karg;
1188         struct obd_trans_info   dummy_oti;
1189         struct oti_req_ack_lock *ack_lock;
1190         struct obdo            *oa;
1191         int                     rw = OBD_BRW_READ;
1192         int                     rc = 0;
1193         int                     i;
1194         ENTRY;
1195
1196         memset(&dummy_oti, 0, sizeof(dummy_oti));
1197
1198         obd = exp->exp_obd;
1199         ec = &obd->u.echo_client;
1200
1201         switch (cmd) {
1202         case OBD_IOC_CREATE:                    /* may create echo object */
1203                 if (!capable (CAP_SYS_ADMIN))
1204                         GOTO (out, rc = -EPERM);
1205
1206                 rc = echo_create_object (obd, 1, &data->ioc_obdo1,
1207                                          data->ioc_pbuf1, data->ioc_plen1,
1208                                          &dummy_oti);
1209                 GOTO(out, rc);
1210
1211         case OBD_IOC_DESTROY:
1212                 if (!capable (CAP_SYS_ADMIN))
1213                         GOTO (out, rc = -EPERM);
1214
1215                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1216                 if (rc == 0) {
1217                         oa = &data->ioc_obdo1;
1218                         oa->o_gr = FILTER_GROUP_ECHO;
1219                         oa->o_valid |= OBD_MD_FLGROUP;
1220                         rc = obd_destroy(ec->ec_exp, oa, eco->eco_lsm,
1221                                          &dummy_oti);
1222                         if (rc == 0)
1223                                 eco->eco_deleted = 1;
1224                         echo_put_object(eco);
1225                 }
1226                 GOTO(out, rc);
1227
1228         case OBD_IOC_GETATTR:
1229                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1230                 if (rc == 0) {
1231                         rc = obd_getattr(ec->ec_exp, &data->ioc_obdo1,
1232                                          eco->eco_lsm);
1233                         echo_put_object(eco);
1234                 }
1235                 GOTO(out, rc);
1236
1237         case OBD_IOC_SETATTR:
1238                 if (!capable (CAP_SYS_ADMIN))
1239                         GOTO (out, rc = -EPERM);
1240
1241                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1242                 if (rc == 0) {
1243                         rc = obd_setattr(ec->ec_exp, &data->ioc_obdo1,
1244                                          eco->eco_lsm, NULL);
1245                         echo_put_object(eco);
1246                 }
1247                 GOTO(out, rc);
1248
1249         case OBD_IOC_BRW_WRITE:
1250                 if (!capable (CAP_SYS_ADMIN))
1251                         GOTO (out, rc = -EPERM);
1252
1253                 rw = OBD_BRW_WRITE;
1254                 /* fall through */
1255         case OBD_IOC_BRW_READ:
1256                 rc = echo_client_brw_ioctl(rw, exp, data);
1257                 GOTO(out, rc);
1258
1259         case ECHO_IOC_GET_STRIPE:
1260                 rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1261                 if (rc == 0) {
1262                         rc = echo_copyout_lsm(eco->eco_lsm, data->ioc_pbuf1,
1263                                               data->ioc_plen1);
1264                         echo_put_object(eco);
1265                 }
1266                 GOTO(out, rc);
1267
1268         case ECHO_IOC_SET_STRIPE:
1269                 if (!capable (CAP_SYS_ADMIN))
1270                         GOTO (out, rc = -EPERM);
1271
1272                 if (data->ioc_pbuf1 == NULL) {  /* unset */
1273                         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1274                         if (rc == 0) {
1275                                 eco->eco_deleted = 1;
1276                                 echo_put_object(eco);
1277                         }
1278                 } else {
1279                         rc = echo_create_object(obd, 0, &data->ioc_obdo1,
1280                                                 data->ioc_pbuf1,
1281                                                 data->ioc_plen1, &dummy_oti);
1282                 }
1283                 GOTO (out, rc);
1284
1285         case ECHO_IOC_ENQUEUE:
1286                 if (!capable (CAP_SYS_ADMIN))
1287                         GOTO (out, rc = -EPERM);
1288
1289                 rc = echo_client_enqueue(exp, &data->ioc_obdo1,
1290                                          data->ioc_conn1, /* lock mode */
1291                                    data->ioc_offset, data->ioc_count);/*extent*/
1292                 GOTO (out, rc);
1293
1294         case ECHO_IOC_CANCEL:
1295                 rc = echo_client_cancel(exp, &data->ioc_obdo1);
1296                 GOTO (out, rc);
1297
1298         default:
1299                 CERROR ("echo_ioctl(): unrecognised ioctl %#x\n", cmd);
1300                 GOTO (out, rc = -ENOTTY);
1301         }
1302
1303         EXIT;
1304  out:
1305
1306         /* XXX this should be in a helper also called by target_send_reply */
1307         for (ack_lock = dummy_oti.oti_ack_locks, i = 0; i < 4;
1308              i++, ack_lock++) {
1309                 if (!ack_lock->mode)
1310                         break;
1311                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1312         }
1313
1314         return rc;
1315 }
1316
1317 static int
1318 echo_client_setup(struct obd_device *obddev, obd_count len, void *buf)
1319 {
1320         struct lustre_cfg* lcfg = buf;
1321         struct echo_client_obd *ec = &obddev->u.echo_client;
1322         struct obd_device *tgt;
1323         struct lustre_handle conn = {0, };
1324         struct obd_uuid echo_uuid = { "ECHO_UUID" };
1325         int rc;
1326         ENTRY;
1327
1328         if (lcfg->lcfg_inllen1 < 1) {
1329                 CERROR("requires a TARGET OBD name\n");
1330                 RETURN(-EINVAL);
1331         }
1332
1333         tgt = class_name2obd(lcfg->lcfg_inlbuf1);
1334         if (!tgt || !tgt->obd_attached || !tgt->obd_set_up) {
1335                 CERROR("device not attached or not set up (%s)\n",
1336                        lcfg->lcfg_inlbuf1);
1337                 RETURN(-EINVAL);
1338         }
1339
1340         spin_lock_init (&ec->ec_lock);
1341         INIT_LIST_HEAD (&ec->ec_objects);
1342         ec->ec_unique = 0;
1343
1344         rc = obd_connect(&conn, tgt, &echo_uuid);
1345         if (rc) {
1346                 CERROR("fail to connect to device %s\n", lcfg->lcfg_inlbuf1);
1347                 return (rc);
1348         }
1349         ec->ec_exp = class_conn2export(&conn);
1350
1351         RETURN(rc);
1352 }
1353
1354 static int echo_client_cleanup(struct obd_device *obddev, int flags)
1355 {
1356         struct list_head       *el;
1357         struct ec_object       *eco;
1358         struct echo_client_obd *ec = &obddev->u.echo_client;
1359         int rc;
1360         ENTRY;
1361
1362         if (!list_empty(&obddev->obd_exports)) {
1363                 CERROR("still has clients!\n");
1364                 RETURN(-EBUSY);
1365         }
1366
1367         /* XXX assuming sole access */
1368         while (!list_empty(&ec->ec_objects)) {
1369                 el = ec->ec_objects.next;
1370                 eco = list_entry(el, struct ec_object, eco_obj_chain);
1371
1372                 LASSERT(eco->eco_refcount == 0);
1373                 eco->eco_refcount = 1;
1374                 eco->eco_deleted = 1;
1375                 echo_put_object(eco);
1376         }
1377
1378         rc = obd_disconnect(ec->ec_exp, 0);
1379         if (rc != 0)
1380                 CERROR("fail to disconnect device: %d\n", rc);
1381
1382         RETURN(rc);
1383 }
1384
1385 static int echo_client_connect(struct lustre_handle *conn,
1386                                struct obd_device *src, struct obd_uuid *cluuid)
1387 {
1388         struct obd_export *exp;
1389         int                rc;
1390
1391         rc = class_connect(conn, src, cluuid);
1392         if (rc == 0) {
1393                 exp = class_conn2export(conn);
1394                 INIT_LIST_HEAD(&exp->exp_ec_data.eced_locks);
1395                 class_export_put(exp);
1396         }
1397
1398         RETURN (rc);
1399 }
1400
1401 static int echo_client_disconnect(struct obd_export *exp, int flags)
1402 {
1403         struct obd_device      *obd;
1404         struct echo_client_obd *ec;
1405         struct ec_lock         *ecl;
1406         int                     rc;
1407         ENTRY;
1408
1409         if (exp == NULL)
1410                 GOTO(out, rc = -EINVAL);
1411
1412         obd = exp->exp_obd;
1413         ec = &obd->u.echo_client;
1414
1415         /* no more contention on export's lock list */
1416         while (!list_empty (&exp->exp_ec_data.eced_locks)) {
1417                 ecl = list_entry (exp->exp_ec_data.eced_locks.next,
1418                                   struct ec_lock, ecl_exp_chain);
1419                 list_del (&ecl->ecl_exp_chain);
1420
1421                 rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm,
1422                                  ecl->ecl_mode, &ecl->ecl_lock_handle);
1423
1424                 CDEBUG (D_INFO, "Cancel lock on object "LPX64" on disconnect "
1425                         "(%d)\n", ecl->ecl_object->eco_id, rc);
1426
1427                 echo_put_object (ecl->ecl_object);
1428                 OBD_FREE (ecl, sizeof (*ecl));
1429         }
1430
1431         rc = class_disconnect(exp, 0);
1432         GOTO(out, rc);
1433  out:
1434         return rc;
1435 }
1436
1437 static struct obd_ops echo_obd_ops = {
1438         .o_owner       = THIS_MODULE,
1439         .o_setup       = echo_client_setup,
1440         .o_cleanup     = echo_client_cleanup,
1441         .o_iocontrol   = echo_client_iocontrol,
1442         .o_connect     = echo_client_connect,
1443         .o_disconnect  = echo_client_disconnect
1444 };
1445
1446 int echo_client_init(void)
1447 {
1448         struct lprocfs_static_vars lvars;
1449
1450         lprocfs_init_vars(echo, &lvars);
1451         return class_register_type(&echo_obd_ops, NULL, lvars.module_vars,
1452                                    OBD_ECHO_CLIENT_DEVICENAME);
1453 }
1454
1455 void echo_client_exit(void)
1456 {
1457         class_unregister_type(OBD_ECHO_CLIENT_DEVICENAME);
1458 }