lustre/obdecho/echo_client.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #define DEBUG_SUBSYSTEM S_ECHO
  23 #ifdef __KERNEL__
  24 #include <linux/version.h>
  25 #include <linux/module.h>
  26 #include <linux/fs.h>
  27 #include <linux/completion.h>
  28 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  29 #include <linux/iobuf.h>
  30 #endif
  31 #include <asm/div64.h>
  32 #else
  33 #include <liblustre.h>
  34 #endif
  35
  36 #include <linux/obd.h>
  37 #include <linux/obd_support.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/obd_echo.h>
  40 #include <linux/lustre_debug.h>
  41 #include <linux/lprocfs_status.h>
  42
  43 static obd_id last_object_id;
  44
  45 #if 0
  46 static void
  47 echo_printk_object (char *msg, struct ec_object *eco)
  48 {
  49         struct lov_stripe_md *lsm = eco->eco_lsm;
  50         int                   i;
  51
  52         printk (KERN_INFO "Lustre: %s: object %p: "LPX64", refs %d%s: "LPX64
  53                 "=%u!%u\n", msg, eco, eco->eco_id, eco->eco_refcount,
  54                 eco->eco_deleted ? "(deleted) " : "",
  55                 lsm->lsm_object_id, lsm->lsm_stripe_size,
  56                 lsm->lsm_stripe_count);
  57
  58         for (i = 0; i < lsm->lsm_stripe_count; i++)
  59                 printk (KERN_INFO "Lustre:   @%2u:"LPX64"\n",
  60                         lsm->lsm_oinfo[i].loi_ost_idx,
  61                         lsm->lsm_oinfo[i].loi_id);
  62 }
  63 #endif
  64
  65 static struct ec_object *
  66 echo_find_object_locked (struct obd_device *obd, obd_id id)
  67 {
  68         struct echo_client_obd *ec = &obd->u.echo_client;
  69         struct ec_object       *eco = NULL;
  70         struct list_head       *el;
  71
  72         list_for_each (el, &ec->ec_objects) {
  73                 eco = list_entry (el, struct ec_object, eco_obj_chain);
  74
  75                 if (eco->eco_id == id)
  76                         return (eco);
  77         }
  78         return (NULL);
  79 }
  80
  81 static int
  82 echo_copyout_lsm (struct lov_stripe_md *lsm, void *ulsm, int ulsm_nob)
  83 {
  84         int nob;
  85
  86         nob = offsetof (struct lov_stripe_md, lsm_oinfo[lsm->lsm_stripe_count]);
  87         if (nob > ulsm_nob)
  88                 return (-EINVAL);
  89
  90         if (copy_to_user (ulsm, lsm, nob))
  91                 return (-EFAULT);
  92
  93         return (0);
  94 }
  95
  96 static int
  97 echo_copyin_lsm (struct obd_device *obd, struct lov_stripe_md *lsm,
  98                  void *ulsm, int ulsm_nob)
  99 {
 100         struct echo_client_obd *ec = &obd->u.echo_client;
 101         int                     nob;
 102
 103         if (ulsm_nob < sizeof (*lsm))
 104                 return (-EINVAL);
 105
 106         if (copy_from_user (lsm, ulsm, sizeof (*lsm)))
 107                 return (-EFAULT);
 108
 109         nob = lsm->lsm_stripe_count * sizeof (lsm->lsm_oinfo[0]);
 110
 111         if (ulsm_nob < nob ||
 112             lsm->lsm_stripe_count > ec->ec_nstripes ||
 113             lsm->lsm_magic != LOV_MAGIC ||
 114             (lsm->lsm_stripe_size & (PAGE_SIZE - 1)) != 0 ||
 115             ((__u64)lsm->lsm_stripe_size * lsm->lsm_stripe_count > ~0UL))
 116                 return (-EINVAL);
 117
 118         if (copy_from_user(lsm->lsm_oinfo,
 119                            ((struct lov_stripe_md *)ulsm)->lsm_oinfo, nob))
 120                 return (-EFAULT);
 121
 122         return (0);
 123 }
 124
 125 static struct ec_object *
 126 echo_allocate_object (struct obd_device *obd)
 127 {
 128         struct echo_client_obd *ec = &obd->u.echo_client;
 129         struct ec_object       *eco;
 130         int rc;
 131
 132         OBD_ALLOC(eco, sizeof (*eco));
 133         if (eco == NULL)
 134                 return NULL;
 135
 136         rc = obd_alloc_memmd(ec->ec_exp, &eco->eco_lsm);
 137         if (rc < 0) {
 138                 OBD_FREE(eco, sizeof (*eco));
 139                 return NULL;
 140         }
 141
 142         eco->eco_device = obd;
 143         eco->eco_deleted = 0;
 144         eco->eco_refcount = 0;
 145         eco->eco_lsm->lsm_magic = LOV_MAGIC;
 146         /* leave stripe count 0 by default */
 147
 148         return (eco);
 149 }
 150
 151 static void
 152 echo_free_object (struct ec_object *eco)
 153 {
 154         struct obd_device      *obd = eco->eco_device;
 155         struct echo_client_obd *ec = &obd->u.echo_client;
 156
 157         LASSERT (eco->eco_refcount == 0);
 158         obd_free_memmd(ec->ec_exp, &eco->eco_lsm);
 159         OBD_FREE (eco, sizeof (*eco));
 160 }
 161
 162 static int echo_create_object(struct obd_device *obd, int on_target,
 163                               struct obdo *oa, void *ulsm, int ulsm_nob,
 164                               struct obd_trans_info *oti)
 165 {
 166         struct echo_client_obd *ec = &obd->u.echo_client;
 167         struct ec_object       *eco2;
 168         struct ec_object       *eco;
 169         struct lov_stripe_md   *lsm;
 170         int                     rc;
 171         int                     i, idx;
 172
 173         if ((oa->o_valid & OBD_MD_FLID) == 0 && /* no obj id */
 174             (on_target ||                       /* set_stripe */
 175              ec->ec_nstripes != 0)) {           /* LOV */
 176                 CERROR ("No valid oid\n");
 177                 return (-EINVAL);
 178         }
 179
 180         if (ulsm != NULL) {
 181                 eco = echo_allocate_object (obd);
 182                 if (eco == NULL)
 183                         return (-ENOMEM);
 184
 185                 lsm = eco->eco_lsm;
 186
 187                 rc = echo_copyin_lsm (obd, lsm, ulsm, ulsm_nob);
 188                 if (rc != 0)
 189                         goto failed;
 190
 191                 /* setup object ID here for !on_target and LOV hint */
 192                 if ((oa->o_valid & OBD_MD_FLID) != 0)
 193                         eco->eco_id = lsm->lsm_object_id = oa->o_id;
 194
 195                 if (lsm->lsm_stripe_count == 0)
 196                         lsm->lsm_stripe_count = ec->ec_nstripes;
 197
 198                 if (lsm->lsm_stripe_size == 0)
 199                         lsm->lsm_stripe_size = PAGE_SIZE;
 200
 201                 idx = ll_insecure_random_int();
 202
 203                 /* setup stripes: indices + default ids if required */
 204                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 205                         if (lsm->lsm_oinfo[i].loi_id == 0)
 206                                 lsm->lsm_oinfo[i].loi_id = lsm->lsm_object_id;
 207
 208                         lsm->lsm_oinfo[i].loi_ost_idx =
 209                                 (idx + i) % ec->ec_nstripes;
 210                 }
 211         } else {
 212                 OBD_ALLOC(eco, sizeof(*eco));
 213                 eco->eco_device = obd;
 214                 lsm = NULL;
 215         }
 216
 217         if (oa->o_id == 0)
 218                 oa->o_id = ++last_object_id;
 219
 220         if (on_target) {
 221                 /* XXX get some filter group constants */
 222                 oa->o_gr = 2;
 223                 oa->o_valid |= OBD_MD_FLGROUP;
 224                 rc = obd_create(ec->ec_exp, oa, &lsm, oti);
 225                 if (rc != 0)
 226                         goto failed;
 227
 228                 /* See what object ID we were given */
 229                 eco->eco_id = oa->o_id = lsm->lsm_object_id;
 230                 oa->o_valid |= OBD_MD_FLID;
 231
 232                 LASSERT(eco->eco_lsm == NULL || eco->eco_lsm == lsm);
 233                 eco->eco_lsm = lsm;
 234         }
 235
 236         spin_lock (&ec->ec_lock);
 237
 238         eco2 = echo_find_object_locked (obd, oa->o_id);
 239         if (eco2 != NULL) {                     /* conflict */
 240                 spin_unlock (&ec->ec_lock);
 241
 242                 CERROR ("Can't create object id "LPX64": id already exists%s\n",
 243                         oa->o_id, on_target ? " (undoing create)" : "");
 244
 245                 if (on_target)
 246                         obd_destroy(ec->ec_exp, oa, lsm, oti);
 247
 248                 rc = -EEXIST;
 249                 goto failed;
 250         }
 251
 252         list_add (&eco->eco_obj_chain, &ec->ec_objects);
 253         spin_unlock (&ec->ec_lock);
 254         CDEBUG (D_INFO,
 255                 "created %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 256                 eco, eco->eco_id,
 257                 eco->eco_lsm->lsm_stripe_size,
 258                 eco->eco_lsm->lsm_stripe_count,
 259                 eco->eco_lsm->lsm_oinfo[0].loi_ost_idx,
 260                 eco->eco_refcount, eco->eco_deleted);
 261         return (0);
 262
 263  failed:
 264         echo_free_object (eco);
 265         return (rc);
 266 }
 267
 268 static int
 269 echo_get_object (struct ec_object **ecop, struct obd_device *obd,
 270                  struct obdo *oa)
 271 {
 272         struct echo_client_obd *ec = &obd->u.echo_client;
 273         struct ec_object       *eco;
 274         struct ec_object       *eco2;
 275         int                     rc;
 276
 277         if ((oa->o_valid & OBD_MD_FLID) == 0 ||
 278             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.echo_client;
 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                         rc = rc2;
 485         }
 486
 487         kunmap(page);
 488         return rc;
 489 }
 490
 491 static int echo_client_kbrw(struct obd_device *obd, int rw, struct obdo *oa,
 492                             struct lov_stripe_md *lsm, obd_off offset,
 493                             obd_size count, struct obd_trans_info *oti)
 494 {
 495         struct echo_client_obd *ec = &obd->u.echo_client;
 496         obd_count               npages;
 497         struct brw_page        *pga;
 498         struct brw_page        *pgp;
 499         obd_off                 off;
 500         int                     i;
 501         int                     rc;
 502         int                     verify = 0;
 503         int                     gfp_mask;
 504
 505         /* oa_id == ECHO_PERSISTENT_OBJID => speed test (no verification).
 506          * oa & 1                         => use HIGHMEM */
 507
 508         verify = (oa->o_id) != ECHO_PERSISTENT_OBJID;
 509         gfp_mask = ((oa->o_id & 2) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 510
 511         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 512
 513         if (count <= 0 ||
 514             (count & (PAGE_SIZE - 1)) != 0 ||
 515             (lsm != NULL &&
 516              lsm->lsm_object_id != oa->o_id))
 517                 return (-EINVAL);
 518
 519         /* XXX think again with misaligned I/O */
 520         npages = count >> PAGE_SHIFT;
 521
 522         OBD_ALLOC(pga, npages * sizeof(*pga));
 523         if (pga == NULL)
 524                 return (-ENOMEM);
 525
 526         for (i = 0, pgp = pga, off = offset;
 527              i < npages;
 528              i++, pgp++, off += PAGE_SIZE) {
 529
 530                 LASSERT (pgp->pg == NULL);      /* for cleanup */
 531
 532                 rc = -ENOMEM;
 533                 pgp->pg = alloc_pages (gfp_mask, 0);
 534                 if (pgp->pg == NULL)
 535                         goto out;
 536
 537                 pgp->count = PAGE_SIZE;
 538                 pgp->off = off;
 539                 pgp->flag = 0;
 540
 541                 if (verify)
 542                         echo_client_page_debug_setup(lsm, pgp->pg, rw,
 543                                                      oa->o_id, off, pgp->count);
 544         }
 545
 546         rc = obd_brw(rw, ec->ec_exp, oa, lsm, npages, pga, oti);
 547
 548  out:
 549         if (rc != 0 || rw != OBD_BRW_READ)
 550                 verify = 0;
 551
 552         for (i = 0, pgp = pga; i < npages; i++, pgp++) {
 553                 if (pgp->pg == NULL)
 554                         continue;
 555
 556                 if (verify) {
 557                         int vrc;
 558                         vrc = echo_client_page_debug_check(lsm, pgp->pg, oa->o_id,
 559                                                            pgp->off, pgp->count);
 560                         if (vrc != 0 && rc == 0)
 561                                 rc = vrc;
 562                 }
 563                 __free_pages(pgp->pg, 0);
 564         }
 565         OBD_FREE(pga, npages * sizeof(*pga));
 566         return (rc);
 567 }
 568
 569 #ifdef __KERNEL__
 570 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 571 static int echo_client_ubrw(struct obd_device *obd, int rw,
 572                             struct obdo *oa, struct lov_stripe_md *lsm,
 573                             obd_off offset, obd_size count, char *buffer,
 574                             struct obd_trans_info *oti)
 575 {
 576         struct echo_client_obd *ec = &obd->u.echo_client;
 577         obd_count               npages;
 578         struct brw_page        *pga;
 579         struct brw_page        *pgp;
 580         obd_off                 off;
 581         struct kiobuf          *kiobuf;
 582         int                     i;
 583         int                     rc;
 584
 585         LASSERT (rw == OBD_BRW_WRITE ||
 586                  rw == OBD_BRW_READ);
 587
 588         /* NB: for now, only whole pages, page aligned */
 589
 590         if (count <= 0 ||
 591             ((long)buffer & (PAGE_SIZE - 1)) != 0 ||
 592             (count & (PAGE_SIZE - 1)) != 0 ||
 593             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 594                 return (-EINVAL);
 595
 596         /* XXX think again with misaligned I/O */
 597         npages = count >> PAGE_SHIFT;
 598
 599         OBD_ALLOC(pga, npages * sizeof(*pga));
 600         if (pga == NULL)
 601                 return (-ENOMEM);
 602
 603         rc = alloc_kiovec (1, &kiobuf);
 604         if (rc != 0)
 605                 goto out_1;
 606
 607         rc = map_user_kiobuf ((rw == OBD_BRW_READ) ? READ : WRITE,
 608                               kiobuf, (unsigned long)buffer, count);
 609         if (rc != 0)
 610                 goto out_2;
 611
 612         LASSERT (kiobuf->offset == 0);
 613         LASSERT (kiobuf->nr_pages == npages);
 614
 615         for (i = 0, off = offset, pgp = pga;
 616              i < npages;
 617              i++, off += PAGE_SIZE, pgp++) {
 618                 pgp->off = off;
 619                 pgp->pg = kiobuf->maplist[i];
 620                 pgp->count = PAGE_SIZE;
 621                 pgp->flag = 0;
 622         }
 623
 624         rc = obd_brw(rw, ec->ec_exp, oa, lsm, npages, pga, oti);
 625
 626         //        if (rw == OBD_BRW_READ)
 627         //                mark_dirty_kiobuf (kiobuf, count);
 628
 629         unmap_kiobuf (kiobuf);
 630  out_2:
 631         free_kiovec (1, &kiobuf);
 632  out_1:
 633         OBD_FREE(pga, npages * sizeof(*pga));
 634         return (rc);
 635 }
 636 #else
 637 static int echo_client_ubrw(struct obd_device *obd, int rw,
 638                             struct obdo *oa, struct lov_stripe_md *lsm,
 639                             obd_off offset, obd_size count, char *buffer,
 640                             struct obd_trans_info *oti)
 641 {
 642 #warning "echo_client_ubrw() needs to be ported on 2.6 yet"
 643         LBUG();
 644         return 0;
 645 }
 646 #endif
 647 #endif
 648
 649 struct echo_async_state;
 650
 651 #define EAP_MAGIC 79277927
 652 struct echo_async_page {
 653         int                     eap_magic;
 654         struct page             *eap_page;
 655         void                    *eap_cookie;
 656         obd_off                 eap_off;
 657         struct echo_async_state *eap_eas;
 658         struct list_head        eap_item;
 659 };
 660
 661 struct echo_async_state {
 662         spinlock_t              eas_lock;
 663         obd_off                 eas_next_offset;
 664         obd_off                 eas_end_offset;
 665         int                     eas_in_flight;
 666         int                     eas_rc;
 667         wait_queue_head_t       eas_waitq;
 668         struct list_head        eas_avail;
 669         struct obdo             eas_oa;
 670         struct lov_stripe_md    *eas_lsm;
 671 };
 672
 673 static int eas_should_wake(struct echo_async_state *eas)
 674 {
 675         unsigned long flags;
 676         int rc = 0;
 677         spin_lock_irqsave(&eas->eas_lock, flags);
 678         if (eas->eas_rc == 0 && !list_empty(&eas->eas_avail))
 679             rc = 1;
 680         spin_unlock_irqrestore(&eas->eas_lock, flags);
 681         return rc;
 682 };
 683
 684 struct echo_async_page *eap_from_cookie(void *cookie)
 685 {
 686         struct echo_async_page *eap = cookie;
 687         if (eap->eap_magic != EAP_MAGIC)
 688                 return ERR_PTR(-EINVAL);
 689         return eap;
 690 };
 691
 692 static int ec_ap_make_ready(void *data, int cmd)
 693 {
 694         /* our pages are issued ready */
 695         LBUG();
 696         return 0;
 697 }
 698 static int ec_ap_refresh_count(void *data, int cmd)
 699 {
 700         /* our pages are issued with a stable count */
 701         LBUG();
 702         return PAGE_SIZE;
 703 }
 704 static void ec_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 705 {
 706         struct echo_async_page *eap;
 707         eap = eap_from_cookie(data);
 708         if (IS_ERR(eap))
 709                 return;
 710
 711         memcpy(oa, &eap->eap_eas->eas_oa, sizeof(*oa));
 712 }
 713
 714 static void ec_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 715 {
 716         struct echo_async_page *eap = eap_from_cookie(data);
 717         struct echo_async_state *eas;
 718         unsigned long flags;
 719
 720         if (IS_ERR(eap))
 721                 return;
 722         eas = eap->eap_eas;
 723
 724         if (cmd == OBD_BRW_READ &&
 725             eas->eas_oa.o_id != ECHO_PERSISTENT_OBJID)
 726                 echo_client_page_debug_check(eas->eas_lsm, eap->eap_page,
 727                                              eas->eas_oa.o_id, eap->eap_off,
 728                                              PAGE_SIZE);
 729
 730         spin_lock_irqsave(&eas->eas_lock, flags);
 731         if (rc && !eas->eas_rc)
 732                 eas->eas_rc = rc;
 733         eas->eas_in_flight--;
 734         list_add(&eap->eap_item, &eas->eas_avail);
 735         wake_up(&eas->eas_waitq);
 736         spin_unlock_irqrestore(&eas->eas_lock, flags);
 737 }
 738
 739 static struct obd_async_page_ops ec_async_page_ops = {
 740         .ap_make_ready =        ec_ap_make_ready,
 741         .ap_refresh_count =     ec_ap_refresh_count,
 742         .ap_fill_obdo =         ec_ap_fill_obdo,
 743         .ap_completion =        ec_ap_completion,
 744 };
 745
 746 static int echo_client_async_page(struct obd_export *exp, int rw,
 747                                    struct obdo *oa, struct lov_stripe_md *lsm,
 748                                    obd_off offset, obd_size count,
 749                                    obd_size batching)
 750 {
 751         obd_count npages, i;
 752         struct echo_async_page *eap;
 753         struct echo_async_state eas;
 754         struct list_head *pos, *n;
 755         int rc = 0;
 756         unsigned long flags;
 757         LIST_HEAD(pages);
 758 #if 0
 759         int                     verify;
 760         int                     gfp_mask;
 761         /* oa_id  == 0    => speed test (no verification) else...
 762          * oa & 1         => use HIGHMEM
 763          */
 764         verify = (oa->o_id != 0);
 765         gfp_mask = ((oa->o_id & 1) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 766 #endif
 767
 768         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 769
 770         if (count <= 0 ||
 771             (count & (PAGE_SIZE - 1)) != 0 ||
 772             (lsm != NULL &&
 773              lsm->lsm_object_id != oa->o_id))
 774                 return (-EINVAL);
 775
 776         /* XXX think again with misaligned I/O */
 777         npages = batching >> PAGE_SHIFT;
 778
 779         memcpy(&eas.eas_oa, oa, sizeof(*oa));
 780         eas.eas_next_offset = offset;
 781         eas.eas_end_offset = offset + count;
 782         spin_lock_init(&eas.eas_lock);
 783         init_waitqueue_head(&eas.eas_waitq);
 784         eas.eas_in_flight = 0;
 785         eas.eas_rc = 0;
 786         eas.eas_lsm = lsm;
 787         INIT_LIST_HEAD(&eas.eas_avail);
 788
 789         /* prepare the group of pages that we're going to be keeping
 790          * in flight */
 791         for (i = 0; i < npages; i++) {
 792                 struct page *page = alloc_page(GFP_KERNEL);
 793                 if (page == NULL)
 794                         GOTO(out, rc = -ENOMEM);
 795
 796                 page->private = 0;
 797                 list_add_tail(&PAGE_LIST(page), &pages);
 798
 799                 OBD_ALLOC(eap, sizeof(*eap));
 800                 if (eap == NULL)
 801                         GOTO(out, rc = -ENOMEM);
 802
 803                 eap->eap_magic = EAP_MAGIC;
 804                 eap->eap_page = page;
 805                 eap->eap_eas = &eas;
 806                 page->private = (unsigned long)eap;
 807                 list_add_tail(&eap->eap_item, &eas.eas_avail);
 808         }
 809
 810         /* first we spin queueing io and being woken by its completion */
 811         spin_lock_irqsave(&eas.eas_lock, flags);
 812         for(;;) {
 813                 int rc;
 814
 815                 /* sleep until we have a page to send */
 816                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 817                 rc = wait_event_interruptible(eas.eas_waitq,
 818                                               eas_should_wake(&eas));
 819                 spin_lock_irqsave(&eas.eas_lock, flags);
 820                 if (rc && !eas.eas_rc)
 821                         eas.eas_rc = rc;
 822                 if (eas.eas_rc)
 823                         break;
 824                 if (list_empty(&eas.eas_avail))
 825                         continue;
 826                 eap = list_entry(eas.eas_avail.next, struct echo_async_page,
 827                                  eap_item);
 828                 list_del(&eap->eap_item);
 829                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 830
 831                 /* unbind the eap from its old page offset */
 832                 if (eap->eap_cookie != NULL) {
 833                         obd_teardown_async_page(exp, lsm, NULL,
 834                                                 eap->eap_cookie);
 835                         eap->eap_cookie = NULL;
 836                 }
 837
 838                 eas.eas_next_offset += PAGE_SIZE;
 839                 eap->eap_off = eas.eas_next_offset;
 840
 841                 rc = obd_prep_async_page(exp, lsm, NULL, eap->eap_page,
 842                                          eap->eap_off, &ec_async_page_ops,
 843                                          eap, &eap->eap_cookie);
 844                 if (rc) {
 845                         spin_lock_irqsave(&eas.eas_lock, flags);
 846                         eas.eas_rc = rc;
 847                         break;
 848                 }
 849
 850                 if (oa->o_id != ECHO_PERSISTENT_OBJID)
 851                         echo_client_page_debug_setup(lsm, eap->eap_page, rw,
 852                                                      oa->o_id,
 853                                                      eap->eap_off, PAGE_SIZE);
 854
 855                 /* always asserts urgent, which isn't quite right */
 856                 rc = obd_queue_async_io(exp, lsm, NULL, eap->eap_cookie,
 857                                         rw, 0, PAGE_SIZE, 0,
 858                                         ASYNC_READY | ASYNC_URGENT |
 859                                         ASYNC_COUNT_STABLE);
 860                 spin_lock_irqsave(&eas.eas_lock, flags);
 861                 if (rc && !eas.eas_rc) {
 862                         eas.eas_rc = rc;
 863                         break;
 864                 }
 865                 eas.eas_in_flight++;
 866                 if (eas.eas_next_offset == eas.eas_end_offset)
 867                         break;
 868         }
 869
 870         /* still hold the eas_lock here.. */
 871
 872         /* now we just spin waiting for all the rpcs to complete */
 873         while(eas.eas_in_flight) {
 874                 spin_unlock_irqrestore(&eas.eas_lock, flags);
 875                 wait_event_interruptible(eas.eas_waitq,
 876                                          eas.eas_in_flight == 0);
 877                 spin_lock_irqsave(&eas.eas_lock, flags);
 878         }
 879         spin_unlock_irqrestore(&eas.eas_lock, flags);
 880
 881 out:
 882         list_for_each_safe(pos, n, &pages) {
 883                 struct page *page = list_entry(pos, struct page,
 884                                                PAGE_LIST_ENTRY);
 885
 886                 list_del(&PAGE_LIST(page));
 887                 if (page->private != 0) {
 888                         eap = (struct echo_async_page *)page->private;
 889                         if (eap->eap_cookie != NULL)
 890                                 obd_teardown_async_page(exp, lsm, NULL,
 891                                                         eap->eap_cookie);
 892                         OBD_FREE(eap, sizeof(*eap));
 893                 }
 894                 __free_page(page);
 895         }
 896
 897         RETURN(rc);
 898 }
 899
 900 static int echo_client_prep_commit(struct obd_export *exp, int rw,
 901                                    struct obdo *oa, struct lov_stripe_md *lsm,
 902                                    obd_off offset, obd_size count,
 903                                    obd_size batch, struct obd_trans_info *oti)
 904 {
 905         struct obd_ioobj ioo;
 906         struct niobuf_local *lnb;
 907         struct niobuf_remote *rnb;
 908         obd_off off;
 909         obd_size npages, tot_pages;
 910         int i, ret = 0;
 911         ENTRY;
 912
 913         if (count <= 0 || (count & (PAGE_SIZE - 1)) != 0 ||
 914             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 915                 RETURN(-EINVAL);
 916
 917         npages = batch >> PAGE_SHIFT;
 918         tot_pages = count >> PAGE_SHIFT;
 919
 920         OBD_ALLOC(lnb, npages * sizeof(struct niobuf_local));
 921         OBD_ALLOC(rnb, npages * sizeof(struct niobuf_remote));
 922
 923         if (lnb == NULL || rnb == NULL)
 924                 GOTO(out, ret = -ENOMEM);
 925
 926         obdo_to_ioobj(oa, &ioo);
 927
 928         off = offset;
 929
 930         for(; tot_pages; tot_pages -= npages) {
 931                 if (tot_pages < npages)
 932                         npages = tot_pages;
 933
 934                 for (i = 0; i < npages; i++, off += PAGE_SIZE) {
 935                         rnb[i].offset = off;
 936                         rnb[i].len = PAGE_SIZE;
 937                 }
 938
 939                 /* XXX this can't be the best.. */
 940                 memset(oti, 0, sizeof(*oti));
 941                 ioo.ioo_bufcnt = npages;
 942
 943                 ret = obd_preprw(rw, exp, oa, 1, &ioo, npages, rnb, lnb, oti);
 944                 if (ret != 0)
 945                         GOTO(out, ret);
 946
 947                 for (i = 0; i < npages; i++) {
 948                         struct page *page = lnb[i].page;
 949
 950                         /* read past eof? */
 951                         if (page == NULL && lnb[i].rc == 0)
 952                                 continue;
 953
 954                         if (oa->o_id == ECHO_PERSISTENT_OBJID)
 955                                 continue;
 956
 957                         if (rw == OBD_BRW_WRITE)
 958                                 echo_client_page_debug_setup(lsm, page, rw,
 959                                                              oa->o_id,
 960                                                              rnb[i].offset,
 961                                                              rnb[i].len);
 962                         else
 963                                 echo_client_page_debug_check(lsm, page,
 964                                                              oa->o_id,
 965                                                              rnb[i].offset,
 966                                                              rnb[i].len);
 967                 }
 968
 969                 ret = obd_commitrw(rw, exp, oa, 1, &ioo, npages, lnb, oti, ret);
 970                 if (ret != 0)
 971                         GOTO(out, ret);
 972         }
 973
 974 out:
 975         if (lnb)
 976                 OBD_FREE(lnb, npages * sizeof(struct niobuf_local));
 977         if (rnb)
 978                 OBD_FREE(rnb, npages * sizeof(struct niobuf_remote));
 979         RETURN(ret);
 980 }
 981
 982 int echo_client_brw_ioctl(int rw, struct obd_export *exp,
 983                           struct obd_ioctl_data *data)
 984 {
 985         struct obd_device *obd = class_exp2obd(exp);
 986         struct echo_client_obd *ec = &obd->u.echo_client;
 987         struct obd_trans_info dummy_oti;
 988         struct ec_object *eco;
 989         int rc;
 990         ENTRY;
 991
 992         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
 993         if (rc)
 994                 RETURN(rc);
 995
 996         memset(&dummy_oti, 0, sizeof(dummy_oti));
 997
 998         data->ioc_obdo1.o_valid &= ~OBD_MD_FLHANDLE;
 999         data->ioc_obdo1.o_valid |= OBD_MD_FLGROUP;
1000         data->ioc_obdo1.o_gr = 2;
1001
1002         switch((long)data->ioc_pbuf1) {
1003         case 1:
1004                 if (data->ioc_pbuf2 == NULL) { // NULL user data pointer
1005                         rc = echo_client_kbrw(obd, rw, &data->ioc_obdo1,
1006                                               eco->eco_lsm, data->ioc_offset,
1007                                               data->ioc_count, &dummy_oti);
1008                 } else {
1009 #ifdef __KERNEL__
1010                         rc = echo_client_ubrw(obd, rw, &data->ioc_obdo1,
1011                                               eco->eco_lsm, data->ioc_offset,
1012                                               data->ioc_count, data->ioc_pbuf2,
1013                                               &dummy_oti);
1014 #endif
1015                 }
1016                 break;
1017         case 2:
1018                 rc = echo_client_async_page(ec->ec_exp, rw, &data->ioc_obdo1,
1019                                            eco->eco_lsm, data->ioc_offset,
1020                                            data->ioc_count, data->ioc_plen1);
1021                 break;
1022         case 3:
1023                 rc = echo_client_prep_commit(ec->ec_exp, rw, &data->ioc_obdo1,
1024                                             eco->eco_lsm, data->ioc_offset,
1025                                             data->ioc_count, data->ioc_plen1,
1026                                             &dummy_oti);
1027                 break;
1028         default:
1029                 rc = -EINVAL;
1030         }
1031         echo_put_object(eco);
1032         RETURN(rc);
1033 }
1034
1035 static int
1036 echo_ldlm_callback (struct ldlm_lock *lock, struct ldlm_lock_desc *new,
1037                     void *data, int flag)
1038 {
1039         struct ec_object       *eco = (struct ec_object *)data;
1040         struct echo_client_obd *ec = &(eco->eco_device->u.echo_client);
1041         struct lustre_handle    lockh;
1042         struct list_head       *el;
1043         int                     found = 0;
1044         int                     rc;
1045
1046         ldlm_lock2handle (lock, &lockh);
1047
1048         /* #ifdef this out if we're not feeling paranoid */
1049         spin_lock (&ec->ec_lock);
1050         list_for_each (el, &ec->ec_objects) {
1051                 found = (eco == list_entry(el, struct ec_object,
1052                                            eco_obj_chain));
1053                 if (found)
1054                         break;
1055         }
1056         spin_unlock (&ec->ec_lock);
1057         LASSERT (found);
1058
1059         switch (flag) {
1060         case LDLM_CB_BLOCKING:
1061                 CDEBUG(D_INFO, "blocking callback on "LPX64", handle "LPX64"\n",
1062                        eco->eco_id, lockh.cookie);
1063                 rc = ldlm_cli_cancel (&lockh);
1064                 if (rc != ELDLM_OK)
1065                         CERROR ("ldlm_cli_cancel failed: %d\n", rc);
1066                 break;
1067
1068         case LDLM_CB_CANCELING:
1069                 CDEBUG(D_INFO, "cancel callback on "LPX64", handle "LPX64"\n",
1070                        eco->eco_id, lockh.cookie);
1071                 break;
1072
1073         default:
1074                 LBUG ();
1075         }
1076
1077         return (0);
1078 }
1079
1080 static int
1081 echo_client_enqueue(struct obd_export *exp, struct obdo *oa,
1082                     int mode, obd_off offset, obd_size nob)
1083 {
1084         struct obd_device      *obd = exp->exp_obd;
1085         struct echo_client_obd *ec = &obd->u.echo_client;
1086         struct lustre_handle   *ulh = obdo_handle (oa);
1087         struct ec_object       *eco;
1088         struct ec_lock         *ecl;
1089         int                     flags;
1090         int                     rc;
1091
1092         if (!(mode == LCK_PR || mode == LCK_PW))
1093                 return -EINVAL;
1094
1095         if ((offset & (PAGE_SIZE - 1)) != 0 ||
1096             (nob & (PAGE_SIZE - 1)) != 0)
1097                 return -EINVAL;
1098
1099         rc = echo_get_object (&eco, obd, oa);
1100         if (rc != 0)
1101                 return rc;
1102
1103         rc = -ENOMEM;
1104         OBD_ALLOC (ecl, sizeof (*ecl));
1105         if (ecl == NULL)
1106                 goto failed_0;
1107
1108         ecl->ecl_mode = mode;
1109         ecl->ecl_object = eco;
1110         ecl->ecl_policy.l_extent.start = offset;
1111         ecl->ecl_policy.l_extent.end =
1112                 (nob == 0) ? ((obd_off) -1) : (offset + nob - 1);
1113
1114         flags = 0;
1115         rc = obd_enqueue(ec->ec_exp, eco->eco_lsm, LDLM_EXTENT,
1116                          &ecl->ecl_policy, mode, &flags, echo_ldlm_callback,
1117                          ldlm_completion_ast, NULL, eco, sizeof(struct ost_lvb),
1118                          lustre_swab_ost_lvb, &ecl->ecl_lock_handle);
1119         if (rc != 0)
1120                 goto failed_1;
1121
1122         CDEBUG(D_INFO, "enqueue handle "LPX64"\n", ecl->ecl_lock_handle.cookie);
1123
1124         /* NB ecl takes object ref from echo_get_object() above */
1125         spin_lock(&ec->ec_lock);
1126
1127         list_add(&ecl->ecl_exp_chain, &exp->exp_ec_data.eced_locks);
1128         ulh->cookie = ecl->ecl_cookie = ec->ec_unique++;
1129
1130         spin_unlock(&ec->ec_lock);
1131
1132         oa->o_valid |= OBD_MD_FLHANDLE;
1133         return 0;
1134
1135  failed_1:
1136         OBD_FREE (ecl, sizeof (*ecl));
1137  failed_0:
1138         echo_put_object (eco);
1139         return (rc);
1140 }
1141
1142 static int
1143 echo_client_cancel(struct obd_export *exp, struct obdo *oa)
1144 {
1145         struct obd_device      *obd = exp->exp_obd;
1146         struct echo_client_obd *ec = &obd->u.echo_client;
1147         struct lustre_handle   *ulh = obdo_handle (oa);
1148         struct ec_lock         *ecl = NULL;
1149         int                     found = 0;
1150         struct list_head       *el;
1151         int                     rc;
1152
1153         if ((oa->o_valid & OBD_MD_FLHANDLE) == 0)
1154                 return -EINVAL;
1155
1156         spin_lock (&ec->ec_lock);
1157
1158         list_for_each (el, &exp->exp_ec_data.eced_locks) {
1159                 ecl = list_entry (el, struct ec_lock, ecl_exp_chain);
1160                 found = (ecl->ecl_cookie == ulh->cookie);
1161                 if (found) {
1162                         list_del (&ecl->ecl_exp_chain);
1163                         break;
1164                 }
1165         }
1166
1167         spin_unlock (&ec->ec_lock);
1168
1169         if (!found)
1170                 return (-ENOENT);
1171
1172         rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm, ecl->ecl_mode,
1173                         &ecl->ecl_lock_handle);
1174
1175         echo_put_object (ecl->ecl_object);
1176         OBD_FREE (ecl, sizeof (*ecl));
1177
1178         return rc;
1179 }
1180
1181 static int
1182 echo_client_iocontrol(unsigned int cmd, struct obd_export *exp,
1183                       int len, void *karg, void *uarg)
1184 {
1185         struct obd_device      *obd;
1186         struct echo_client_obd *ec;
1187         struct ec_object       *eco;
1188         struct obd_ioctl_data  *data = karg;
1189         struct obd_trans_info   dummy_oti;
1190         struct oti_req_ack_lock *ack_lock;
1191         struct obdo            *oa;
1192         int                     rw = OBD_BRW_READ;
1193         int                     rc = 0;
1194         int                     i;
1195         ENTRY;
1196
1197         memset(&dummy_oti, 0, sizeof(dummy_oti));
1198
1199         obd = exp->exp_obd;
1200         ec = &obd->u.echo_client;
1201
1202         switch (cmd) {
1203         case OBD_IOC_CREATE:                    /* may create echo object */
1204                 if (!capable (CAP_SYS_ADMIN))
1205                         GOTO (out, rc = -EPERM);
1206
1207                 rc = echo_create_object (obd, 1, &data->ioc_obdo1,
1208                                          data->ioc_pbuf1, data->ioc_plen1,
1209                                          &dummy_oti);
1210                 GOTO(out, rc);
1211
1212         case OBD_IOC_DESTROY:
1213                 if (!capable (CAP_SYS_ADMIN))
1214                         GOTO (out, rc = -EPERM);
1215
1216                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1217                 if (rc == 0) {
1218                         oa = &data->ioc_obdo1;
1219                         oa->o_gr = 2;
1220                         oa->o_valid |= OBD_MD_FLGROUP;
1221                         rc = obd_destroy(ec->ec_exp, oa, eco->eco_lsm,
1222                                          &dummy_oti);
1223                         if (rc == 0)
1224                                 eco->eco_deleted = 1;
1225                         echo_put_object(eco);
1226                 }
1227                 GOTO(out, rc);
1228
1229         case OBD_IOC_GETATTR:
1230                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1231                 if (rc == 0) {
1232                         rc = obd_getattr(ec->ec_exp, &data->ioc_obdo1,
1233                                          eco->eco_lsm);
1234                         echo_put_object(eco);
1235                 }
1236                 GOTO(out, rc);
1237
1238         case OBD_IOC_SETATTR:
1239                 if (!capable (CAP_SYS_ADMIN))
1240                         GOTO (out, rc = -EPERM);
1241
1242                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1243                 if (rc == 0) {
1244                         rc = obd_setattr(ec->ec_exp, &data->ioc_obdo1,
1245                                          eco->eco_lsm, NULL);
1246                         echo_put_object(eco);
1247                 }
1248                 GOTO(out, rc);
1249
1250         case OBD_IOC_BRW_WRITE:
1251                 if (!capable (CAP_SYS_ADMIN))
1252                         GOTO (out, rc = -EPERM);
1253
1254                 rw = OBD_BRW_WRITE;
1255                 /* fall through */
1256         case OBD_IOC_BRW_READ:
1257                 rc = echo_client_brw_ioctl(rw, exp, data);
1258                 GOTO(out, rc);
1259
1260         case ECHO_IOC_GET_STRIPE:
1261                 rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1262                 if (rc == 0) {
1263                         rc = echo_copyout_lsm(eco->eco_lsm, data->ioc_pbuf1,
1264                                               data->ioc_plen1);
1265                         echo_put_object(eco);
1266                 }
1267                 GOTO(out, rc);
1268
1269         case ECHO_IOC_SET_STRIPE:
1270                 if (!capable (CAP_SYS_ADMIN))
1271                         GOTO (out, rc = -EPERM);
1272
1273                 if (data->ioc_pbuf1 == NULL) {  /* unset */
1274                         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1275                         if (rc == 0) {
1276                                 eco->eco_deleted = 1;
1277                                 echo_put_object(eco);
1278                         }
1279                 } else {
1280                         rc = echo_create_object(obd, 0, &data->ioc_obdo1,
1281                                                 data->ioc_pbuf1,
1282                                                 data->ioc_plen1, &dummy_oti);
1283                 }
1284                 GOTO (out, rc);
1285
1286         case ECHO_IOC_ENQUEUE:
1287                 if (!capable (CAP_SYS_ADMIN))
1288                         GOTO (out, rc = -EPERM);
1289
1290                 rc = echo_client_enqueue(exp, &data->ioc_obdo1,
1291                                          data->ioc_conn1, /* lock mode */
1292                                    data->ioc_offset, data->ioc_count);/*extent*/
1293                 GOTO (out, rc);
1294
1295         case ECHO_IOC_CANCEL:
1296                 rc = echo_client_cancel(exp, &data->ioc_obdo1);
1297                 GOTO (out, rc);
1298
1299         default:
1300                 CERROR ("echo_ioctl(): unrecognised ioctl %#x\n", cmd);
1301                 GOTO (out, rc = -ENOTTY);
1302         }
1303
1304         EXIT;
1305  out:
1306
1307         /* XXX this should be in a helper also called by target_send_reply */
1308         for (ack_lock = dummy_oti.oti_ack_locks, i = 0; i < 4;
1309              i++, ack_lock++) {
1310                 if (!ack_lock->mode)
1311                         break;
1312                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1313         }
1314
1315         return rc;
1316 }
1317
1318 static int
1319 echo_client_setup(struct obd_device *obddev, obd_count len, void *buf)
1320 {
1321         struct lustre_cfg* lcfg = buf;
1322         struct echo_client_obd *ec = &obddev->u.echo_client;
1323         struct obd_device *tgt;
1324         struct lustre_handle conn = {0, };
1325         struct obd_uuid echo_uuid = { "ECHO_UUID" };
1326         int rc;
1327         ENTRY;
1328
1329         if (lcfg->lcfg_inllen1 < 1) {
1330                 CERROR("requires a TARGET OBD name\n");
1331                 RETURN(-EINVAL);
1332         }
1333
1334         tgt = class_name2obd(lcfg->lcfg_inlbuf1);
1335         if (!tgt || !tgt->obd_attached || !tgt->obd_set_up) {
1336                 CERROR("device not attached or not set up (%s)\n",
1337                        lcfg->lcfg_inlbuf1);
1338                 RETURN(-EINVAL);
1339         }
1340
1341         spin_lock_init (&ec->ec_lock);
1342         INIT_LIST_HEAD (&ec->ec_objects);
1343         ec->ec_unique = 0;
1344
1345         rc = obd_connect(&conn, tgt, &echo_uuid);
1346         if (rc) {
1347                 CERROR("fail to connect to device %s\n", lcfg->lcfg_inlbuf1);
1348                 return (rc);
1349         }
1350         ec->ec_exp = class_conn2export(&conn);
1351
1352         RETURN(rc);
1353 }
1354
1355 static int echo_client_cleanup(struct obd_device *obddev, int flags)
1356 {
1357         struct list_head       *el;
1358         struct ec_object       *eco;
1359         struct echo_client_obd *ec = &obddev->u.echo_client;
1360         int rc;
1361         ENTRY;
1362
1363         if (!list_empty(&obddev->obd_exports)) {
1364                 CERROR("still has clients!\n");
1365                 RETURN(-EBUSY);
1366         }
1367
1368         /* XXX assuming sole access */
1369         while (!list_empty(&ec->ec_objects)) {
1370                 el = ec->ec_objects.next;
1371                 eco = list_entry(el, struct ec_object, eco_obj_chain);
1372
1373                 LASSERT(eco->eco_refcount == 0);
1374                 eco->eco_refcount = 1;
1375                 eco->eco_deleted = 1;
1376                 echo_put_object(eco);
1377         }
1378
1379         rc = obd_disconnect(ec->ec_exp, 0);
1380         if (rc != 0)
1381                 CERROR("fail to disconnect device: %d\n", rc);
1382
1383         RETURN(rc);
1384 }
1385
1386 static int echo_client_connect(struct lustre_handle *conn,
1387                                struct obd_device *src, struct obd_uuid *cluuid)
1388 {
1389         struct obd_export *exp;
1390         int                rc;
1391
1392         rc = class_connect(conn, src, cluuid);
1393         if (rc == 0) {
1394                 exp = class_conn2export(conn);
1395                 INIT_LIST_HEAD(&exp->exp_ec_data.eced_locks);
1396                 class_export_put(exp);
1397         }
1398
1399         RETURN (rc);
1400 }
1401
1402 static int echo_client_disconnect(struct obd_export *exp, int flags)
1403 {
1404         struct obd_device      *obd;
1405         struct echo_client_obd *ec;
1406         struct ec_lock         *ecl;
1407         int                     rc;
1408         ENTRY;
1409
1410         if (exp == NULL)
1411                 GOTO(out, rc = -EINVAL);
1412
1413         obd = exp->exp_obd;
1414         ec = &obd->u.echo_client;
1415
1416         /* no more contention on export's lock list */
1417         while (!list_empty (&exp->exp_ec_data.eced_locks)) {
1418                 ecl = list_entry (exp->exp_ec_data.eced_locks.next,
1419                                   struct ec_lock, ecl_exp_chain);
1420                 list_del (&ecl->ecl_exp_chain);
1421
1422                 rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm,
1423                                  ecl->ecl_mode, &ecl->ecl_lock_handle);
1424
1425                 CDEBUG (D_INFO, "Cancel lock on object "LPX64" on disconnect "
1426                         "(%d)\n", ecl->ecl_object->eco_id, rc);
1427
1428                 echo_put_object (ecl->ecl_object);
1429                 OBD_FREE (ecl, sizeof (*ecl));
1430         }
1431
1432         rc = class_disconnect(exp, 0);
1433         GOTO(out, rc);
1434  out:
1435         return rc;
1436 }
1437
1438 static struct obd_ops echo_obd_ops = {
1439         o_owner:       THIS_MODULE,
1440         o_setup:       echo_client_setup,
1441         o_cleanup:     echo_client_cleanup,
1442         o_iocontrol:   echo_client_iocontrol,
1443         o_connect:     echo_client_connect,
1444         o_disconnect:  echo_client_disconnect
1445 };
1446
1447 int echo_client_init(void)
1448 {
1449         struct lprocfs_static_vars lvars;
1450
1451         lprocfs_init_vars(echo, &lvars);
1452         return class_register_type(&echo_obd_ops, lvars.module_vars,
1453                                    OBD_ECHO_CLIENT_DEVICENAME);
1454 }
1455
1456 void echo_client_exit(void)
1457 {
1458         class_unregister_type(OBD_ECHO_CLIENT_DEVICENAME);
1459 }