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  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_ECHO
  38 #ifdef __KERNEL__
  39 #include <libcfs/libcfs.h>
  40 #else
  41 #include <liblustre.h>
  42 #endif
  43
  44 #include <obd.h>
  45 #include <obd_support.h>
  46 #include <obd_class.h>
  47 #include <obd_echo.h>
  48 #include <lustre_debug.h>
  49 #include <lprocfs_status.h>
  50
  51 static obd_id last_object_id;
  52
  53 #if 0
  54 static void
  55 echo_printk_object (char *msg, struct ec_object *eco)
  56 {
  57         struct lov_stripe_md *lsm = eco->eco_lsm;
  58         int                   i;
  59
  60         printk (KERN_INFO "Lustre: %s: object %p: "LPX64", refs %d%s: "LPX64
  61                 "=%u!%u\n", msg, eco, eco->eco_id, eco->eco_refcount,
  62                 eco->eco_deleted ? "(deleted) " : "",
  63                 lsm->lsm_object_id, lsm->lsm_stripe_size,
  64                 lsm->lsm_stripe_count);
  65
  66         for (i = 0; i < lsm->lsm_stripe_count; i++)
  67                 printk (KERN_INFO "Lustre:   @%2u:"LPX64"\n",
  68                         lsm->lsm_oinfo[i].loi_ost_idx,
  69                         lsm->lsm_oinfo[i].loi_id);
  70 }
  71 #endif
  72
  73 static struct ec_object *
  74 echo_find_object_locked (struct obd_device *obd, obd_id id)
  75 {
  76         struct echo_client_obd *ec = &obd->u.echo_client;
  77         struct ec_object       *eco = NULL;
  78         struct list_head       *el;
  79
  80         list_for_each (el, &ec->ec_objects) {
  81                 eco = list_entry (el, struct ec_object, eco_obj_chain);
  82
  83                 if (eco->eco_id == id)
  84                         return (eco);
  85         }
  86         return (NULL);
  87 }
  88
  89 static int
  90 echo_copyout_lsm (struct lov_stripe_md *lsm, void *_ulsm, int ulsm_nob)
  91 {
  92         struct lov_stripe_md *ulsm = _ulsm;
  93         int nob, i;
  94
  95         nob = offsetof (struct lov_stripe_md, lsm_oinfo[lsm->lsm_stripe_count]);
  96         if (nob > ulsm_nob)
  97                 return (-EINVAL);
  98
  99         if (copy_to_user (ulsm, lsm, sizeof(ulsm)))
 100                 return (-EFAULT);
 101
 102         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 103                 if (copy_to_user (ulsm->lsm_oinfo[i], lsm->lsm_oinfo[i],
 104                                   sizeof(lsm->lsm_oinfo[0])))
 105                         return (-EFAULT);
 106         }
 107         return (0);
 108 }
 109
 110 static int
 111 echo_copyin_lsm (struct obd_device *obd, struct lov_stripe_md *lsm,
 112                  void *ulsm, int ulsm_nob)
 113 {
 114         struct echo_client_obd *ec = &obd->u.echo_client;
 115         int                     i;
 116
 117         if (ulsm_nob < sizeof (*lsm))
 118                 return (-EINVAL);
 119
 120         if (copy_from_user (lsm, ulsm, sizeof (*lsm)))
 121                 return (-EFAULT);
 122
 123         if (lsm->lsm_stripe_count > ec->ec_nstripes ||
 124             lsm->lsm_magic != LOV_MAGIC ||
 125             (lsm->lsm_stripe_size & (~CFS_PAGE_MASK)) != 0 ||
 126             ((__u64)lsm->lsm_stripe_size * lsm->lsm_stripe_count > ~0UL))
 127                 return (-EINVAL);
 128
 129         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 130                 if (copy_from_user(lsm->lsm_oinfo[i],
 131                                    ((struct lov_stripe_md *)ulsm)->lsm_oinfo[i],
 132                                    sizeof(lsm->lsm_oinfo[0])))
 133                         return (-EFAULT);
 134         }
 135
 136         return (0);
 137 }
 138
 139 static struct ec_object *
 140 echo_allocate_object (struct obd_device *obd)
 141 {
 142         struct echo_client_obd *ec = &obd->u.echo_client;
 143         struct ec_object       *eco;
 144         int rc;
 145
 146         OBD_ALLOC(eco, sizeof (*eco));
 147         if (eco == NULL)
 148                 return NULL;
 149
 150         rc = obd_alloc_memmd(ec->ec_exp, &eco->eco_lsm);
 151         if (rc < 0) {
 152                 OBD_FREE(eco, sizeof (*eco));
 153                 return NULL;
 154         }
 155
 156         eco->eco_device = obd;
 157         eco->eco_deleted = 0;
 158         eco->eco_refcount = 0;
 159         eco->eco_lsm->lsm_magic = LOV_MAGIC;
 160         /* leave stripe count 0 by default */
 161
 162         return (eco);
 163 }
 164
 165 static void
 166 echo_free_object (struct ec_object *eco)
 167 {
 168         struct obd_device      *obd = eco->eco_device;
 169         struct echo_client_obd *ec = &obd->u.echo_client;
 170
 171         LASSERT (eco->eco_refcount == 0);
 172         if (!eco->eco_lsm)
 173                 CERROR("No object %s\n", obd->obd_name);
 174         else
 175                 obd_free_memmd(ec->ec_exp, &eco->eco_lsm);
 176         OBD_FREE (eco, sizeof (*eco));
 177 }
 178
 179 static int echo_create_object(struct obd_device *obd, int on_target,
 180                               struct obdo *oa, void *ulsm, int ulsm_nob,
 181                               struct obd_trans_info *oti)
 182 {
 183         struct echo_client_obd *ec = &obd->u.echo_client;
 184         struct ec_object       *eco2;
 185         struct ec_object       *eco;
 186         struct lov_stripe_md   *lsm;
 187         int                     rc;
 188         int                     i, idx;
 189
 190         if ((oa->o_valid & OBD_MD_FLID) == 0 && /* no obj id */
 191             (on_target ||                       /* set_stripe */
 192              ec->ec_nstripes != 0)) {           /* LOV */
 193                 CERROR ("No valid oid\n");
 194                 return (-EINVAL);
 195         }
 196
 197         if (ulsm != NULL) {
 198                 eco = echo_allocate_object (obd);
 199                 if (eco == NULL)
 200                         return (-ENOMEM);
 201
 202                 lsm = eco->eco_lsm;
 203
 204                 rc = echo_copyin_lsm (obd, lsm, ulsm, ulsm_nob);
 205                 if (rc != 0)
 206                         goto failed;
 207
 208                 /* setup object ID here for !on_target and LOV hint */
 209                 if ((oa->o_valid & OBD_MD_FLID) != 0)
 210                         eco->eco_id = lsm->lsm_object_id = oa->o_id;
 211
 212                 if (lsm->lsm_stripe_count == 0)
 213                         lsm->lsm_stripe_count = ec->ec_nstripes;
 214
 215                 if (lsm->lsm_stripe_size == 0)
 216                         lsm->lsm_stripe_size = CFS_PAGE_SIZE;
 217
 218                 idx = ll_rand();
 219
 220                 /* setup stripes: indices + default ids if required */
 221                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 222                         if (lsm->lsm_oinfo[i]->loi_id == 0)
 223                                 lsm->lsm_oinfo[i]->loi_id = lsm->lsm_object_id;
 224
 225                         lsm->lsm_oinfo[i]->loi_ost_idx =
 226                                 (idx + i) % ec->ec_nstripes;
 227                 }
 228         } else {
 229                 OBD_ALLOC(eco, sizeof(*eco));
 230                 if (!eco)
 231                         return (-ENOMEM);
 232                 eco->eco_device = obd;
 233                 lsm = NULL;
 234         }
 235
 236         if (oa->o_id == 0)
 237                 oa->o_id = ++last_object_id;
 238
 239         if (on_target) {
 240                 oa->o_gr = FILTER_GROUP_ECHO;
 241                 oa->o_valid |= OBD_MD_FLGROUP;
 242
 243                 rc = obd_create(ec->ec_exp, oa, &lsm, oti);
 244                 if (rc != 0)
 245                         goto failed;
 246
 247                 /* See what object ID we were given */
 248                 eco->eco_id = oa->o_id = lsm->lsm_object_id;
 249                 oa->o_valid |= OBD_MD_FLID;
 250
 251                 LASSERT(eco->eco_lsm == NULL || eco->eco_lsm == lsm);
 252                 eco->eco_lsm = lsm;
 253         }
 254
 255         spin_lock (&ec->ec_lock);
 256
 257         eco2 = echo_find_object_locked (obd, oa->o_id);
 258         if (eco2 != NULL) {                     /* conflict */
 259                 spin_unlock (&ec->ec_lock);
 260
 261                 CERROR ("Can't create object id "LPX64": id already exists%s\n",
 262                         oa->o_id, on_target ? " (undoing create)" : "");
 263
 264                 if (on_target)
 265                         obd_destroy(ec->ec_exp, oa, lsm, oti, NULL);
 266
 267                 rc = -EEXIST;
 268                 goto failed;
 269         }
 270
 271         list_add (&eco->eco_obj_chain, &ec->ec_objects);
 272         spin_unlock (&ec->ec_lock);
 273         CDEBUG (D_INFO,
 274                 "created %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 275                 eco, eco->eco_id,
 276                 eco->eco_lsm->lsm_stripe_size,
 277                 eco->eco_lsm->lsm_stripe_count,
 278                 eco->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 279                 eco->eco_refcount, eco->eco_deleted);
 280         return (0);
 281
 282  failed:
 283         echo_free_object (eco);
 284         if (rc)
 285                 CERROR("%s: err %d on create\n", obd->obd_name, rc);
 286         return (rc);
 287 }
 288
 289 static int
 290 echo_get_object (struct ec_object **ecop, struct obd_device *obd,
 291                  struct obdo *oa)
 292 {
 293         struct echo_client_obd *ec = &obd->u.echo_client;
 294         struct ec_object       *eco;
 295         struct ec_object       *eco2;
 296         int                     rc;
 297
 298         if ((oa->o_valid & OBD_MD_FLID) == 0 ||
 299             oa->o_id == 0)                      /* disallow use of object id 0 */
 300         {
 301                 CERROR ("No valid oid\n");
 302                 return (-EINVAL);
 303         }
 304
 305         spin_lock (&ec->ec_lock);
 306         eco = echo_find_object_locked (obd, oa->o_id);
 307         if (eco != NULL) {
 308                 if (eco->eco_deleted) {           /* being deleted */
 309                         spin_unlock(&ec->ec_lock);/* (see comment in cleanup) */
 310                         return (-EAGAIN);
 311                 }
 312
 313                 eco->eco_refcount++;
 314                 spin_unlock (&ec->ec_lock);
 315                 *ecop = eco;
 316                 CDEBUG (D_INFO,
 317                         "found %p: "LPX64"=%u#%u@%u refs %d del %d\n",
 318                         eco, eco->eco_id,
 319                         eco->eco_lsm->lsm_stripe_size,
 320                         eco->eco_lsm->lsm_stripe_count,
 321                         eco->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 322                         eco->eco_refcount, eco->eco_deleted);
 323                 return (0);
 324         }
 325         spin_unlock (&ec->ec_lock);
 326
 327         if (ec->ec_nstripes != 0)               /* striping required */
 328                 return (-ENOENT);
 329
 330         eco = echo_allocate_object (obd);
 331         if (eco == NULL)
 332                 return (-ENOMEM);
 333
 334         eco->eco_id = eco->eco_lsm->lsm_object_id = oa->o_id;
 335
 336         spin_lock (&ec->ec_lock);
 337
 338         eco2 = echo_find_object_locked (obd, oa->o_id);
 339         if (eco2 == NULL) {                     /* didn't race */
 340                 list_add (&eco->eco_obj_chain, &ec->ec_objects);
 341                 spin_unlock (&ec->ec_lock);
 342                 eco->eco_refcount = 1;
 343                 *ecop = eco;
 344                 CDEBUG (D_INFO,
 345                         "created %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 346                         eco, eco->eco_id,
 347                         eco->eco_lsm->lsm_stripe_size,
 348                         eco->eco_lsm->lsm_stripe_count,
 349                         eco->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 350                         eco->eco_refcount, eco->eco_deleted);
 351                 return (0);
 352         }
 353
 354         if (eco2->eco_deleted)
 355                 rc = -EAGAIN;                   /* lose race */
 356         else {
 357                 eco2->eco_refcount++;           /* take existing */
 358                 *ecop = eco2;
 359                 rc = 0;
 360                 LASSERT (eco2->eco_id == eco2->eco_lsm->lsm_object_id);
 361                 CDEBUG (D_INFO,
 362                         "found(2) %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 363                         eco2, eco2->eco_id,
 364                         eco2->eco_lsm->lsm_stripe_size,
 365                         eco2->eco_lsm->lsm_stripe_count,
 366                         eco2->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 367                         eco2->eco_refcount, eco2->eco_deleted);
 368         }
 369
 370         spin_unlock (&ec->ec_lock);
 371
 372         echo_free_object (eco);
 373         return (rc);
 374 }
 375
 376 static void
 377 echo_put_object (struct ec_object *eco)
 378 {
 379         struct obd_device      *obd = eco->eco_device;
 380         struct echo_client_obd *ec = &obd->u.echo_client;
 381
 382         /* Release caller's ref on the object.
 383          * delete => mark for deletion when last ref goes
 384          */
 385
 386         spin_lock (&ec->ec_lock);
 387
 388         eco->eco_refcount--;
 389         LASSERT (eco->eco_refcount >= 0);
 390
 391         CDEBUG(D_INFO, "put %p: "LPX64"=%u#%u@%d refs %d del %d\n",
 392                eco, eco->eco_id,
 393                eco->eco_lsm->lsm_stripe_size,
 394                eco->eco_lsm->lsm_stripe_count,
 395                eco->eco_lsm->lsm_oinfo[0]->loi_ost_idx,
 396                eco->eco_refcount, eco->eco_deleted);
 397
 398         if (eco->eco_refcount != 0 || !eco->eco_deleted) {
 399                 spin_unlock (&ec->ec_lock);
 400                 return;
 401         }
 402
 403         spin_unlock (&ec->ec_lock);
 404
 405         /* NB leave obj in the object list.  We must prevent anyone from
 406          * attempting to enqueue on this object number until we can be
 407          * sure there will be no more lock callbacks.
 408          */
 409         obd_cancel_unused(ec->ec_exp, eco->eco_lsm, 0, NULL);
 410
 411         /* now we can let it go */
 412         spin_lock (&ec->ec_lock);
 413         list_del (&eco->eco_obj_chain);
 414         spin_unlock (&ec->ec_lock);
 415
 416         LASSERT (eco->eco_refcount == 0);
 417
 418         echo_free_object (eco);
 419 }
 420
 421 static void
 422 echo_get_stripe_off_id (struct lov_stripe_md *lsm, obd_off *offp, obd_id *idp)
 423 {
 424         unsigned long stripe_count;
 425         unsigned long stripe_size;
 426         unsigned long width;
 427         unsigned long woffset;
 428         int           stripe_index;
 429         obd_off       offset;
 430
 431         if (lsm->lsm_stripe_count <= 1)
 432                 return;
 433
 434         offset       = *offp;
 435         stripe_size  = lsm->lsm_stripe_size;
 436         stripe_count = lsm->lsm_stripe_count;
 437
 438         /* width = # bytes in all stripes */
 439         width = stripe_size * stripe_count;
 440
 441         /* woffset = offset within a width; offset = whole number of widths */
 442         woffset = do_div (offset, width);
 443
 444         stripe_index = woffset / stripe_size;
 445
 446         *idp = lsm->lsm_oinfo[stripe_index]->loi_id;
 447         *offp = offset * stripe_size + woffset % stripe_size;
 448 }
 449
 450 static void
 451 echo_client_page_debug_setup(struct lov_stripe_md *lsm,
 452                              cfs_page_t *page, int rw, obd_id id,
 453                              obd_off offset, obd_off count)
 454 {
 455         char    *addr;
 456         obd_off  stripe_off;
 457         obd_id   stripe_id;
 458         int      delta;
 459
 460         /* no partial pages on the client */
 461         LASSERT(count == CFS_PAGE_SIZE);
 462
 463         addr = cfs_kmap(page);
 464
 465         for (delta = 0; delta < CFS_PAGE_SIZE; delta += OBD_ECHO_BLOCK_SIZE) {
 466                 if (rw == OBD_BRW_WRITE) {
 467                         stripe_off = offset + delta;
 468                         stripe_id = id;
 469                         echo_get_stripe_off_id(lsm, &stripe_off, &stripe_id);
 470                 } else {
 471                         stripe_off = 0xdeadbeef00c0ffeeULL;
 472                         stripe_id = 0xdeadbeef00c0ffeeULL;
 473                 }
 474                 block_debug_setup(addr + delta, OBD_ECHO_BLOCK_SIZE,
 475                                   stripe_off, stripe_id);
 476         }
 477
 478         cfs_kunmap(page);
 479 }
 480
 481 static int
 482 echo_client_page_debug_check(struct lov_stripe_md *lsm,
 483                              cfs_page_t *page, obd_id id,
 484                              obd_off offset, obd_off count)
 485 {
 486         obd_off stripe_off;
 487         obd_id  stripe_id;
 488         char   *addr;
 489         int     delta;
 490         int     rc;
 491         int     rc2;
 492
 493         /* no partial pages on the client */
 494         LASSERT(count == CFS_PAGE_SIZE);
 495
 496         addr = cfs_kmap(page);
 497
 498         for (rc = delta = 0; delta < CFS_PAGE_SIZE; delta += OBD_ECHO_BLOCK_SIZE) {
 499                 stripe_off = offset + delta;
 500                 stripe_id = id;
 501                 echo_get_stripe_off_id (lsm, &stripe_off, &stripe_id);
 502
 503                 rc2 = block_debug_check("test_brw",
 504                                         addr + delta, OBD_ECHO_BLOCK_SIZE,
 505                                         stripe_off, stripe_id);
 506                 if (rc2 != 0) {
 507                         CERROR ("Error in echo object "LPX64"\n", id);
 508                         rc = rc2;
 509                 }
 510         }
 511
 512         cfs_kunmap(page);
 513         return rc;
 514 }
 515
 516 static int echo_client_kbrw(struct obd_device *obd, int rw, struct obdo *oa,
 517                             struct lov_stripe_md *lsm, obd_off offset,
 518                             obd_size count, struct obd_trans_info *oti)
 519 {
 520         struct echo_client_obd *ec = &obd->u.echo_client;
 521         struct obd_info         oinfo = { { { 0 } } };
 522         obd_count               npages;
 523         struct ptlrpc_request_set *set = NULL;
 524         struct brw_page        *pga;
 525         struct brw_page        *pgp;
 526         obd_off                 off;
 527         int                     i;
 528         int                     rc;
 529         int                     verify;
 530         int                     gfp_mask;
 531         int                     brw_flags = 0;
 532
 533         verify = ((oa->o_id) != ECHO_PERSISTENT_OBJID &&
 534                   (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 535                   (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
 536
 537         gfp_mask = ((oa->o_id & 2) == 0) ? CFS_ALLOC_STD : CFS_ALLOC_HIGHUSER;
 538
 539         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 540         LASSERT(lsm != NULL);
 541         LASSERT(lsm->lsm_object_id == oa->o_id);
 542
 543         if (count <= 0 ||
 544             (count & (~CFS_PAGE_MASK)) != 0)
 545                 return (-EINVAL);
 546
 547         if (rw == OBD_BRW_WRITE)
 548                 brw_flags = OBD_BRW_ASYNC;
 549
 550         set =  ptlrpc_prep_set();
 551         if (set == NULL)
 552                 RETURN(-ENOMEM);
 553
 554         /* XXX think again with misaligned I/O */
 555         npages = count >> CFS_PAGE_SHIFT;
 556
 557         OBD_ALLOC(pga, npages * sizeof(*pga));
 558         if (pga == NULL)
 559                 return (-ENOMEM);
 560
 561         for (i = 0, pgp = pga, off = offset;
 562              i < npages;
 563              i++, pgp++, off += CFS_PAGE_SIZE) {
 564
 565                 LASSERT (pgp->pg == NULL);      /* for cleanup */
 566
 567                 rc = -ENOMEM;
 568                 OBD_PAGE_ALLOC(pgp->pg, gfp_mask);
 569                 if (pgp->pg == NULL)
 570                         goto out;
 571
 572                 pgp->count = CFS_PAGE_SIZE;
 573                 pgp->off = off;
 574                 pgp->flag = brw_flags;
 575
 576                 if (verify)
 577                         echo_client_page_debug_setup(lsm, pgp->pg, rw,
 578                                                      oa->o_id, off, pgp->count);
 579         }
 580
 581         oinfo.oi_oa = oa;
 582         oinfo.oi_md = lsm;
 583
 584         /* OST/filter device don't support o_brw_async ops, turn to o_brw ops */
 585         if (ec->ec_exp && ec->ec_exp->exp_obd &&
 586             OBT(ec->ec_exp->exp_obd) && OBP(ec->ec_exp->exp_obd, brw_async)) {
 587                 rc = obd_brw_async(rw, ec->ec_exp, &oinfo, npages, pga, oti,
 588                                    set, 0);
 589                 if (rc == 0) {
 590                         rc = ptlrpc_set_wait(set);
 591                         if (rc)
 592                                 CERROR("error from callback: rc = %d\n", rc);
 593                 }
 594         } else {
 595                 rc = obd_brw(rw, ec->ec_exp, &oinfo, npages, pga, oti);
 596         }
 597         if (rc)
 598                 CDEBUG_LIMIT(rc == -ENOSPC ? D_INODE : D_ERROR,
 599                              "error from obd_brw_async: rc = %d\n", rc);
 600         ptlrpc_set_destroy(set);
 601  out:
 602         if (rc != 0 || rw != OBD_BRW_READ)
 603                 verify = 0;
 604
 605         for (i = 0, pgp = pga; i < npages; i++, pgp++) {
 606                 if (pgp->pg == NULL)
 607                         continue;
 608
 609                 if (verify) {
 610                         int vrc;
 611                         vrc = echo_client_page_debug_check(lsm, pgp->pg, oa->o_id,
 612                                                            pgp->off, pgp->count);
 613                         if (vrc != 0 && rc == 0)
 614                                 rc = vrc;
 615                 }
 616                 OBD_PAGE_FREE(pgp->pg);
 617         }
 618         OBD_FREE(pga, npages * sizeof(*pga));
 619         return (rc);
 620 }
 621
 622 struct echo_async_state;
 623
 624 #define EAP_MAGIC 79277927
 625 struct echo_async_page {
 626         int                     eap_magic;
 627         cfs_page_t             *eap_page;
 628         void                    *eap_cookie;
 629         obd_off                 eap_off;
 630         struct echo_async_state *eap_eas;
 631         struct list_head        eap_item;
 632 };
 633
 634 #define EAP_FROM_COOKIE(c)                                                      \
 635         (LASSERT(((struct echo_async_page *)(c))->eap_magic == EAP_MAGIC),      \
 636          (struct echo_async_page *)(c))
 637
 638 struct echo_async_state {
 639         spinlock_t              eas_lock;
 640         obd_off                 eas_next_offset;
 641         obd_off                 eas_end_offset;
 642         int                     eas_in_flight;
 643         int                     eas_rc;
 644         cfs_waitq_t             eas_waitq;
 645         struct list_head        eas_avail;
 646         struct obdo             eas_oa;
 647         struct lov_stripe_md    *eas_lsm;
 648 };
 649
 650 static int eas_should_wake(struct echo_async_state *eas)
 651 {
 652         int rc = 0;
 653
 654         spin_lock(&eas->eas_lock);
 655         if (eas->eas_rc == 0 && !list_empty(&eas->eas_avail))
 656             rc = 1;
 657         spin_unlock(&eas->eas_lock);
 658         return rc;
 659 };
 660
 661 static int ec_ap_make_ready(void *data, int cmd)
 662 {
 663         /* our pages are issued ready */
 664         LBUG();
 665         return 0;
 666 }
 667 static int ec_ap_refresh_count(void *data, int cmd)
 668 {
 669         /* our pages are issued with a stable count */
 670         LBUG();
 671         return CFS_PAGE_SIZE;
 672 }
 673 static void ec_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 674 {
 675         struct echo_async_page *eap = EAP_FROM_COOKIE(data);
 676
 677         lustre_set_wire_obdo(oa, &eap->eap_eas->eas_oa);
 678 }
 679
 680 static int ec_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 681 {
 682         struct echo_async_page *eap = EAP_FROM_COOKIE(data);
 683         struct echo_async_state *eas;
 684
 685         eas = eap->eap_eas;
 686
 687         if (cmd == OBD_BRW_READ &&
 688             eas->eas_oa.o_id != ECHO_PERSISTENT_OBJID &&
 689             (eas->eas_oa.o_valid & OBD_MD_FLFLAGS) != 0 &&
 690             (eas->eas_oa.o_flags & OBD_FL_DEBUG_CHECK) != 0)
 691                 echo_client_page_debug_check(eas->eas_lsm, eap->eap_page,
 692                                              eas->eas_oa.o_id, eap->eap_off,
 693                                              CFS_PAGE_SIZE);
 694
 695         spin_lock(&eas->eas_lock);
 696         if (rc && !eas->eas_rc)
 697                 eas->eas_rc = rc;
 698         eas->eas_in_flight--;
 699         list_add(&eap->eap_item, &eas->eas_avail);
 700         cfs_waitq_signal(&eas->eas_waitq);
 701         spin_unlock(&eas->eas_lock);
 702         return 0;
 703 }
 704
 705 static struct obd_async_page_ops ec_async_page_ops = {
 706         .ap_make_ready =        ec_ap_make_ready,
 707         .ap_refresh_count =     ec_ap_refresh_count,
 708         .ap_fill_obdo =         ec_ap_fill_obdo,
 709         .ap_completion =        ec_ap_completion,
 710 };
 711
 712 static int echo_client_async_page(struct obd_export *exp, int rw,
 713                                    struct obdo *oa, struct lov_stripe_md *lsm,
 714                                    obd_off offset, obd_size count,
 715                                    obd_size batching)
 716 {
 717         obd_count npages, i;
 718         struct echo_async_page *eap;
 719         struct echo_async_state *eas;
 720         int rc = 0;
 721         struct echo_async_page **aps = NULL;
 722         int brw_flags = 0;
 723
 724         ENTRY;
 725 #if 0
 726         int                     verify;
 727         int                     gfp_mask;
 728
 729         verify = ((oa->o_id) != ECHO_PERSISTENT_OBJID &&
 730                   (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 731                   (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
 732
 733         gfp_mask = ((oa->o_id & 2) == 0) ? GFP_KERNEL : GFP_HIGHUSER;
 734 #endif
 735
 736         LASSERT(rw == OBD_BRW_WRITE || rw == OBD_BRW_READ);
 737
 738         if (count <= 0 ||
 739             (count & (~CFS_PAGE_MASK)) != 0 ||
 740             (lsm != NULL &&
 741              lsm->lsm_object_id != oa->o_id))
 742                 return (-EINVAL);
 743
 744         /* XXX think again with misaligned I/O */
 745         npages = batching >> CFS_PAGE_SHIFT;
 746
 747         OBD_ALLOC_PTR(eas);
 748         if (NULL == eas)
 749                 return(-ENOMEM);
 750
 751         if (rw == OBD_BRW_WRITE)
 752                 brw_flags = OBD_BRW_ASYNC;
 753
 754         memcpy(&eas->eas_oa, oa, sizeof(*oa));
 755         eas->eas_next_offset = offset;
 756         eas->eas_end_offset = offset + count;
 757         spin_lock_init(&eas->eas_lock);
 758         cfs_waitq_init(&eas->eas_waitq);
 759         eas->eas_in_flight = 0;
 760         eas->eas_rc = 0;
 761         eas->eas_lsm = lsm;
 762         CFS_INIT_LIST_HEAD(&eas->eas_avail);
 763
 764         OBD_ALLOC(aps, npages * sizeof aps[0]);
 765         if (aps == NULL)
 766                 GOTO(free_eas, rc = -ENOMEM);
 767
 768         /* prepare the group of pages that we're going to be keeping
 769          * in flight */
 770         for (i = 0; i < npages; i++) {
 771                 cfs_page_t *page;
 772                 OBD_PAGE_ALLOC(page, CFS_ALLOC_STD);
 773                 if (page == NULL)
 774                         GOTO(out, rc = -ENOMEM);
 775
 776                 OBD_ALLOC(eap, sizeof(*eap));
 777                 if (eap == NULL) {
 778                         OBD_PAGE_FREE(page);
 779                         GOTO(out, rc = -ENOMEM);
 780                 }
 781
 782                 eap->eap_magic = EAP_MAGIC;
 783                 eap->eap_page = page;
 784                 eap->eap_eas = eas;
 785                 list_add_tail(&eap->eap_item, &eas->eas_avail);
 786                 aps[i] = eap;
 787         }
 788
 789         /* first we spin queueing io and being woken by its completion */
 790         spin_lock(&eas->eas_lock);
 791         for(;;) {
 792                 int rc;
 793
 794                 /* sleep until we have a page to send */
 795                 spin_unlock(&eas->eas_lock);
 796                 rc = wait_event_interruptible(eas->eas_waitq,
 797                                               eas_should_wake(eas));
 798                 spin_lock(&eas->eas_lock);
 799                 if (rc && !eas->eas_rc)
 800                         eas->eas_rc = rc;
 801                 if (eas->eas_rc)
 802                         break;
 803                 if (list_empty(&eas->eas_avail))
 804                         continue;
 805                 eap = list_entry(eas->eas_avail.next, struct echo_async_page,
 806                                  eap_item);
 807                 list_del(&eap->eap_item);
 808                 spin_unlock(&eas->eas_lock);
 809
 810                 /* unbind the eap from its old page offset */
 811                 if (eap->eap_cookie != NULL) {
 812                         obd_teardown_async_page(exp, lsm, NULL,
 813                                                 eap->eap_cookie);
 814                         eap->eap_cookie = NULL;
 815                 }
 816
 817                 eas->eas_next_offset += CFS_PAGE_SIZE;
 818                 eap->eap_off = eas->eas_next_offset;
 819
 820                 rc = obd_prep_async_page(exp, lsm, NULL, eap->eap_page,
 821                                          eap->eap_off, &ec_async_page_ops,
 822                                          eap, &eap->eap_cookie,
 823                                          OBD_PAGE_NO_CACHE, NULL);
 824                 if (rc) {
 825                         spin_lock(&eas->eas_lock);
 826                         eas->eas_rc = rc;
 827                         break;
 828                 }
 829
 830                 if (oa->o_id != ECHO_PERSISTENT_OBJID &&
 831                     (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
 832                     (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0)
 833                         echo_client_page_debug_setup(lsm, eap->eap_page, rw,
 834                                                      oa->o_id,
 835                                                      eap->eap_off, CFS_PAGE_SIZE);
 836
 837                 /* always asserts urgent, which isn't quite right */
 838                 rc = obd_queue_async_io(exp, lsm, NULL, eap->eap_cookie,
 839                                         rw, 0, CFS_PAGE_SIZE, brw_flags,
 840                                         ASYNC_READY | ASYNC_URGENT |
 841                                         ASYNC_COUNT_STABLE);
 842                 spin_lock(&eas->eas_lock);
 843                 if (rc && !eas->eas_rc) {
 844                         eas->eas_rc = rc;
 845                         break;
 846                 }
 847                 eas->eas_in_flight++;
 848                 if (eas->eas_next_offset == eas->eas_end_offset)
 849                         break;
 850         }
 851
 852         /* still hold the eas_lock here.. */
 853
 854         /* now we just spin waiting for all the rpcs to complete */
 855         while(eas->eas_in_flight) {
 856                 spin_unlock(&eas->eas_lock);
 857                 wait_event_interruptible(eas->eas_waitq,
 858                                          eas->eas_in_flight == 0);
 859                 spin_lock(&eas->eas_lock);
 860         }
 861         spin_unlock(&eas->eas_lock);
 862
 863 out:
 864         if (aps != NULL) {
 865                 for (i = 0; i < npages; ++ i) {
 866                         cfs_page_t *page;
 867
 868                         eap = aps[i];
 869                         page = eap->eap_page;
 870                         if (eap->eap_cookie != NULL)
 871                                 obd_teardown_async_page(exp, lsm, NULL,
 872                                                         eap->eap_cookie);
 873                         OBD_FREE(eap, sizeof(*eap));
 874                         OBD_PAGE_FREE(page);
 875                 }
 876                 OBD_FREE(aps, npages * sizeof aps[0]);
 877         }
 878 free_eas:
 879         OBD_FREE_PTR(eas);
 880
 881         RETURN(rc);
 882 }
 883
 884 static int echo_client_prep_commit(struct obd_export *exp, int rw,
 885                                    struct obdo *oa, struct lov_stripe_md *lsm,
 886                                    obd_off offset, obd_size count,
 887                                    obd_size batch, struct obd_trans_info *oti)
 888 {
 889         struct obd_ioobj ioo;
 890         struct niobuf_local *lnb;
 891         struct niobuf_remote *rnb;
 892         obd_off off;
 893         obd_size npages, tot_pages;
 894         int i, ret = 0;
 895         ENTRY;
 896
 897         if (count <= 0 || (count & (~CFS_PAGE_MASK)) != 0 ||
 898             (lsm != NULL && lsm->lsm_object_id != oa->o_id))
 899                 RETURN(-EINVAL);
 900
 901         npages = batch >> CFS_PAGE_SHIFT;
 902         tot_pages = count >> CFS_PAGE_SHIFT;
 903
 904         OBD_ALLOC(lnb, npages * sizeof(struct niobuf_local));
 905         OBD_ALLOC(rnb, npages * sizeof(struct niobuf_remote));
 906
 907         if (lnb == NULL || rnb == NULL)
 908                 GOTO(out, ret = -ENOMEM);
 909
 910         obdo_to_ioobj(oa, &ioo);
 911
 912         off = offset;
 913
 914         for(; tot_pages; tot_pages -= npages) {
 915                 int lpages;
 916
 917                 if (tot_pages < npages)
 918                         npages = tot_pages;
 919
 920                 for (i = 0; i < npages; i++, off += CFS_PAGE_SIZE) {
 921                         rnb[i].offset = off;
 922                         rnb[i].len = CFS_PAGE_SIZE;
 923                 }
 924
 925                 ioo.ioo_bufcnt = npages;
 926                 oti->oti_transno = 0;
 927
 928                 lpages = npages;
 929                 ret = obd_preprw(rw, exp, oa, 1, &ioo, rnb, &lpages, lnb, oti);
 930                 if (ret != 0)
 931                         GOTO(out, ret);
 932                 LASSERT(lpages == npages);
 933
 934                 for (i = 0; i < lpages; i++) {
 935                         cfs_page_t *page = lnb[i].page;
 936
 937                         /* read past eof? */
 938                         if (page == NULL && lnb[i].rc == 0)
 939                                 continue;
 940
 941                         if (oa->o_id == ECHO_PERSISTENT_OBJID ||
 942                             (oa->o_valid & OBD_MD_FLFLAGS) == 0 ||
 943                             (oa->o_flags & OBD_FL_DEBUG_CHECK) == 0)
 944                                 continue;
 945
 946                         if (rw == OBD_BRW_WRITE)
 947                                 echo_client_page_debug_setup(lsm, page, rw,
 948                                                              oa->o_id,
 949                                                              rnb[i].offset,
 950                                                              rnb[i].len);
 951                         else
 952                                 echo_client_page_debug_check(lsm, page,
 953                                                              oa->o_id,
 954                                                              rnb[i].offset,
 955                                                              rnb[i].len);
 956                 }
 957
 958                 ret = obd_commitrw(rw, exp, oa, 1,&ioo,rnb,npages,lnb,oti,ret);
 959                 if (ret != 0)
 960                         GOTO(out, ret);
 961         }
 962
 963 out:
 964         if (lnb)
 965                 OBD_FREE(lnb, npages * sizeof(struct niobuf_local));
 966         if (rnb)
 967                 OBD_FREE(rnb, npages * sizeof(struct niobuf_remote));
 968         RETURN(ret);
 969 }
 970
 971 int echo_client_brw_ioctl(int rw, struct obd_export *exp,
 972                           struct obd_ioctl_data *data)
 973 {
 974         struct obd_device *obd = class_exp2obd(exp);
 975         struct echo_client_obd *ec = &obd->u.echo_client;
 976         struct obd_trans_info dummy_oti = { .oti_thread = NULL };
 977         struct ec_object *eco;
 978         int rc;
 979         ENTRY;
 980
 981         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
 982         if (rc)
 983                 RETURN(rc);
 984
 985         data->ioc_obdo1.o_valid &= ~OBD_MD_FLHANDLE;
 986         data->ioc_obdo1.o_valid |= OBD_MD_FLGROUP;
 987         data->ioc_obdo1.o_gr = FILTER_GROUP_ECHO;
 988
 989         switch((long)data->ioc_pbuf1) {
 990         case 1:
 991                 rc = echo_client_kbrw(obd, rw, &data->ioc_obdo1,
 992                                       eco->eco_lsm, data->ioc_offset,
 993                                       data->ioc_count, &dummy_oti);
 994                 break;
 995         case 2:
 996                 rc = echo_client_async_page(ec->ec_exp, rw, &data->ioc_obdo1,
 997                                            eco->eco_lsm, data->ioc_offset,
 998                                            data->ioc_count, data->ioc_plen1);
 999                 break;
1000         case 3:
1001                 rc = echo_client_prep_commit(ec->ec_exp, rw, &data->ioc_obdo1,
1002                                             eco->eco_lsm, data->ioc_offset,
1003                                             data->ioc_count, data->ioc_plen1,
1004                                             &dummy_oti);
1005                 break;
1006         default:
1007                 rc = -EINVAL;
1008         }
1009         echo_put_object(eco);
1010         RETURN(rc);
1011 }
1012
1013 static int
1014 echo_ldlm_callback (struct ldlm_lock *lock, struct ldlm_lock_desc *new,
1015                     void *data, int flag)
1016 {
1017         struct ec_object       *eco = (struct ec_object *)data;
1018         struct echo_client_obd *ec = &(eco->eco_device->u.echo_client);
1019         struct lustre_handle    lockh;
1020         struct list_head       *el;
1021         int                     found = 0;
1022         int                     rc;
1023
1024         ldlm_lock2handle (lock, &lockh);
1025
1026         /* #ifdef this out if we're not feeling paranoid */
1027         spin_lock (&ec->ec_lock);
1028         list_for_each (el, &ec->ec_objects) {
1029                 found = (eco == list_entry(el, struct ec_object,
1030                                            eco_obj_chain));
1031                 if (found)
1032                         break;
1033         }
1034         spin_unlock (&ec->ec_lock);
1035         LASSERT (found);
1036
1037         switch (flag) {
1038         case LDLM_CB_BLOCKING:
1039                 CDEBUG(D_INFO, "blocking callback on "LPX64", handle "LPX64"\n",
1040                        eco->eco_id, lockh.cookie);
1041                 rc = ldlm_cli_cancel (&lockh);
1042                 if (rc != ELDLM_OK)
1043                         CERROR ("ldlm_cli_cancel failed: %d\n", rc);
1044                 break;
1045
1046         case LDLM_CB_CANCELING:
1047                 CDEBUG(D_INFO, "cancel callback on "LPX64", handle "LPX64"\n",
1048                        eco->eco_id, lockh.cookie);
1049                 break;
1050
1051         default:
1052                 LBUG ();
1053         }
1054
1055         return (0);
1056 }
1057
1058 static int
1059 echo_client_enqueue(struct obd_export *exp, struct obdo *oa,
1060                     int mode, obd_off offset, obd_size nob)
1061 {
1062         struct obd_device      *obd = exp->exp_obd;
1063         struct echo_client_obd *ec = &obd->u.echo_client;
1064         struct lustre_handle   *ulh = &oa->o_handle;
1065         struct ldlm_enqueue_info einfo = { 0 };
1066         struct obd_info oinfo = { { { 0 } } };
1067         struct ec_object       *eco;
1068         struct ec_lock         *ecl;
1069         int                     rc;
1070
1071         if (!(mode == LCK_PR || mode == LCK_PW))
1072                 return -EINVAL;
1073
1074         if ((offset & (~CFS_PAGE_MASK)) != 0 ||
1075             (nob & (~CFS_PAGE_MASK)) != 0)
1076                 return -EINVAL;
1077
1078         rc = echo_get_object (&eco, obd, oa);
1079         if (rc != 0)
1080                 return rc;
1081
1082         rc = -ENOMEM;
1083         OBD_ALLOC (ecl, sizeof (*ecl));
1084         if (ecl == NULL)
1085                 goto failed_0;
1086
1087         ecl->ecl_mode = mode;
1088         ecl->ecl_object = eco;
1089         ecl->ecl_policy.l_extent.start = offset;
1090         ecl->ecl_policy.l_extent.end =
1091                 (nob == 0) ? ((obd_off) -1) : (offset + nob - 1);
1092
1093         einfo.ei_type = LDLM_EXTENT;
1094         einfo.ei_mode = mode;
1095         einfo.ei_cb_bl = echo_ldlm_callback;
1096         einfo.ei_cb_cp = ldlm_completion_ast;
1097         einfo.ei_cb_gl = NULL;
1098         einfo.ei_cbdata = eco;
1099
1100         oinfo.oi_policy = ecl->ecl_policy;
1101         oinfo.oi_lockh = &ecl->ecl_lock_handle;
1102         oinfo.oi_md = eco->eco_lsm;
1103         rc = obd_enqueue(ec->ec_exp, &oinfo, &einfo, NULL);
1104         if (rc != 0)
1105                 goto failed_1;
1106
1107         CDEBUG(D_INFO, "enqueue handle "LPX64"\n", ecl->ecl_lock_handle.cookie);
1108
1109         /* NB ecl takes object ref from echo_get_object() above */
1110         spin_lock(&ec->ec_lock);
1111
1112         list_add(&ecl->ecl_exp_chain, &exp->exp_ec_data.eced_locks);
1113         ulh->cookie = ecl->ecl_cookie = ec->ec_unique++;
1114
1115         spin_unlock(&ec->ec_lock);
1116
1117         oa->o_valid |= OBD_MD_FLHANDLE;
1118         return 0;
1119
1120  failed_1:
1121         OBD_FREE (ecl, sizeof (*ecl));
1122  failed_0:
1123         echo_put_object (eco);
1124         return (rc);
1125 }
1126
1127 static int
1128 echo_client_cancel(struct obd_export *exp, struct obdo *oa)
1129 {
1130         struct obd_device      *obd = exp->exp_obd;
1131         struct echo_client_obd *ec = &obd->u.echo_client;
1132         struct lustre_handle   *ulh = &oa->o_handle;
1133         struct ec_lock         *ecl = NULL;
1134         int                     found = 0;
1135         struct list_head       *el;
1136         int                     rc;
1137
1138         if ((oa->o_valid & OBD_MD_FLHANDLE) == 0)
1139                 return -EINVAL;
1140
1141         spin_lock (&ec->ec_lock);
1142
1143         list_for_each (el, &exp->exp_ec_data.eced_locks) {
1144                 ecl = list_entry (el, struct ec_lock, ecl_exp_chain);
1145                 found = (ecl->ecl_cookie == ulh->cookie);
1146                 if (found) {
1147                         list_del (&ecl->ecl_exp_chain);
1148                         break;
1149                 }
1150         }
1151
1152         spin_unlock (&ec->ec_lock);
1153
1154         if (!found)
1155                 return (-ENOENT);
1156
1157         rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm, ecl->ecl_mode,
1158                         &ecl->ecl_lock_handle, 0, 0);
1159
1160         echo_put_object (ecl->ecl_object);
1161         OBD_FREE (ecl, sizeof (*ecl));
1162
1163         return rc;
1164 }
1165
1166 static int
1167 echo_client_iocontrol(unsigned int cmd, struct obd_export *exp,
1168                       int len, void *karg, void *uarg)
1169 {
1170         struct obd_device      *obd;
1171         struct echo_client_obd *ec;
1172         struct ec_object       *eco;
1173         struct obd_ioctl_data  *data = karg;
1174         struct obd_trans_info   dummy_oti;
1175         struct oti_req_ack_lock *ack_lock;
1176         struct obdo            *oa;
1177         int                     rw = OBD_BRW_READ;
1178         int                     rc = 0;
1179         int                     i;
1180         ENTRY;
1181
1182         unlock_kernel();
1183
1184         memset(&dummy_oti, 0, sizeof(dummy_oti));
1185
1186         obd = exp->exp_obd;
1187         ec = &obd->u.echo_client;
1188
1189         switch (cmd) {
1190         case OBD_IOC_CREATE:                    /* may create echo object */
1191                 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
1192                         GOTO (out, rc = -EPERM);
1193
1194                 rc = echo_create_object (obd, 1, &data->ioc_obdo1,
1195                                          data->ioc_pbuf1, data->ioc_plen1,
1196                                          &dummy_oti);
1197                 GOTO(out, rc);
1198
1199         case OBD_IOC_DESTROY:
1200                 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
1201                         GOTO (out, rc = -EPERM);
1202                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1203                 if (rc == 0) {
1204                         oa = &data->ioc_obdo1;
1205                         oa->o_gr = FILTER_GROUP_ECHO;
1206                         oa->o_valid |= OBD_MD_FLGROUP;
1207                         rc = obd_destroy(ec->ec_exp, oa, eco->eco_lsm,
1208                                          &dummy_oti, NULL);
1209                         if (rc == 0)
1210                                 eco->eco_deleted = 1;
1211                         echo_put_object(eco);
1212                 }
1213                 GOTO(out, rc);
1214
1215         case OBD_IOC_GETATTR:
1216                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1217                 if (rc == 0) {
1218                         struct obd_info oinfo = { { { 0 } } };
1219                         oinfo.oi_md = eco->eco_lsm;
1220                         oinfo.oi_oa = &data->ioc_obdo1;
1221                         rc = obd_getattr(ec->ec_exp, &oinfo);
1222                         echo_put_object(eco);
1223                 }
1224                 GOTO(out, rc);
1225
1226         case OBD_IOC_SETATTR:
1227                 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
1228                         GOTO (out, rc = -EPERM);
1229
1230                 rc = echo_get_object (&eco, obd, &data->ioc_obdo1);
1231                 if (rc == 0) {
1232                         struct obd_info oinfo = { { { 0 } } };
1233                         oinfo.oi_oa = &data->ioc_obdo1;
1234                         oinfo.oi_md = eco->eco_lsm;
1235
1236                         rc = obd_setattr(ec->ec_exp, &oinfo, NULL);
1237                         echo_put_object(eco);
1238                 }
1239                 GOTO(out, rc);
1240
1241         case OBD_IOC_BRW_WRITE:
1242                 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
1243                         GOTO (out, rc = -EPERM);
1244
1245                 rw = OBD_BRW_WRITE;
1246                 /* fall through */
1247         case OBD_IOC_BRW_READ:
1248                 rc = echo_client_brw_ioctl(rw, exp, data);
1249                 GOTO(out, rc);
1250
1251         case ECHO_IOC_GET_STRIPE:
1252                 rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1253                 if (rc == 0) {
1254                         rc = echo_copyout_lsm(eco->eco_lsm, data->ioc_pbuf1,
1255                                               data->ioc_plen1);
1256                         echo_put_object(eco);
1257                 }
1258                 GOTO(out, rc);
1259
1260         case ECHO_IOC_SET_STRIPE:
1261                 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
1262                         GOTO (out, rc = -EPERM);
1263
1264                 if (data->ioc_pbuf1 == NULL) {  /* unset */
1265                         rc = echo_get_object(&eco, obd, &data->ioc_obdo1);
1266                         if (rc == 0) {
1267                                 eco->eco_deleted = 1;
1268                                 echo_put_object(eco);
1269                         }
1270                 } else {
1271                         rc = echo_create_object(obd, 0, &data->ioc_obdo1,
1272                                                 data->ioc_pbuf1,
1273                                                 data->ioc_plen1, &dummy_oti);
1274                 }
1275                 GOTO (out, rc);
1276
1277         case ECHO_IOC_ENQUEUE:
1278                 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
1279                         GOTO (out, rc = -EPERM);
1280
1281                 rc = echo_client_enqueue(exp, &data->ioc_obdo1,
1282                                          data->ioc_conn1, /* lock mode */
1283                                    data->ioc_offset, data->ioc_count);/*extent*/
1284                 GOTO (out, rc);
1285
1286         case ECHO_IOC_CANCEL:
1287                 rc = echo_client_cancel(exp, &data->ioc_obdo1);
1288                 GOTO (out, rc);
1289
1290         default:
1291                 CERROR ("echo_ioctl(): unrecognised ioctl %#x\n", cmd);
1292                 GOTO (out, rc = -ENOTTY);
1293         }
1294
1295         EXIT;
1296  out:
1297
1298         /* XXX this should be in a helper also called by target_send_reply */
1299         for (ack_lock = dummy_oti.oti_ack_locks, i = 0; i < 4;
1300              i++, ack_lock++) {
1301                 if (!ack_lock->mode)
1302                         break;
1303                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1304         }
1305
1306         lock_kernel();
1307
1308         return rc;
1309 }
1310
1311 static int
1312 echo_client_setup(struct obd_device *obddev, obd_count len, void *buf)
1313 {
1314         struct lustre_cfg* lcfg = buf;
1315         struct echo_client_obd *ec = &obddev->u.echo_client;
1316         struct obd_device *tgt;
1317         struct lustre_handle conn = {0, };
1318         struct obd_uuid echo_uuid = { "ECHO_UUID" };
1319         struct obd_connect_data *ocd = NULL;
1320         int rc;
1321         ENTRY;
1322
1323         if (lcfg->lcfg_bufcount < 2 || LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1324                 CERROR("requires a TARGET OBD name\n");
1325                 RETURN(-EINVAL);
1326         }
1327
1328         tgt = class_name2obd(lustre_cfg_string(lcfg, 1));
1329         if (!tgt || !tgt->obd_attached || !tgt->obd_set_up) {
1330                 CERROR("device not attached or not set up (%s)\n",
1331                        lustre_cfg_string(lcfg, 1));
1332                 RETURN(-EINVAL);
1333         }
1334
1335         spin_lock_init (&ec->ec_lock);
1336         CFS_INIT_LIST_HEAD (&ec->ec_objects);
1337         ec->ec_unique = 0;
1338
1339         ec->ec_exp = lustre_hash_lookup(tgt->obd_uuid_hash, &echo_uuid);
1340         if (ec->ec_exp)
1341                 RETURN(0);
1342
1343         OBD_ALLOC(ocd, sizeof(*ocd));
1344         if (ocd == NULL) {
1345                 CERROR("Can't alloc ocd connecting to %s\n",
1346                        lustre_cfg_string(lcfg, 1));
1347                 return -ENOMEM;
1348         }
1349         ocd->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_REQPORTAL;
1350         ocd->ocd_version = LUSTRE_VERSION_CODE;
1351
1352         if ((strncmp(tgt->obd_type->typ_name, LUSTRE_OSC_NAME,
1353                      strlen(LUSTRE_OSC_NAME)) == 0) ||
1354             (strncmp(tgt->obd_type->typ_name, LUSTRE_LOV_NAME,
1355                      strlen(LUSTRE_LOV_NAME)) == 0)) {
1356                 rc = obd_connect(&conn, tgt, &echo_uuid, ocd, &ec->ec_exp);
1357         } else {
1358                 rc = obd_connect(&conn, tgt, &echo_uuid, ocd, NULL);
1359                 if (rc == 0) {
1360                         ec->ec_exp = class_conn2export(&conn);
1361
1362                         /* Turn off pinger because it connects to tgt obd directly */
1363                         spin_lock(&tgt->obd_dev_lock);
1364                         list_del_init(&ec->ec_exp->exp_obd_chain_timed);
1365                         spin_unlock(&tgt->obd_dev_lock);
1366                 }
1367         }
1368
1369         OBD_FREE(ocd, sizeof(*ocd));
1370
1371         if (rc == -EALREADY && (strncmp(tgt->obd_type->typ_name,LUSTRE_OSC_NAME,
1372                                         strlen(LUSTRE_OSC_NAME)) == 0)) {
1373                 /* OSC obd forbid reconnect already connected import,
1374                  * so we hack creating another export here */
1375                 down_write(&tgt->u.cli.cl_sem);
1376                 rc = class_connect(&conn, tgt, &echo_uuid);
1377                 if (rc == 0) {
1378                         ++tgt->u.cli.cl_conn_count;
1379                         ec->ec_exp = class_conn2export(&conn);
1380                 }
1381                 up_write(&tgt->u.cli.cl_sem);
1382         }
1383
1384         if (rc != 0)
1385                 CERROR("fail to connect to device %s\n",
1386                        lustre_cfg_string(lcfg, 1));
1387
1388         RETURN(rc);
1389 }
1390
1391 static int echo_client_cleanup(struct obd_device *obddev)
1392 {
1393         struct list_head       *el;
1394         struct ec_object       *eco;
1395         struct echo_client_obd *ec = &obddev->u.echo_client;
1396         int rc;
1397         ENTRY;
1398
1399         if (!list_empty(&obddev->obd_exports)) {
1400                 CERROR("still has clients!\n");
1401                 RETURN(-EBUSY);
1402         }
1403
1404         /* XXX assuming sole access */
1405         while (!list_empty(&ec->ec_objects)) {
1406                 el = ec->ec_objects.next;
1407                 eco = list_entry(el, struct ec_object, eco_obj_chain);
1408
1409                 LASSERT(eco->eco_refcount == 0);
1410                 eco->eco_refcount = 1;
1411                 eco->eco_deleted = 1;
1412                 echo_put_object(eco);
1413         }
1414
1415         rc = obd_disconnect(ec->ec_exp);
1416         if (rc != 0)
1417                 CERROR("fail to disconnect device: %d\n", rc);
1418
1419         RETURN(rc);
1420 }
1421
1422 static int echo_client_connect(struct lustre_handle *conn,
1423                                struct obd_device *src, struct obd_uuid *cluuid,
1424                                struct obd_connect_data *data, void *localdata)
1425 {
1426         struct obd_export *exp;
1427         int                rc;
1428
1429         ENTRY;
1430         rc = class_connect(conn, src, cluuid);
1431         if (rc == 0) {
1432                 exp = class_conn2export(conn);
1433                 CFS_INIT_LIST_HEAD(&exp->exp_ec_data.eced_locks);
1434                 class_export_put(exp);
1435         }
1436
1437         RETURN (rc);
1438 }
1439
1440 static int echo_client_disconnect(struct obd_export *exp)
1441 {
1442         struct obd_device      *obd;
1443         struct echo_client_obd *ec;
1444         struct ec_lock         *ecl;
1445         int                     rc;
1446         ENTRY;
1447
1448         if (exp == NULL)
1449                 GOTO(out, rc = -EINVAL);
1450
1451         obd = exp->exp_obd;
1452         ec = &obd->u.echo_client;
1453
1454         /* no more contention on export's lock list */
1455         while (!list_empty (&exp->exp_ec_data.eced_locks)) {
1456                 ecl = list_entry (exp->exp_ec_data.eced_locks.next,
1457                                   struct ec_lock, ecl_exp_chain);
1458                 list_del (&ecl->ecl_exp_chain);
1459
1460                 rc = obd_cancel(ec->ec_exp, ecl->ecl_object->eco_lsm,
1461                                  ecl->ecl_mode, &ecl->ecl_lock_handle, 0, 0);
1462
1463                 CDEBUG (D_INFO, "Cancel lock on object "LPX64" on disconnect "
1464                         "(%d)\n", ecl->ecl_object->eco_id, rc);
1465
1466                 echo_put_object (ecl->ecl_object);
1467                 OBD_FREE (ecl, sizeof (*ecl));
1468         }
1469
1470         rc = class_disconnect(exp);
1471         GOTO(out, rc);
1472  out:
1473         return rc;
1474 }
1475
1476 static struct obd_ops echo_obd_ops = {
1477         .o_owner       = THIS_MODULE,
1478         .o_setup       = echo_client_setup,
1479         .o_cleanup     = echo_client_cleanup,
1480         .o_iocontrol   = echo_client_iocontrol,
1481         .o_connect     = echo_client_connect,
1482         .o_disconnect  = echo_client_disconnect
1483 };
1484
1485 int echo_client_init(void)
1486 {
1487         struct lprocfs_static_vars lvars = { 0 };
1488
1489         lprocfs_echo_init_vars(&lvars);
1490         return class_register_type(&echo_obd_ops, lvars.module_vars,
1491                                    LUSTRE_ECHO_CLIENT_NAME);
1492 }
1493
1494 void echo_client_exit(void)
1495 {
1496         class_unregister_type(LUSTRE_ECHO_CLIENT_NAME);
1497 }