lustre/obdecho/echo.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  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Andreas Dilger <adilger@clusterfs.com>
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #ifndef EXPORT_SYMTAB
  25 # define EXPORT_SYMTAB
  26 #endif
  27
  28 #include <linux/version.h>
  29 #include <linux/module.h>
  30 #include <linux/mm.h>
  31 #include <linux/highmem.h>
  32 #include <linux/fs.h>
  33 #include <linux/stat.h>
  34 #include <linux/sched.h>
  35 #include <linux/smp_lock.h>
  36 #include <linux/ext2_fs.h>
  37 #include <linux/quotaops.h>
  38 #include <linux/proc_fs.h>
  39 #include <linux/init.h>
  40 #include <asm/unistd.h>
  41
  42 #define DEBUG_SUBSYSTEM S_ECHO
  43
  44 #include <linux/obd_support.h>
  45 #include <linux/obd_class.h>
  46 #include <linux/obd_echo.h>
  47 #include <linux/lustre_debug.h>
  48 #include <linux/lustre_dlm.h>
  49 #include <linux/lprocfs_status.h>
  50
  51 #define ECHO_INIT_OBJID      0x1000000000000000ULL
  52 #define ECHO_HANDLE_MAGIC    0xabcd0123fedc9876ULL
  53
  54 #define ECHO_OBJECT0_NPAGES  16
  55 static struct page *echo_object0_pages[ECHO_OBJECT0_NPAGES];
  56
  57 enum {
  58         LPROC_ECHO_READ_BYTES = 1,
  59         LPROC_ECHO_WRITE_BYTES = 2,
  60         LPROC_ECHO_LAST = LPROC_ECHO_WRITE_BYTES +1
  61 };
  62
  63 static int echo_connect(struct lustre_handle *conn, struct obd_device *obd,
  64                         struct obd_uuid *cluuid)
  65 {
  66         return class_connect(conn, obd, cluuid);
  67 }
  68
  69 static int echo_disconnect(struct obd_export *exp, int flags)
  70 {
  71         unsigned long irqflags;
  72
  73         LASSERT (exp != NULL);
  74
  75         ldlm_cancel_locks_for_export(exp);
  76
  77         spin_lock_irqsave(&exp->exp_lock, irqflags);
  78         exp->exp_flags = flags;
  79         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
  80
  81         return class_disconnect(exp, flags);
  82 }
  83
  84 static int echo_destroy_export(struct obd_export *exp)
  85 {
  86         ENTRY;
  87
  88         target_destroy_export(exp);
  89
  90         RETURN(0);
  91 }
  92
  93 static __u64 echo_next_id(struct obd_device *obddev)
  94 {
  95         obd_id id;
  96
  97         spin_lock(&obddev->u.echo.eo_lock);
  98         id = ++obddev->u.echo.eo_lastino;
  99         spin_unlock(&obddev->u.echo.eo_lock);
 100
 101         return id;
 102 }
 103
 104 int echo_create(struct obd_export *exp, struct obdo *oa,
 105                 struct lov_stripe_md **ea, struct obd_trans_info *oti)
 106 {
 107         struct obd_device *obd = class_exp2obd(exp);
 108
 109         if (!obd) {
 110                 CERROR("invalid client cookie "LPX64"\n",
 111                        exp->exp_handle.h_cookie);
 112                 return -EINVAL;
 113         }
 114
 115         if (!(oa->o_mode && S_IFMT)) {
 116                 CERROR("echo obd: no type!\n");
 117                 return -ENOENT;
 118         }
 119
 120         if (!(oa->o_valid & OBD_MD_FLTYPE)) {
 121                 CERROR("invalid o_valid %08x\n", oa->o_valid);
 122                 return -EINVAL;
 123         }
 124
 125         oa->o_id = echo_next_id(obd);
 126         oa->o_valid = OBD_MD_FLID;
 127         atomic_inc(&obd->u.echo.eo_create);
 128
 129         return 0;
 130 }
 131
 132 int echo_destroy(struct obd_export *exp, struct obdo *oa,
 133                  struct lov_stripe_md *ea, struct obd_trans_info *oti)
 134 {
 135         struct obd_device *obd = class_exp2obd(exp);
 136
 137         if (!obd) {
 138                 CERROR("invalid client cookie "LPX64"\n",
 139                        exp->exp_handle.h_cookie);
 140                 RETURN(-EINVAL);
 141         }
 142
 143         if (!(oa->o_valid & OBD_MD_FLID)) {
 144                 CERROR("obdo missing FLID valid flag: %08x\n", oa->o_valid);
 145                 RETURN(-EINVAL);
 146         }
 147
 148         if (oa->o_id > obd->u.echo.eo_lastino || oa->o_id < ECHO_INIT_OBJID) {
 149                 CERROR("bad destroy objid: "LPX64"\n", oa->o_id);
 150                 RETURN(-EINVAL);
 151         }
 152
 153         atomic_inc(&obd->u.echo.eo_destroy);
 154
 155         return 0;
 156 }
 157
 158 static int echo_getattr(struct obd_export *exp, struct obdo *oa,
 159                         struct lov_stripe_md *md)
 160 {
 161         struct obd_device *obd = class_exp2obd(exp);
 162         obd_id id = oa->o_id;
 163
 164         if (!obd) {
 165                 CERROR("invalid client cookie "LPX64"\n",
 166                        exp->exp_handle.h_cookie);
 167                 RETURN(-EINVAL);
 168         }
 169
 170         if (!(oa->o_valid & OBD_MD_FLID)) {
 171                 CERROR("obdo missing FLID valid flag: %08x\n", oa->o_valid);
 172                 RETURN(-EINVAL);
 173         }
 174
 175         obdo_cpy_md(oa, &obd->u.echo.oa, oa->o_valid);
 176         oa->o_id = id;
 177
 178         return 0;
 179 }
 180
 181 static int echo_setattr(struct obd_export *exp, struct obdo *oa,
 182                         struct lov_stripe_md *md, struct obd_trans_info *oti)
 183 {
 184         struct obd_device *obd = class_exp2obd(exp);
 185
 186         if (!obd) {
 187                 CERROR("invalid client cookie "LPX64"\n",
 188                        exp->exp_handle.h_cookie);
 189                 RETURN(-EINVAL);
 190         }
 191
 192         if (!(oa->o_valid & OBD_MD_FLID)) {
 193                 CERROR("obdo missing FLID valid flag: %08x\n", oa->o_valid);
 194                 RETURN(-EINVAL);
 195         }
 196
 197         memcpy(&obd->u.echo.oa, oa, sizeof(*oa));
 198
 199         atomic_inc(&obd->u.echo.eo_setattr);
 200
 201         return 0;
 202 }
 203
 204 /* This allows us to verify that desc_private is passed unmolested */
 205 #define DESC_PRIV 0x10293847
 206
 207 int echo_preprw(int cmd, struct obd_export *export, struct obdo *oa,
 208                 int objcount, struct obd_ioobj *obj, int niocount,
 209                 struct niobuf_remote *nb, struct niobuf_local *res,
 210                 struct obd_trans_info *oti)
 211 {
 212         struct obd_device *obd;
 213         struct niobuf_local *r = res;
 214         int tot_bytes = 0;
 215         int rc = 0;
 216         int i;
 217         ENTRY;
 218
 219         obd = export->exp_obd;
 220         if (obd == NULL)
 221                 RETURN(-EINVAL);
 222
 223         /* Temp fix to stop falling foul of osc_announce_cached() */
 224         oa->o_valid &= ~(OBD_MD_FLBLOCKS | OBD_MD_FLGRANT);
 225
 226         memset(res, 0, sizeof(*res) * niocount);
 227
 228         CDEBUG(D_PAGE, "%s %d obdos with %d IOs\n",
 229                cmd == OBD_BRW_READ ? "reading" : "writing", objcount, niocount);
 230
 231         if (oti)
 232                 oti->oti_handle = (void *)DESC_PRIV;
 233
 234         for (i = 0; i < objcount; i++, obj++) {
 235                 int gfp_mask = (obj->ioo_id & 1) ? GFP_HIGHUSER : GFP_KERNEL;
 236                 int isobj0 = obj->ioo_id == 0;
 237                 int verify = !isobj0;
 238                 int j;
 239
 240                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, nb++, r++) {
 241
 242                         if (isobj0 &&
 243                             (nb->offset >> PAGE_SHIFT) < ECHO_OBJECT0_NPAGES) {
 244                                 r->page = echo_object0_pages[nb->offset >>
 245                                                              PAGE_SHIFT];
 246                                 /* Take extra ref so __free_pages() can be called OK */
 247                                 get_page (r->page);
 248                         } else {
 249                                 r->page = alloc_pages(gfp_mask, 0);
 250                                 if (r->page == NULL) {
 251                                         CERROR("can't get page %u/%u for id "
 252                                                LPU64"\n",
 253                                                j, obj->ioo_bufcnt, obj->ioo_id);
 254                                         GOTO(preprw_cleanup, rc = -ENOMEM);
 255                                 }
 256                         }
 257
 258                         tot_bytes += nb->len;
 259
 260                         atomic_inc(&obd->u.echo.eo_prep);
 261
 262                         r->offset = nb->offset;
 263                         r->len = nb->len;
 264                         LASSERT((r->offset & ~PAGE_MASK) + r->len <= PAGE_SIZE);
 265
 266                         CDEBUG(D_PAGE, "$$$$ get page %p @ "LPU64" for %d\n",
 267                                r->page, r->offset, r->len);
 268
 269                         if (cmd & OBD_BRW_READ) {
 270                                 r->rc = r->len;
 271                                 if (verify) {
 272                                         page_debug_setup(kmap (r->page), r->len,
 273                                                          r->offset,obj->ioo_id);
 274                                         kunmap (r->page);
 275                                 }
 276                                 r->rc = r->len;
 277                         } else {
 278                                 if (verify) {
 279                                         page_debug_setup(kmap (r->page), r->len,
 280                                                          0xecc0ecc0ecc0ecc0ULL,
 281                                                          0xecc0ecc0ecc0ecc0ULL);
 282                                         kunmap (r->page);
 283                                 }
 284                         }
 285                 }
 286         }
 287         if (cmd & OBD_BRW_READ)
 288                 lprocfs_counter_add(obd->obd_stats, LPROC_ECHO_READ_BYTES,
 289                                     tot_bytes);
 290         else
 291                 lprocfs_counter_add(obd->obd_stats, LPROC_ECHO_WRITE_BYTES,
 292                                     tot_bytes);
 293
 294         CDEBUG(D_PAGE, "%d pages allocated after prep\n",
 295                atomic_read(&obd->u.echo.eo_prep));
 296
 297         RETURN(0);
 298
 299 preprw_cleanup:
 300         /* It is possible that we would rather handle errors by  allow
 301          * any already-set-up pages to complete, rather than tearing them
 302          * all down again.  I believe that this is what the in-kernel
 303          * prep/commit operations do.
 304          */
 305         CERROR("cleaning up %ld pages (%d obdos)\n", (long)(r - res), objcount);
 306         while (r-- > res) {
 307                 kunmap(r->page);
 308                 /* NB if this is an 'object0' page, __free_pages will just
 309                  * lose the extra ref gained above */
 310                 __free_pages(r->page, 0);
 311                 atomic_dec(&obd->u.echo.eo_prep);
 312         }
 313         memset(res, 0, sizeof(*res) * niocount);
 314
 315         return rc;
 316 }
 317
 318 int echo_commitrw(int cmd, struct obd_export *export, struct obdo *oa,
 319                   int objcount, struct obd_ioobj *obj, int niocount,
 320                   struct niobuf_local *res, struct obd_trans_info *oti)
 321 {
 322         struct obd_device *obd;
 323         struct niobuf_local *r = res;
 324         int i, vrc = 0, rc = 0;
 325         ENTRY;
 326
 327         obd = export->exp_obd;
 328         if (obd == NULL)
 329                 RETURN(-EINVAL);
 330
 331         if ((cmd & OBD_BRW_RWMASK) == OBD_BRW_READ) {
 332                 CDEBUG(D_PAGE, "reading %d obdos with %d IOs\n",
 333                        objcount, niocount);
 334         } else {
 335                 CDEBUG(D_PAGE, "writing %d obdos with %d IOs\n",
 336                        objcount, niocount);
 337         }
 338
 339         if (niocount && !r) {
 340                 CERROR("NULL res niobuf with niocount %d\n", niocount);
 341                 RETURN(-EINVAL);
 342         }
 343
 344         LASSERT(oti == NULL || oti->oti_handle == (void *)DESC_PRIV);
 345
 346         for (i = 0; i < objcount; i++, obj++) {
 347                 int verify = obj->ioo_id != 0;
 348                 int j;
 349
 350                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, r++) {
 351                         struct page *page = r->page;
 352                         void *addr;
 353
 354                         if (!page || !(addr = kmap(page)) ||
 355                             !kern_addr_valid((unsigned long)addr)) {
 356
 357                                 CERROR("bad page objid "LPU64":%p, buf %d/%d\n",
 358                                        obj->ioo_id, page, j, obj->ioo_bufcnt);
 359                                 kunmap(page);
 360                                 GOTO(commitrw_cleanup, rc = -EFAULT);
 361                         }
 362
 363                         CDEBUG(D_PAGE, "$$$$ use page %p, addr %p@"LPU64"\n",
 364                                r->page, addr, r->offset);
 365
 366                         if (verify) {
 367                                 vrc = page_debug_check("echo", addr, r->len,
 368                                                        r->offset, obj->ioo_id);
 369                                 /* check all the pages always */
 370                                 if (vrc != 0 && rc == 0)
 371                                         rc = vrc;
 372                         }
 373
 374                         kunmap(page);
 375                         /* NB see comment above regarding object0 pages */
 376                         __free_pages(page, 0);
 377                         atomic_dec(&obd->u.echo.eo_prep);
 378                 }
 379         }
 380         CDEBUG(D_PAGE, "%d pages remain after commit\n",
 381                atomic_read(&obd->u.echo.eo_prep));
 382         RETURN(rc);
 383
 384 commitrw_cleanup:
 385         CERROR("cleaning up %ld pages (%d obdos)\n",
 386                niocount - (long)(r - res) - 1, objcount);
 387         while (++r < res + niocount) {
 388                 struct page *page = r->page;
 389
 390                 /* NB see comment above regarding object0 pages */
 391                 __free_pages(page, 0);
 392                 atomic_dec(&obd->u.echo.eo_prep);
 393         }
 394         return rc;
 395 }
 396
 397 static int echo_setup(struct obd_device *obddev, obd_count len, void *buf)
 398 {
 399         ENTRY;
 400
 401         spin_lock_init(&obddev->u.echo.eo_lock);
 402         obddev->u.echo.eo_lastino = ECHO_INIT_OBJID;
 403
 404         obddev->obd_namespace =
 405                 ldlm_namespace_new("echo-tgt", LDLM_NAMESPACE_SERVER);
 406         if (obddev->obd_namespace == NULL) {
 407                 LBUG();
 408                 RETURN(-ENOMEM);
 409         }
 410
 411         ptlrpc_init_client (LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
 412                             "echo_ldlm_cb_client", &obddev->obd_ldlm_client);
 413         RETURN(0);
 414 }
 415
 416 static int echo_cleanup(struct obd_device *obddev, int flags)
 417 {
 418         int     leaked;
 419         ENTRY;
 420
 421         ldlm_namespace_free(obddev->obd_namespace, flags & OBD_OPT_FORCE);
 422
 423         leaked = atomic_read(&obddev->u.echo.eo_prep);
 424         if (leaked != 0)
 425                 CERROR("%d prep/commitrw pages leaked\n", leaked);
 426
 427         RETURN(0);
 428 }
 429
 430 int echo_attach(struct obd_device *obd, obd_count len, void *data)
 431 {
 432         struct lprocfs_static_vars lvars;
 433         int rc;
 434
 435         lprocfs_init_vars(echo, &lvars);
 436         rc = lprocfs_obd_attach(obd, lvars.obd_vars);
 437         if (rc != 0)
 438                 return rc;
 439         rc = lprocfs_alloc_obd_stats(obd, LPROC_ECHO_LAST);
 440         if (rc != 0)
 441                 return rc;
 442
 443         lprocfs_counter_init(obd->obd_stats, LPROC_ECHO_READ_BYTES,
 444                              LPROCFS_CNTR_AVGMINMAX, "read_bytes", "bytes");
 445         lprocfs_counter_init(obd->obd_stats, LPROC_ECHO_WRITE_BYTES,
 446                              LPROCFS_CNTR_AVGMINMAX, "write_bytes", "bytes");
 447         return rc;
 448 }
 449
 450 int echo_detach(struct obd_device *dev)
 451 {
 452         lprocfs_free_obd_stats(dev);
 453         return lprocfs_obd_detach(dev);
 454 }
 455
 456 static struct obd_ops echo_obd_ops = {
 457         o_owner:           THIS_MODULE,
 458         o_attach:          echo_attach,
 459         o_detach:          echo_detach,
 460         o_connect:         echo_connect,
 461         o_disconnect:      echo_disconnect,
 462         o_destroy_export:  echo_destroy_export,
 463         o_create:          echo_create,
 464         o_destroy:         echo_destroy,
 465         o_getattr:         echo_getattr,
 466         o_setattr:         echo_setattr,
 467         o_preprw:          echo_preprw,
 468         o_commitrw:        echo_commitrw,
 469         o_setup:           echo_setup,
 470         o_cleanup:         echo_cleanup
 471 };
 472
 473 extern int echo_client_init(void);
 474 extern void echo_client_exit(void);
 475
 476 static void
 477 echo_object0_pages_fini (void)
 478 {
 479         int     i;
 480
 481         for (i = 0; i < ECHO_OBJECT0_NPAGES; i++)
 482                 if (echo_object0_pages[i] != NULL) {
 483                         __free_pages (echo_object0_pages[i], 0);
 484                         echo_object0_pages[i] = NULL;
 485                 }
 486 }
 487
 488 static int
 489 echo_object0_pages_init (void)
 490 {
 491         struct page *pg;
 492         int          i;
 493
 494         for (i = 0; i < ECHO_OBJECT0_NPAGES; i++) {
 495                 int gfp_mask = (i < ECHO_OBJECT0_NPAGES/2) ?
 496                         GFP_KERNEL : GFP_HIGHUSER;
 497
 498                 pg = alloc_pages (gfp_mask, 0);
 499                 if (pg == NULL) {
 500                         echo_object0_pages_fini ();
 501                         return (-ENOMEM);
 502                 }
 503
 504                 memset (kmap (pg), 0, PAGE_SIZE);
 505                 kunmap (pg);
 506
 507                 echo_object0_pages[i] = pg;
 508         }
 509
 510         return (0);
 511 }
 512
 513 static int __init obdecho_init(void)
 514 {
 515         struct lprocfs_static_vars lvars;
 516         int rc;
 517
 518         printk(KERN_INFO "Lustre: Echo OBD driver; info@clusterfs.com\n");
 519
 520         lprocfs_init_vars(echo, &lvars);
 521
 522         rc = echo_object0_pages_init ();
 523         if (rc != 0)
 524                 goto failed_0;
 525
 526         rc = class_register_type(&echo_obd_ops, lvars.module_vars,
 527                                  OBD_ECHO_DEVICENAME);
 528         if (rc != 0)
 529                 goto failed_1;
 530
 531         rc = echo_client_init();
 532         if (rc == 0)
 533                 RETURN (0);
 534
 535         class_unregister_type(OBD_ECHO_DEVICENAME);
 536  failed_1:
 537         echo_object0_pages_fini ();
 538  failed_0:
 539         RETURN(rc);
 540 }
 541
 542 static void /*__exit*/ obdecho_exit(void)
 543 {
 544         echo_client_exit();
 545         class_unregister_type(OBD_ECHO_DEVICENAME);
 546         echo_object0_pages_fini ();
 547 }
 548
 549 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 550 MODULE_DESCRIPTION("Lustre Testing Echo OBD driver");
 551 MODULE_LICENSE("GPL");
 552
 553 module_init(obdecho_init);
 554 module_exit(obdecho_exit);