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/proc_fs.h>
  37 #include <linux/init.h>
  38 #include <asm/unistd.h>
  39
  40 #define DEBUG_SUBSYSTEM S_ECHO
  41
  42 #include <linux/obd_support.h>
  43 #include <linux/obd_class.h>
  44 #include <linux/obd_echo.h>
  45 #include <linux/lustre_debug.h>
  46 #include <linux/lustre_dlm.h>
  47 #include <linux/lprocfs_status.h>
  48
  49 #define ECHO_INIT_OBJID      0x1000000000000000ULL
  50 #define ECHO_HANDLE_MAGIC    0xabcd0123fedc9876ULL
  51
  52 #define ECHO_PERSISTENT_PAGES (ECHO_PERSISTENT_SIZE/PAGE_SIZE)
  53 static struct page *echo_persistent_pages[ECHO_PERSISTENT_PAGES];
  54
  55 enum {
  56         LPROC_ECHO_READ_BYTES = 1,
  57         LPROC_ECHO_WRITE_BYTES = 2,
  58         LPROC_ECHO_LAST = LPROC_ECHO_WRITE_BYTES +1
  59 };
  60
  61 static int echo_connect(struct lustre_handle *conn, struct obd_device *obd,
  62                         struct obd_uuid *cluuid)
  63 {
  64         return class_connect(conn, obd, cluuid);
  65 }
  66
  67 static int echo_disconnect(struct obd_export *exp, int flags)
  68 {
  69         unsigned long irqflags;
  70
  71         LASSERT (exp != NULL);
  72
  73         ldlm_cancel_locks_for_export(exp);
  74
  75         spin_lock_irqsave(&exp->exp_lock, irqflags);
  76         exp->exp_flags = flags;
  77         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
  78
  79         return class_disconnect(exp, flags);
  80 }
  81
  82 static int echo_destroy_export(struct obd_export *exp)
  83 {
  84         ENTRY;
  85
  86         target_destroy_export(exp);
  87
  88         RETURN(0);
  89 }
  90
  91  static __u64 echo_next_id(struct obd_device *obddev)
  92 {
  93         obd_id id;
  94
  95         spin_lock(&obddev->u.echo.eo_lock);
  96         id = ++obddev->u.echo.eo_lastino;
  97         spin_unlock(&obddev->u.echo.eo_lock);
  98
  99         return id;
 100 }
 101
 102 int echo_create(struct obd_export *exp, struct obdo *oa,
 103                 struct lov_stripe_md **ea, struct obd_trans_info *oti)
 104 {
 105         struct obd_device *obd = class_exp2obd(exp);
 106
 107         if (!obd) {
 108                 CERROR("invalid client cookie "LPX64"\n",
 109                        exp->exp_handle.h_cookie);
 110                 return -EINVAL;
 111         }
 112
 113         if (!(oa->o_mode && S_IFMT)) {
 114                 CERROR("echo obd: no type!\n");
 115                 return -ENOENT;
 116         }
 117
 118         if (!(oa->o_valid & OBD_MD_FLTYPE)) {
 119                 CERROR("invalid o_valid %08x\n", oa->o_valid);
 120                 return -EINVAL;
 121         }
 122
 123         oa->o_id = echo_next_id(obd);
 124         oa->o_valid = OBD_MD_FLID;
 125         atomic_inc(&obd->u.echo.eo_create);
 126
 127         return 0;
 128 }
 129
 130 int echo_destroy(struct obd_export *exp, struct obdo *oa,
 131                  struct lov_stripe_md *ea, struct obd_trans_info *oti)
 132 {
 133         struct obd_device *obd = class_exp2obd(exp);
 134
 135         if (!obd) {
 136                 CERROR("invalid client cookie "LPX64"\n",
 137                        exp->exp_handle.h_cookie);
 138                 RETURN(-EINVAL);
 139         }
 140
 141         if (!(oa->o_valid & OBD_MD_FLID)) {
 142                 CERROR("obdo missing FLID valid flag: %08x\n", oa->o_valid);
 143                 RETURN(-EINVAL);
 144         }
 145
 146         if (oa->o_id > obd->u.echo.eo_lastino || oa->o_id < ECHO_INIT_OBJID) {
 147                 CERROR("bad destroy objid: "LPX64"\n", oa->o_id);
 148                 RETURN(-EINVAL);
 149         }
 150
 151         atomic_inc(&obd->u.echo.eo_destroy);
 152
 153         return 0;
 154 }
 155
 156 static int echo_getattr(struct obd_export *exp, struct obdo *oa,
 157                         struct lov_stripe_md *md)
 158 {
 159         struct obd_device *obd = class_exp2obd(exp);
 160         obd_id id = oa->o_id;
 161
 162         if (!obd) {
 163                 CERROR("invalid client cookie "LPX64"\n",
 164                        exp->exp_handle.h_cookie);
 165                 RETURN(-EINVAL);
 166         }
 167
 168         if (!(oa->o_valid & OBD_MD_FLID)) {
 169                 CERROR("obdo missing FLID valid flag: %08x\n", oa->o_valid);
 170                 RETURN(-EINVAL);
 171         }
 172
 173         obdo_cpy_md(oa, &obd->u.echo.oa, oa->o_valid);
 174         oa->o_id = id;
 175
 176         return 0;
 177 }
 178
 179 static int echo_setattr(struct obd_export *exp, struct obdo *oa,
 180                         struct lov_stripe_md *md, struct obd_trans_info *oti)
 181 {
 182         struct obd_device *obd = class_exp2obd(exp);
 183
 184         if (!obd) {
 185                 CERROR("invalid client cookie "LPX64"\n",
 186                        exp->exp_handle.h_cookie);
 187                 RETURN(-EINVAL);
 188         }
 189
 190         if (!(oa->o_valid & OBD_MD_FLID)) {
 191                 CERROR("obdo missing FLID valid flag: %08x\n", oa->o_valid);
 192                 RETURN(-EINVAL);
 193         }
 194
 195         memcpy(&obd->u.echo.oa, oa, sizeof(*oa));
 196
 197         atomic_inc(&obd->u.echo.eo_setattr);
 198
 199         return 0;
 200 }
 201
 202 static void
 203 echo_page_debug_setup(struct page *page, int rw, obd_id id,
 204                       __u64 offset, int len)
 205 {
 206         int   page_offset = offset & (PAGE_SIZE - 1);
 207         char *addr        = ((char *)kmap(page)) + page_offset;
 208
 209         if (len % OBD_ECHO_BLOCK_SIZE != 0)
 210                 CERROR("Unexpected block size %d\n", len);
 211
 212         while (len > 0) {
 213                 if (rw & OBD_BRW_READ)
 214                         block_debug_setup(addr, OBD_ECHO_BLOCK_SIZE,
 215                                           offset, id);
 216                 else
 217                         block_debug_setup(addr, OBD_ECHO_BLOCK_SIZE,
 218                                           0xecc0ecc0ecc0ecc0ULL,
 219                                           0xecc0ecc0ecc0ecc0ULL);
 220
 221                 addr   += OBD_ECHO_BLOCK_SIZE;
 222                 offset += OBD_ECHO_BLOCK_SIZE;
 223                 len    -= OBD_ECHO_BLOCK_SIZE;
 224         }
 225
 226         kunmap(page);
 227 }
 228
 229 static int
 230 echo_page_debug_check(struct page *page, obd_id id,
 231                       __u64 offset, int len)
 232 {
 233         int   page_offset = offset & (PAGE_SIZE - 1);
 234         char *addr        = ((char *)kmap(page)) + page_offset;
 235         int   rc          = 0;
 236         int   rc2;
 237
 238         if (len % OBD_ECHO_BLOCK_SIZE != 0)
 239                 CERROR("Unexpected block size %d\n", len);
 240
 241         while (len > 0) {
 242                 rc2 = block_debug_check("echo", addr, OBD_ECHO_BLOCK_SIZE,
 243                                         offset, id);
 244
 245                 if (rc2 != 0 && rc == 0)
 246                         rc = rc2;
 247
 248                 addr   += OBD_ECHO_BLOCK_SIZE;
 249                 offset += OBD_ECHO_BLOCK_SIZE;
 250                 len    -= OBD_ECHO_BLOCK_SIZE;
 251         }
 252
 253         kunmap(page);
 254
 255         return (rc);
 256 }
 257
 258 /* This allows us to verify that desc_private is passed unmolested */
 259 #define DESC_PRIV 0x10293847
 260
 261 int echo_preprw(int cmd, struct obd_export *export, struct obdo *oa,
 262                 int objcount, struct obd_ioobj *obj, int niocount,
 263                 struct niobuf_remote *nb, struct niobuf_local *res,
 264                 struct obd_trans_info *oti)
 265 {
 266         struct obd_device *obd;
 267         struct niobuf_local *r = res;
 268         int tot_bytes = 0;
 269         int rc = 0;
 270         int i;
 271         ENTRY;
 272
 273         obd = export->exp_obd;
 274         if (obd == NULL)
 275                 RETURN(-EINVAL);
 276
 277         /* Temp fix to stop falling foul of osc_announce_cached() */
 278         oa->o_valid &= ~(OBD_MD_FLBLOCKS | OBD_MD_FLGRANT);
 279
 280         memset(res, 0, sizeof(*res) * niocount);
 281
 282         CDEBUG(D_PAGE, "%s %d obdos with %d IOs\n",
 283                cmd == OBD_BRW_READ ? "reading" : "writing", objcount, niocount);
 284
 285         if (oti)
 286                 oti->oti_handle = (void *)DESC_PRIV;
 287
 288         for (i = 0; i < objcount; i++, obj++) {
 289                 int gfp_mask = (obj->ioo_id & 1) ? GFP_HIGHUSER : GFP_KERNEL;
 290                 int ispersistent = obj->ioo_id == ECHO_PERSISTENT_OBJID;
 291                 int j;
 292
 293                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, nb++, r++) {
 294
 295                         if (ispersistent &&
 296                             (nb->offset >> PAGE_SHIFT) < ECHO_PERSISTENT_PAGES) {
 297                                 r->page = echo_persistent_pages[nb->offset >>
 298                                                                 PAGE_SHIFT];
 299                                 /* Take extra ref so __free_pages() can be called OK */
 300                                 get_page (r->page);
 301                         } else {
 302                                 r->page = alloc_pages(gfp_mask, 0);
 303                                 if (r->page == NULL) {
 304                                         CERROR("can't get page %u/%u for id "
 305                                                LPU64"\n",
 306                                                j, obj->ioo_bufcnt, obj->ioo_id);
 307                                         GOTO(preprw_cleanup, rc = -ENOMEM);
 308                                 }
 309                         }
 310
 311                         tot_bytes += nb->len;
 312
 313                         atomic_inc(&obd->u.echo.eo_prep);
 314
 315                         r->offset = nb->offset;
 316                         r->len = nb->len;
 317                         LASSERT((r->offset & ~PAGE_MASK) + r->len <= PAGE_SIZE);
 318
 319                         CDEBUG(D_PAGE, "$$$$ get page %p @ "LPU64" for %d\n",
 320                                r->page, r->offset, r->len);
 321
 322                         if (cmd & OBD_BRW_READ)
 323                                 r->rc = r->len;
 324
 325                         if (!ispersistent)
 326                                 echo_page_debug_setup(r->page, cmd, obj->ioo_id,
 327                                                       r->offset, r->len);
 328                 }
 329         }
 330         if (cmd & OBD_BRW_READ)
 331                 lprocfs_counter_add(obd->obd_stats, LPROC_ECHO_READ_BYTES,
 332                                     tot_bytes);
 333         else
 334                 lprocfs_counter_add(obd->obd_stats, LPROC_ECHO_WRITE_BYTES,
 335                                     tot_bytes);
 336
 337         CDEBUG(D_PAGE, "%d pages allocated after prep\n",
 338                atomic_read(&obd->u.echo.eo_prep));
 339
 340         RETURN(0);
 341
 342 preprw_cleanup:
 343         /* It is possible that we would rather handle errors by  allow
 344          * any already-set-up pages to complete, rather than tearing them
 345          * all down again.  I believe that this is what the in-kernel
 346          * prep/commit operations do.
 347          */
 348         CERROR("cleaning up %ld pages (%d obdos)\n", (long)(r - res), objcount);
 349         while (r-- > res) {
 350                 kunmap(r->page);
 351                 /* NB if this is a persistent page, __free_pages will just
 352                  * lose the extra ref gained above */
 353                 __free_pages(r->page, 0);
 354                 atomic_dec(&obd->u.echo.eo_prep);
 355         }
 356         memset(res, 0, sizeof(*res) * niocount);
 357
 358         return rc;
 359 }
 360
 361 int echo_commitrw(int cmd, struct obd_export *export, struct obdo *oa,
 362                   int objcount, struct obd_ioobj *obj, int niocount,
 363                   struct niobuf_local *res, struct obd_trans_info *oti, int rc)
 364 {
 365         struct obd_device *obd;
 366         struct niobuf_local *r = res;
 367         int i, vrc = 0;
 368         ENTRY;
 369
 370         obd = export->exp_obd;
 371         if (obd == NULL)
 372                 RETURN(-EINVAL);
 373
 374         if (rc)
 375                 GOTO(commitrw_cleanup, rc);
 376
 377         if ((cmd & OBD_BRW_RWMASK) == OBD_BRW_READ) {
 378                 CDEBUG(D_PAGE, "reading %d obdos with %d IOs\n",
 379                        objcount, niocount);
 380         } else {
 381                 CDEBUG(D_PAGE, "writing %d obdos with %d IOs\n",
 382                        objcount, niocount);
 383         }
 384
 385         if (niocount && !r) {
 386                 CERROR("NULL res niobuf with niocount %d\n", niocount);
 387                 RETURN(-EINVAL);
 388         }
 389
 390         LASSERT(oti == NULL || oti->oti_handle == (void *)DESC_PRIV);
 391
 392         for (i = 0; i < objcount; i++, obj++) {
 393                 int verify = (rc == 0 &&
 394                               obj->ioo_id != ECHO_PERSISTENT_OBJID);
 395                 int j;
 396
 397                 for (j = 0 ; j < obj->ioo_bufcnt ; j++, r++) {
 398                         struct page *page = r->page;
 399                         void *addr;
 400
 401                         if (page == NULL) {
 402                                 CERROR("null page objid "LPU64":%p, buf %d/%d\n",
 403                                        obj->ioo_id, page, j, obj->ioo_bufcnt);
 404                                 GOTO(commitrw_cleanup, rc = -EFAULT);
 405                         }
 406
 407                         addr = kmap(page);
 408
 409                         CDEBUG(D_PAGE, "$$$$ use page %p, addr %p@"LPU64"\n",
 410                                r->page, addr, r->offset);
 411
 412                         if (verify) {
 413                                 vrc = echo_page_debug_check(page, obj->ioo_id,
 414                                                             r->offset, r->len);
 415                                 /* check all the pages always */
 416                                 if (vrc != 0 && rc == 0)
 417                                         rc = vrc;
 418                         }
 419
 420                         kunmap(page);
 421                         /* NB see comment above regarding persistent pages */
 422                         __free_pages(page, 0);
 423                         atomic_dec(&obd->u.echo.eo_prep);
 424                 }
 425         }
 426         CDEBUG(D_PAGE, "%d pages remain after commit\n",
 427                atomic_read(&obd->u.echo.eo_prep));
 428         RETURN(rc);
 429
 430 commitrw_cleanup:
 431         CERROR("cleaning up %ld pages (%d obdos)\n",
 432                niocount - (long)(r - res) - 1, objcount);
 433         while (++r < res + niocount) {
 434                 struct page *page = r->page;
 435
 436                 /* NB see comment above regarding persistent pages */
 437                 __free_pages(page, 0);
 438                 atomic_dec(&obd->u.echo.eo_prep);
 439         }
 440         return rc;
 441 }
 442
 443 static int echo_setup(struct obd_device *obd, obd_count len, void *buf)
 444 {
 445         struct lprocfs_static_vars lvars;
 446         ENTRY;
 447
 448         spin_lock_init(&obd->u.echo.eo_lock);
 449         obd->u.echo.eo_lastino = ECHO_INIT_OBJID;
 450
 451         obd->obd_namespace = ldlm_namespace_new("echo-tgt",
 452                                                 LDLM_NAMESPACE_SERVER);
 453         if (obd->obd_namespace == NULL) {
 454                 LBUG();
 455                 RETURN(-ENOMEM);
 456         }
 457
 458         lprocfs_init_vars(echo, &lvars);
 459         if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
 460             lprocfs_alloc_obd_stats(obd, LPROC_ECHO_LAST) == 0) {
 461                 lprocfs_counter_init(obd->obd_stats, LPROC_ECHO_READ_BYTES,
 462                                      LPROCFS_CNTR_AVGMINMAX,
 463                                      "read_bytes", "bytes");
 464                 lprocfs_counter_init(obd->obd_stats, LPROC_ECHO_WRITE_BYTES,
 465                                      LPROCFS_CNTR_AVGMINMAX,
 466                                      "write_bytes", "bytes");
 467         }
 468
 469         ptlrpc_init_client (LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
 470                             "echo_ldlm_cb_client", &obd->obd_ldlm_client);
 471         RETURN(0);
 472 }
 473
 474 static int echo_cleanup(struct obd_device *obd, int flags)
 475 {
 476         int leaked;
 477         ENTRY;
 478
 479         lprocfs_free_obd_stats(obd);
 480         lprocfs_obd_cleanup(obd);
 481
 482         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
 483
 484         leaked = atomic_read(&obd->u.echo.eo_prep);
 485         if (leaked != 0)
 486                 CERROR("%d prep/commitrw pages leaked\n", leaked);
 487
 488         RETURN(0);
 489 }
 490
 491 static struct obd_ops echo_obd_ops = {
 492         o_owner:           THIS_MODULE,
 493         o_connect:         echo_connect,
 494         o_disconnect:      echo_disconnect,
 495         o_destroy_export:  echo_destroy_export,
 496         o_create:          echo_create,
 497         o_destroy:         echo_destroy,
 498         o_getattr:         echo_getattr,
 499         o_setattr:         echo_setattr,
 500         o_preprw:          echo_preprw,
 501         o_commitrw:        echo_commitrw,
 502         o_setup:           echo_setup,
 503         o_cleanup:         echo_cleanup
 504 };
 505
 506 extern int echo_client_init(void);
 507 extern void echo_client_exit(void);
 508
 509 static void
 510 echo_persistent_pages_fini (void)
 511 {
 512         int     i;
 513
 514         for (i = 0; i < ECHO_PERSISTENT_PAGES; i++)
 515                 if (echo_persistent_pages[i] != NULL) {
 516                         __free_pages (echo_persistent_pages[i], 0);
 517                         echo_persistent_pages[i] = NULL;
 518                 }
 519 }
 520
 521 static int
 522 echo_persistent_pages_init (void)
 523 {
 524         struct page *pg;
 525         int          i;
 526
 527         for (i = 0; i < ECHO_PERSISTENT_PAGES; i++) {
 528                 int gfp_mask = (i < ECHO_PERSISTENT_PAGES/2) ?
 529                         GFP_KERNEL : GFP_HIGHUSER;
 530
 531                 pg = alloc_pages (gfp_mask, 0);
 532                 if (pg == NULL) {
 533                         echo_persistent_pages_fini ();
 534                         return (-ENOMEM);
 535                 }
 536
 537                 memset (kmap (pg), 0, PAGE_SIZE);
 538                 kunmap (pg);
 539
 540                 echo_persistent_pages[i] = pg;
 541         }
 542
 543         return (0);
 544 }
 545
 546 static int __init obdecho_init(void)
 547 {
 548         struct lprocfs_static_vars lvars;
 549         int rc;
 550
 551         printk(KERN_INFO "Lustre: Echo OBD driver; info@clusterfs.com\n");
 552
 553         LASSERT(PAGE_SIZE % OBD_ECHO_BLOCK_SIZE == 0);
 554
 555         lprocfs_init_vars(echo, &lvars);
 556
 557         rc = echo_persistent_pages_init ();
 558         if (rc != 0)
 559                 goto failed_0;
 560
 561         rc = class_register_type(&echo_obd_ops, lvars.module_vars,
 562                                  OBD_ECHO_DEVICENAME);
 563         if (rc != 0)
 564                 goto failed_1;
 565
 566         rc = echo_client_init();
 567         if (rc == 0)
 568                 RETURN (0);
 569
 570         class_unregister_type(OBD_ECHO_DEVICENAME);
 571  failed_1:
 572         echo_persistent_pages_fini ();
 573  failed_0:
 574         RETURN(rc);
 575 }
 576
 577 static void /*__exit*/ obdecho_exit(void)
 578 {
 579         echo_client_exit();
 580         class_unregister_type(OBD_ECHO_DEVICENAME);
 581         echo_persistent_pages_fini ();
 582 }
 583
 584 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 585 MODULE_DESCRIPTION("Lustre Testing Echo OBD driver");
 586 MODULE_LICENSE("GPL");
 587
 588 module_init(obdecho_init);
 589 module_exit(obdecho_exit);