lustre/obdclass/class_obd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
   5  *
   6  * This code is issued under the GNU General Public License.
   7  * See the file COPYING in this distribution
   8  *
   9  * These are the only exported functions, they provide some generic
  10  * infrastructure for managing object devices
  11  *
  12  * Object Devices Class Driver
  13  */
  14
  15 #define EXPORT_SYMTAB
  16 #include <linux/config.h> /* for CONFIG_PROC_FS */
  17 #include <linux/module.h>
  18 #include <linux/errno.h>
  19 #include <linux/kernel.h>
  20 #include <linux/major.h>
  21 #include <linux/sched.h>
  22 #include <linux/lp.h>
  23 #include <linux/slab.h>
  24 #include <linux/ioport.h>
  25 #include <linux/fcntl.h>
  26 #include <linux/delay.h>
  27 #include <linux/skbuff.h>
  28 #include <linux/proc_fs.h>
  29 #include <linux/fs.h>
  30 #include <linux/poll.h>
  31 #include <linux/init.h>
  32 #include <linux/list.h>
  33 #include <linux/highmem.h>
  34 #include <asm/io.h>
  35 #include <asm/ioctls.h>
  36 #include <asm/system.h>
  37 #include <asm/poll.h>
  38 #include <asm/uaccess.h>
  39 #include <linux/miscdevice.h>
  40
  41 #define DEBUG_SUBSYSTEM S_CLASS
  42
  43 #include <linux/obd_support.h>
  44 #include <linux/obd_class.h>
  45 #include <linux/lustre_debug.h>
  46 #include <linux/smp_lock.h>
  47 #include <linux/lprocfs_status.h>
  48 #include <portals/lib-types.h> /* for PTL_MD_MAX_IOV */
  49
  50 struct semaphore obd_conf_sem;   /* serialize configuration commands */
  51 struct obd_device obd_dev[MAX_OBD_DEVICES];
  52 struct list_head obd_types;
  53 unsigned long obd_memory;
  54
  55 /* The following are visible and mutable through /proc/sys/lustre/. */
  56 unsigned long obd_fail_loc;
  57 unsigned long obd_timeout = 100;
  58 char obd_recovery_upcall[128] = "/usr/lib/lustre/ha_assist";
  59
  60 extern struct obd_type *class_nm_to_type(char *nm);
  61
  62 /*  opening /dev/obd */
  63 static int obd_class_open(struct inode * inode, struct file * file)
  64 {
  65         ENTRY;
  66
  67         file->private_data = NULL;
  68         CDEBUG(D_IOCTL, "MOD_INC_USE for open: count = %d\n",
  69                atomic_read(&(THIS_MODULE)->uc.usecount));
  70         MOD_INC_USE_COUNT;
  71         RETURN(0);
  72 }
  73
  74 /*  closing /dev/obd */
  75 static int obd_class_release(struct inode * inode, struct file * file)
  76 {
  77         ENTRY;
  78
  79         // XXX drop lsm, connections here
  80         if (file->private_data)
  81                 file->private_data = NULL;
  82
  83         CDEBUG(D_IOCTL, "MOD_DEC_USE for close: count = %d\n",
  84                atomic_read(&(THIS_MODULE)->uc.usecount) - 1);
  85         MOD_DEC_USE_COUNT;
  86         RETURN(0);
  87 }
  88
  89
  90 static inline void obd_data2conn(struct lustre_handle *conn,
  91                                  struct obd_ioctl_data *data)
  92 {
  93         conn->addr = data->ioc_addr;
  94         conn->cookie = data->ioc_cookie;
  95 }
  96
  97
  98 static inline void obd_conn2data(struct obd_ioctl_data *data,
  99                                  struct lustre_handle *conn)
 100 {
 101         data->ioc_addr = conn->addr;
 102         data->ioc_cookie = conn->cookie;
 103 }
 104
 105 static void forcibly_detach_exports(struct obd_device *obd)
 106 {
 107         int rc;
 108         struct list_head *tmp, *n;
 109         struct lustre_handle fake_conn;
 110
 111         CDEBUG(D_IOCTL, "OBD device %d (%p) has exports, "
 112                "disconnecting them", obd->obd_minor, obd);
 113         list_for_each_safe(tmp, n, &obd->obd_exports) {
 114                 struct obd_export *exp = list_entry(tmp, struct obd_export,
 115                                                     exp_obd_chain);
 116                 fake_conn.addr = (__u64)(unsigned long)exp;
 117                 fake_conn.cookie = exp->exp_cookie;
 118                 rc = obd_disconnect(&fake_conn);
 119                 if (rc) {
 120                         CDEBUG(D_IOCTL, "disconnecting export %p failed: %d\n",
 121                                exp, rc);
 122                 } else {
 123                         CDEBUG(D_IOCTL, "export %p disconnected\n", exp);
 124                 }
 125         }
 126 }
 127
 128 /* to control /dev/obd */
 129 static int obd_class_ioctl (struct inode * inode, struct file * filp,
 130                             unsigned int cmd, unsigned long arg)
 131 {
 132         char *buf = NULL;
 133         int len = 0;
 134         struct obd_ioctl_data *data;
 135         struct obd_device *obd = filp->private_data;
 136
 137         struct lustre_handle conn;
 138         int rw = OBD_BRW_READ;
 139         int err = 0;
 140         int serialised = 0;
 141
 142         ENTRY;
 143
 144         switch (cmd)
 145         {
 146         case OBD_IOC_BRW_WRITE:
 147         case OBD_IOC_BRW_READ:
 148         case OBD_IOC_GETATTR:
 149                 break;
 150         default:
 151                 down(&obd_conf_sem);
 152                 serialised = 1;
 153                 break;
 154         }
 155
 156         if (!obd && cmd != OBD_IOC_DEVICE && cmd != TCGETS &&
 157             cmd != OBD_IOC_LIST &&
 158             cmd != OBD_IOC_NAME2DEV && cmd != OBD_IOC_NEWDEV) {
 159                 CERROR("OBD ioctl: No device\n");
 160                 GOTO(out, err=-EINVAL);
 161         }
 162         if (obd_ioctl_getdata(&buf, &len, (void *)arg)) {
 163                 CERROR("OBD ioctl: data error\n");
 164                 GOTO(out, err=-EINVAL);
 165         }
 166         data = (struct obd_ioctl_data *)buf;
 167
 168         switch (cmd) {
 169         case TCGETS:
 170                 GOTO(out, err=-EINVAL);
 171         case OBD_IOC_DEVICE: {
 172                 CDEBUG(D_IOCTL, "\n");
 173                 if (data->ioc_dev >= MAX_OBD_DEVICES || data->ioc_dev < 0) {
 174                         CERROR("OBD ioctl: DEVICE insufficient devices\n");
 175                         GOTO(out, err=-EINVAL);
 176                 }
 177                 CDEBUG(D_IOCTL, "device %d\n", data->ioc_dev);
 178
 179                 filp->private_data = &obd_dev[data->ioc_dev];
 180                 GOTO(out, err=0);
 181         }
 182
 183         case OBD_IOC_LIST: {
 184                 int i;
 185                 char *buf2 = data->ioc_bulk;
 186                 int remains = data->ioc_inllen1;
 187
 188                 if (!data->ioc_inlbuf1) {
 189                         CERROR("No buffer passed!\n");
 190                         GOTO(out, err=-EINVAL);
 191                 }
 192
 193
 194                 for (i = 0 ; i < MAX_OBD_DEVICES ; i++) {
 195                         int l;
 196                         char *status;
 197                         struct obd_device *obd = &obd_dev[i];
 198                         if (!obd->obd_type)
 199                                 continue;
 200                         if (obd->obd_flags & OBD_SET_UP)
 201                                 status = "UP";
 202                         else if (obd->obd_flags & OBD_ATTACHED)
 203                                 status = "AT";
 204                         else
 205                                 status = "-";
 206                         l = snprintf(buf2, remains, "%2d %s %s %s %s %d\n",
 207                                      i, status, obd->obd_type->typ_name,
 208                                      obd->obd_name, obd->obd_uuid, obd->obd_type->typ_refcnt);
 209                         buf2 +=l;
 210                         remains -=l;
 211                         if (remains <= 0) {
 212                                 CERROR("not enough space for device listing\n");
 213                                 break;
 214                         }
 215                 }
 216
 217                 err = copy_to_user((void *)arg, data, len);
 218                 if (err)
 219                         err = -EFAULT;
 220                 GOTO(out, err);
 221         }
 222
 223
 224         case OBD_IOC_NAME2DEV: {
 225                 /* Resolve a device name.  This does not change the
 226                  * currently selected device.
 227                  */
 228                 int dev;
 229
 230                 if (!data->ioc_inllen1 || !data->ioc_inlbuf1 ) {
 231                         CERROR("No name passed,!\n");
 232                         GOTO(out, err=-EINVAL);
 233                 }
 234                 if (data->ioc_inlbuf1[data->ioc_inllen1-1] !=0) {
 235                         CERROR("Name not nul terminated!\n");
 236                         GOTO(out, err=-EINVAL);
 237                 }
 238
 239                 CDEBUG(D_IOCTL, "device name %s\n", data->ioc_inlbuf1);
 240                 dev = class_name2dev(data->ioc_inlbuf1);
 241                 data->ioc_dev = dev;
 242                 if (dev == -1) {
 243                         CDEBUG(D_IOCTL, "No device for name %s!\n",
 244                                data->ioc_inlbuf1);
 245                         GOTO(out, err=-EINVAL);
 246                 }
 247
 248                 CDEBUG(D_IOCTL, "device name %s, dev %d\n", data->ioc_inlbuf1,
 249                        dev);
 250                 err = copy_to_user((void *)arg, data, sizeof(*data));
 251                 if (err)
 252                         err = -EFAULT;
 253                 GOTO(out, err);
 254         }
 255
 256         case OBD_IOC_UUID2DEV: {
 257                 /* Resolve a device uuid.  This does not change the
 258                  * currently selected device.
 259                  */
 260                 int dev;
 261
 262                 if (!data->ioc_inllen1 || !data->ioc_inlbuf1) {
 263                         CERROR("No UUID passed!\n");
 264                         GOTO(out, err=-EINVAL);
 265                 }
 266                 if (data->ioc_inlbuf1[data->ioc_inllen1-1] !=0) {
 267                         CERROR("Name not nul terminated!\n");
 268                         GOTO(out, err=-EINVAL);
 269                 }
 270
 271                 CDEBUG(D_IOCTL, "device name %s\n", data->ioc_inlbuf1);
 272                 dev = class_uuid2dev(data->ioc_inlbuf1);
 273                 data->ioc_dev = dev;
 274                 if (dev == -1) {
 275                         CDEBUG(D_IOCTL, "No device for name %s!\n",
 276                                data->ioc_inlbuf1);
 277                         GOTO(out, err=-EINVAL);
 278                 }
 279
 280                 CDEBUG(D_IOCTL, "device name %s, dev %d\n", data->ioc_inlbuf1,
 281                        dev);
 282                 err = copy_to_user((void *)arg, data, sizeof(*data));
 283                 if (err)
 284                         err = -EFAULT;
 285                 GOTO(out, err);
 286         }
 287
 288         case OBD_IOC_NEWDEV: {
 289                 int dev = -1;
 290                 int i;
 291
 292                 filp->private_data = NULL;
 293                 for (i = 0 ; i < MAX_OBD_DEVICES ; i++) {
 294                         struct obd_device *obd = &obd_dev[i];
 295                         if (!obd->obd_type) {
 296                                 filp->private_data = obd;
 297                                 dev = i;
 298                                 break;
 299                         }
 300                 }
 301
 302
 303                 data->ioc_dev = dev;
 304                 if (dev == -1)
 305                         GOTO(out, err=-EINVAL);
 306
 307                 err = copy_to_user((void *)arg, data, sizeof(*data));
 308                 if (err)
 309                         err = -EFAULT;
 310                 GOTO(out, err);
 311         }
 312
 313         case OBD_IOC_ATTACH: {
 314                 struct obd_type *type;
 315                 int minor;
 316
 317                 /* have we attached a type to this device */
 318                 if (obd->obd_flags & OBD_ATTACHED || obd->obd_type) {
 319                         CERROR("OBD: Device %d already typed as %s.\n",
 320                                obd->obd_minor, MKSTR(obd->obd_type->typ_name));
 321                         GOTO(out, err=-EBUSY);
 322                 }
 323
 324                 if (!data->ioc_inllen1 || !data->ioc_inlbuf1) {
 325                         CERROR("No type passed!\n");
 326                         GOTO(out, err=-EINVAL);
 327                 }
 328                 if (data->ioc_inlbuf1[data->ioc_inllen1-1] !=0) {
 329                         CERROR("Type not nul terminated!\n");
 330                         GOTO(out, err=-EINVAL);
 331                 }
 332
 333                 CDEBUG(D_IOCTL, "attach type %s name: %s uuid: %s\n",
 334                        MKSTR(data->ioc_inlbuf1),
 335                        MKSTR(data->ioc_inlbuf2), MKSTR(data->ioc_inlbuf3));
 336
 337                 /* find the type */
 338                 type = class_nm_to_type(data->ioc_inlbuf1);
 339                 if (!type) {
 340                         CERROR("OBD: unknown type dev %d\n", obd->obd_minor);
 341                         GOTO(out, err=-EINVAL);
 342                 }
 343
 344                 minor = obd->obd_minor;
 345                 memset(obd, 0, sizeof(*obd));
 346                 obd->obd_minor = minor;
 347                 obd->obd_type = type;
 348                 INIT_LIST_HEAD(&obd->obd_exports);
 349                 INIT_LIST_HEAD(&obd->obd_imports);
 350                 spin_lock_init(&obd->obd_dev_lock);
 351
 352                         if (data->ioc_inlbuf2) {
 353                                 int len = strlen(data->ioc_inlbuf2) + 1;
 354                                 OBD_ALLOC(obd->obd_name, len);
 355                                 if (!obd->obd_name) {
 356                                         CERROR("no memory\n");
 357                                         LBUG();
 358                                 }
 359                                 memcpy(obd->obd_name, data->ioc_inlbuf2, len);
 360                         } else {
 361                                 CERROR("WARNING: unnamed obd device\n");
 362                         }
 363                         if (data->ioc_inlbuf3) {
 364                                 int len = strlen(data->ioc_inlbuf3);
 365                                 if (len >= sizeof(obd->obd_uuid)) {
 366                                         CERROR("uuid must be < %d bytes long\n",
 367                                                sizeof(obd->obd_uuid));
 368                                         if (obd->obd_name)
 369                                                 OBD_FREE(obd->obd_name,
 370                                                          strlen(obd->obd_name) + 1);
 371                                         GOTO(out, err=-EINVAL);
 372                                 }
 373                                 memcpy(obd->obd_uuid, data->ioc_inlbuf3, len);
 374                         }
 375                 /* do the attach */
 376                 if (OBP(obd, attach))
 377                         err = OBP(obd,attach)(obd, sizeof(*data), data);
 378                 if (err) {
 379                         if(data->ioc_inlbuf2)
 380                                 OBD_FREE(obd->obd_name, strlen(obd->obd_name)+1);
 381                         obd->obd_type = NULL;
 382
 383                         } else {
 384                         obd->obd_flags |= OBD_ATTACHED;
 385
 386                         type->typ_refcnt++;
 387                         CDEBUG(D_IOCTL, "OBD: dev %d attached type %s\n",
 388                                obd->obd_minor, data->ioc_inlbuf1);
 389
 390                         CDEBUG(D_IOCTL, "MOD_INC_USE for attach: count = %d\n",
 391                                atomic_read(&(THIS_MODULE)->uc.usecount));
 392                         MOD_INC_USE_COUNT;
 393                 }
 394
 395                 GOTO(out, err);
 396         }
 397
 398         case OBD_IOC_DETACH: {
 399                 ENTRY;
 400                 if (obd->obd_flags & OBD_SET_UP) {
 401                         CERROR("OBD device %d still set up\n", obd->obd_minor);
 402                         GOTO(out, err=-EBUSY);
 403                 }
 404                 if (!(obd->obd_flags & OBD_ATTACHED) ) {
 405                         CERROR("OBD device %d not attached\n", obd->obd_minor);
 406                         GOTO(out, err=-ENODEV);
 407                 }
 408                 if (!list_empty(&obd->obd_exports)) {
 409                         if (!data->ioc_inlbuf1 || data->ioc_inlbuf1[0] != 'F') {
 410                                 CERROR("OBD device %d (%p) has exports\n",
 411                                        obd->obd_minor, obd);
 412                                 GOTO(out, err=-EBUSY);
 413                         }
 414                         forcibly_detach_exports(obd);
 415                 }
 416                    if (OBP(obd, detach))
 417                         err=OBP(obd,detach)(obd);
 418
 419                 if (obd->obd_name) {
 420                         OBD_FREE(obd->obd_name, strlen(obd->obd_name)+1);
 421                         obd->obd_name = NULL;
 422                 }
 423
 424                 obd->obd_flags &= ~OBD_ATTACHED;
 425                 obd->obd_type->typ_refcnt--;
 426                 obd->obd_type = NULL;
 427                 CDEBUG(D_IOCTL, "MOD_DEC_USE for detach: count = %d\n",
 428                        atomic_read(&(THIS_MODULE)->uc.usecount) - 1);
 429                 MOD_DEC_USE_COUNT;
 430                 GOTO(out, err = 0);
 431         }
 432
 433         case OBD_IOC_SETUP: {
 434                 /* have we attached a type to this device? */
 435                 if (!(obd->obd_flags & OBD_ATTACHED)) {
 436                         CERROR("Device %d not attached\n", obd->obd_minor);
 437                         GOTO(out, err=-ENODEV);
 438                 }
 439
 440                 /* has this been done already? */
 441                 if ( obd->obd_flags & OBD_SET_UP ) {
 442                         CERROR("Device %d already setup (type %s)\n",
 443                                obd->obd_minor, obd->obd_type->typ_name);
 444                         GOTO(out, err=-EBUSY);
 445                 }
 446
 447                 if ( OBT(obd) && OBP(obd, setup) )
 448                         err = obd_setup(obd, sizeof(*data), data);
 449
 450                 if (!err) {
 451                         obd->obd_type->typ_refcnt++;
 452                         obd->obd_flags |= OBD_SET_UP;
 453                 }
 454
 455                 GOTO(out, err);
 456         }
 457         case OBD_IOC_CLEANUP: {
 458                 /* have we attached a type to this device? */
 459                 if (!(obd->obd_flags & OBD_ATTACHED)) {
 460                         CERROR("Device %d not attached\n", obd->obd_minor);
 461                         GOTO(out, err=-ENODEV);
 462                 }
 463
 464                 if ( OBT(obd) && OBP(obd, cleanup) )
 465                         err = obd_cleanup(obd);
 466
 467                 if (!err) {
 468                         obd->obd_flags &= ~OBD_SET_UP;
 469                         obd->obd_type->typ_refcnt--;
 470                 }
 471                 GOTO(out, err);
 472         }
 473
 474         case OBD_IOC_CONNECT: {
 475                 char * cluuid = "OBD_CLASS_UUID";
 476                 obd_data2conn(&conn, data);
 477
 478                 err = obd_connect(&conn, obd, cluuid, NULL, NULL);
 479
 480                 CDEBUG(D_IOCTL, "assigned export "LPX64"\n", conn.addr);
 481                 obd_conn2data(data, &conn);
 482                 if (err)
 483                         GOTO(out, err);
 484
 485                 err = copy_to_user((void *)arg, data, sizeof(*data));
 486                 if (err)
 487                         err = -EFAULT;
 488                 // XXX save connection data into file handle
 489                 GOTO(out, err);
 490         }
 491
 492         case OBD_IOC_DISCONNECT: {
 493                 obd_data2conn(&conn, data);
 494                 err = obd_disconnect(&conn);
 495                 GOTO(out, err);
 496         }
 497
 498         case OBD_IOC_DEC_USE_COUNT: {
 499                 CDEBUG(D_IOCTL, "MOD_DEC_USE for force dec: count = %d\n",
 500                        atomic_read(&(THIS_MODULE)->uc.usecount) - 1);
 501                 MOD_DEC_USE_COUNT;
 502                 GOTO(out, err=0);
 503         }
 504
 505         case OBD_IOC_CREATE: {
 506                 struct lov_stripe_md *lsm = NULL;
 507                 obd_data2conn(&conn, data);
 508
 509 #warning FIXME: save lsm into file handle for other ops, release on close
 510                 err = obd_create(&conn, &data->ioc_obdo1, &lsm);
 511                 if (!err) {
 512                         err = copy_to_user((void *)arg, data, sizeof(*data));
 513                         if (err)
 514                                 err = -EFAULT;
 515                 }
 516                 GOTO(out, err);
 517         }
 518
 519         case OBD_IOC_GETATTR: {
 520                 obd_data2conn(&conn, data);
 521                 err = obd_getattr(&conn, &data->ioc_obdo1, NULL);
 522                 if (!err) {
 523                         err = copy_to_user((void *)arg, data, sizeof(*data));
 524                         if (err)
 525                                 err = -EFAULT;
 526                 }
 527                 GOTO(out, err);
 528         }
 529
 530         case OBD_IOC_SETATTR: {
 531                 obd_data2conn(&conn, data);
 532                 err = obd_setattr(&conn, &data->ioc_obdo1, NULL);
 533                 if (!err) {
 534                         err = copy_to_user((void *)arg, data, sizeof(*data));
 535                         if (err)
 536                                 err = -EFAULT;
 537                 }
 538                 GOTO(out, err);
 539         }
 540
 541         case OBD_IOC_DESTROY: {
 542                 //void *ea;
 543                 obd_data2conn(&conn, data);
 544
 545                 err = obd_destroy(&conn, &data->ioc_obdo1, NULL);
 546                 if (!err) {
 547                         err = copy_to_user((void *)arg, data, sizeof(*data));
 548                         if (err)
 549                                 err = -EFAULT;
 550                 }
 551                 GOTO(out, err);
 552         }
 553
 554         case OBD_IOC_OPEN: {
 555                 struct lov_stripe_md *lsm = NULL; // XXX fill in from create
 556
 557                 obd_data2conn(&conn, data);
 558                 err = obd_open(&conn, &data->ioc_obdo1, lsm);
 559                 if (!err) {
 560                         err = copy_to_user((void *)arg, data, sizeof(*data));
 561                         if (err)
 562                                 err = -EFAULT;
 563                 }
 564                 GOTO(out, err);
 565         }
 566
 567         case OBD_IOC_CLOSE: {
 568                 struct lov_stripe_md *lsm = NULL; // XXX fill in from create
 569
 570                 obd_data2conn(&conn, data);
 571                 err = obd_close(&conn, &data->ioc_obdo1, lsm);
 572                 GOTO(out, err);
 573         }
 574
 575         case OBD_IOC_BRW_WRITE:
 576                 rw = OBD_BRW_WRITE;
 577         case OBD_IOC_BRW_READ: {
 578                 struct lov_stripe_md tmp_lsm; // XXX fill in from create
 579                 struct lov_stripe_md *lsm = &tmp_lsm; // XXX fill in from create
 580                 struct brw_cb_data *brw_cbd = ll_init_brw_cb_data();
 581                 obd_count       pages = 0;
 582                 struct brw_page *pga, *pgp;
 583                 __u64 id = data->ioc_obdo1.o_id;
 584                 int gfp_mask = (id & 1) ? GFP_HIGHUSER : GFP_KERNEL;
 585                 int verify = (id != 0);
 586                 __u64 off;
 587                 int j;
 588
 589                 if (!brw_cbd)
 590                         GOTO(out, err = -ENOMEM);
 591
 592                 obd_data2conn(&conn, data);
 593
 594                 pages = data->ioc_count / PAGE_SIZE;
 595                 off = data->ioc_offset;
 596
 597                 CDEBUG(D_INODE, "BRW %s with %d pages @ "LPX64"\n",
 598                        rw == OBD_BRW_READ ? "read" : "write", pages, off);
 599                 OBD_ALLOC(pga, pages * sizeof(*pga));
 600                 if (!pga) {
 601                         CERROR("no memory for %d BRW per-page data\n", pages);
 602                         GOTO(brw_free, err = -ENOMEM);
 603                 }
 604
 605                 memset(lsm, 0, sizeof(*lsm)); // XXX don't do this later
 606                 lsm->lsm_object_id = id; // ensure id == lsm->lsm_object_id
 607
 608                 for (j = 0, pgp = pga; j < pages; j++, off += PAGE_SIZE, pgp++){
 609                         pgp->pg = alloc_pages(gfp_mask, 0);
 610                         if (!pgp->pg) {
 611                                 CERROR("no memory for brw pages\n");
 612                                 GOTO(brw_cleanup, err = -ENOMEM);
 613                         }
 614                         pgp->count = PAGE_SIZE;
 615                         pgp->off = off;
 616                         pgp->flag = 0;
 617
 618                         if (verify) {
 619                                 void *addr = kmap(pgp->pg);
 620
 621                                 if (rw == OBD_BRW_WRITE)
 622                                         page_debug_setup(addr, pgp->count,
 623                                                          pgp->off, id);
 624                                 else
 625                                         page_debug_setup(addr, pgp->count,
 626                                                          0xdeadbeef00c0ffee,
 627                                                          0xdeadbeef00c0ffee);
 628                                 kunmap(pgp->pg);
 629                         }
 630                 }
 631
 632                 err = obd_brw(rw, &conn, lsm, j, pga, ll_sync_brw_cb, brw_cbd);
 633                 if (err)
 634                         CERROR("test_brw: error from obd_brw: err = %d\n", err);
 635                 EXIT;
 636         brw_cleanup:
 637                 for (j = 0, pgp = pga; j < pages; j++, pgp++) {
 638                         if (pgp->pg != NULL) {
 639                                 if (verify && !err) {
 640                                         void *addr = kmap(pgp->pg);
 641
 642                                         err = page_debug_check("test_brw",
 643                                                                addr,
 644                                                                PAGE_SIZE,
 645                                                                pgp->off,id);
 646                                         kunmap(pgp->pg);
 647                                 }
 648                                 __free_pages(pgp->pg, 0);
 649                         }
 650                 }
 651         brw_free:
 652                 OBD_FREE(pga, pages * sizeof(*pga));
 653                 GOTO(out, err);
 654         }
 655         default:
 656                 obd_data2conn(&conn, data);
 657
 658                 err = obd_iocontrol(cmd, &conn, len, data, NULL);
 659                 if (err)
 660                         GOTO(out, err);
 661
 662                 err = copy_to_user((void *)arg, data, len);
 663                 if (err)
 664                         err = -EFAULT;
 665                 GOTO(out, err);
 666         }
 667
 668  out:
 669         if (buf)
 670                 OBD_FREE(buf, len);
 671         if (serialised)
 672                 up(&obd_conf_sem);
 673         RETURN(err);
 674 } /* obd_class_ioctl */
 675
 676
 677
 678 /* declare character device */
 679 static struct file_operations obd_psdev_fops = {
 680         ioctl: obd_class_ioctl,       /* ioctl */
 681         open: obd_class_open,        /* open */
 682         release: obd_class_release,     /* release */
 683 };
 684
 685 /* modules setup */
 686 #define OBD_MINOR 241
 687 static struct miscdevice obd_psdev = {
 688         OBD_MINOR,
 689         "obd_psdev",
 690         &obd_psdev_fops
 691 };
 692
 693 void (*class_signal_connection_failure)(struct ptlrpc_connection *);
 694
 695 #ifdef CONFIG_HIGHMEM
 696 #warning "using kmap accounting for deadlock avoidance"
 697 /* Allow at most 3/4 of the kmap mappings to be consumed by vector I/O
 698  * requests.  This avoids deadlocks on servers which have a lot of clients
 699  * doing vector I/O.  We don't need to do this for non-vector I/O requests
 700  * because singleton requests will just block on the kmap itself and never
 701  * deadlock waiting for additional kmaps to complete.
 702  *
 703  * If we are a "server" task, we can have at most a single reservation
 704  * in excess of the maximum.  This avoids a deadlock when multiple client
 705  * threads are on the same machine as the server threads, and the clients
 706  * have consumed all of the available mappings.  As long as a single server
 707  * thread is can make progress, we are guaranteed to avoid deadlock.
 708  */
 709 #define OBD_KMAP_MAX (LAST_PKMAP * 3 / 4)
 710 static atomic_t obd_kmap_count = ATOMIC_INIT(OBD_KMAP_MAX);
 711 static DECLARE_WAIT_QUEUE_HEAD(obd_kmap_waitq);
 712
 713 void obd_kmap_get(int count, int server)
 714 {
 715         //CERROR("getting %d kmap counts (%d/%d)\n", count,
 716         //       atomic_read(&obd_kmap_count), OBD_KMAP_MAX);
 717         if (count == 1)
 718                 atomic_dec(&obd_kmap_count);
 719         else while (atomic_add_negative(-count, &obd_kmap_count)) {
 720                 static long next_show = 0;
 721                 static int skipped = 0;
 722
 723                 if (server && atomic_read(&obd_kmap_count) >= -PTL_MD_MAX_IOV)
 724                         break;
 725
 726                 CDEBUG(D_OTHER, "negative kmap reserved count: %d\n",
 727                        atomic_read(&obd_kmap_count));
 728                 atomic_add(count, &obd_kmap_count);
 729
 730                 if (time_after(jiffies, next_show)) {
 731                         CERROR("blocking %s (and %d others) for kmaps\n",
 732                                current->comm, skipped);
 733                         next_show = jiffies + 5*HZ;
 734                         skipped = 0;
 735                 } else
 736                         skipped++;
 737                 wait_event(obd_kmap_waitq,
 738                            atomic_read(&obd_kmap_count) >= count);
 739         }
 740 }
 741
 742 void obd_kmap_put(int count)
 743 {
 744         atomic_add(count, &obd_kmap_count);
 745         /* Wake up sleepers.  Sadly, this wakes up all of the tasks at once.
 746          * We could have something smarter here like:
 747         while (atomic_read(&obd_kmap_count) > 0)
 748                 wake_up_nr(obd_kmap_waitq, 1);
 749         although we would need to set somewhere (probably obd_class_init):
 750         obd_kmap_waitq.flags |= WQ_FLAG_EXCLUSIVE;
 751         For now the wait_event() condition will handle this OK I believe.
 752          */
 753         if (atomic_read(&obd_kmap_count) > 0)
 754                 wake_up(&obd_kmap_waitq);
 755 }
 756
 757 EXPORT_SYMBOL(obd_kmap_get);
 758 EXPORT_SYMBOL(obd_kmap_put);
 759 #endif
 760
 761 EXPORT_SYMBOL(obd_dev);
 762 EXPORT_SYMBOL(obdo_cachep);
 763 EXPORT_SYMBOL(obd_memory);
 764 EXPORT_SYMBOL(obd_fail_loc);
 765 EXPORT_SYMBOL(obd_timeout);
 766 EXPORT_SYMBOL(obd_recovery_upcall);
 767 EXPORT_SYMBOL(ptlrpc_put_connection_superhack);
 768
 769 EXPORT_SYMBOL(class_register_type);
 770 EXPORT_SYMBOL(class_unregister_type);
 771 EXPORT_SYMBOL(class_name2dev);
 772 EXPORT_SYMBOL(class_uuid2dev);
 773 EXPORT_SYMBOL(class_uuid2obd);
 774 EXPORT_SYMBOL(class_new_export);
 775 EXPORT_SYMBOL(class_destroy_export);
 776 EXPORT_SYMBOL(class_connect);
 777 EXPORT_SYMBOL(class_conn2export);
 778 EXPORT_SYMBOL(class_conn2obd);
 779 EXPORT_SYMBOL(class_conn2cliimp);
 780 EXPORT_SYMBOL(class_conn2ldlmimp);
 781 EXPORT_SYMBOL(class_disconnect);
 782 EXPORT_SYMBOL(class_disconnect_all);
 783 //EXPORT_SYMBOL(class_uuid_parse);
 784 EXPORT_SYMBOL(class_uuid_unparse);
 785 //EXPORT_SYMBOL(class_multi_setup);
 786 //EXPORT_SYMBOL(class_multi_cleanup);
 787
 788 EXPORT_SYMBOL(class_signal_connection_failure);
 789 EXPORT_SYMBOL(ll_sync_brw_cb);
 790 EXPORT_SYMBOL(ll_init_brw_cb_data);
 791 EXPORT_SYMBOL(class_nm_to_type);
 792
 793 static int __init init_obdclass(void)
 794 {
 795         struct obd_device *obd;
 796         int err;
 797         int i;
 798
 799         printk(KERN_INFO "OBD class driver  v0.9, info@clusterfs.com\n");
 800
 801         sema_init(&obd_conf_sem, 1);
 802         INIT_LIST_HEAD(&obd_types);
 803
 804         if ((err = misc_register(&obd_psdev))) {
 805                 CERROR("cannot register %d err %d\n", OBD_MINOR, err);
 806                 return err;
 807         }
 808
 809         /* This struct is already zerod for us (static global) */
 810         for (i = 0, obd = obd_dev; i < MAX_OBD_DEVICES; i++, obd++)
 811                 obd->obd_minor = i;
 812
 813         err = obd_init_caches();
 814
 815         if (err)
 816                 return err;
 817         obd_sysctl_init();
 818
 819         err=lprocfs_reg_main();
 820
 821         return 0;
 822 }
 823
 824 static void __exit cleanup_obdclass(void)
 825 {
 826         int i, err;
 827         ENTRY;
 828
 829         misc_deregister(&obd_psdev);
 830         for (i = 0; i < MAX_OBD_DEVICES; i++) {
 831                 struct obd_device *obd = &obd_dev[i];
 832                 if (obd->obd_type && (obd->obd_flags & OBD_SET_UP) &&
 833                     OBT(obd) && OBP(obd, detach)) {
 834                         /* XXX should this call generic detach otherwise? */
 835                         OBP(obd, detach)(obd);
 836                 }
 837         }
 838
 839         obd_cleanup_caches();
 840         obd_sysctl_clean();
 841
 842         err = lprocfs_dereg_main();
 843
 844         CERROR("obd memory leaked: %ld bytes\n", obd_memory);
 845         EXIT;
 846 }
 847
 848 /* Check that we're building against the appropriate version of the Lustre
 849  * kernel patch */
 850 #include <linux/lustre_version.h>
 851 #if (LUSTRE_KERNEL_VERSION != 2)
 852 # error Cannot continue: Your Lustre kernel patch is out of date
 853 #endif
 854
 855 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 856 MODULE_DESCRIPTION("Lustre Class Driver v1.0");
 857 MODULE_LICENSE("GPL");
 858
 859 module_init(init_obdclass);
 860 module_exit(cleanup_obdclass);