lustre/obdclass/class_obd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 1999, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_CLASS
  33
  34 #include <linux/miscdevice.h>
  35 #include <linux/user_namespace.h>
  36 #include <linux/uidgid.h>
  37 #include <linux/atomic.h>
  38 #include <linux/list.h>
  39 #include <linux/oom.h>
  40
  41 #include <obd_support.h>
  42 #include <obd_class.h>
  43 #include <uapi/linux/lnet/lnetctl.h>
  44 #include <lustre_kernelcomm.h>
  45 #include <lprocfs_status.h>
  46 #include <cl_object.h>
  47 #ifdef HAVE_SERVER_SUPPORT
  48 # include <dt_object.h>
  49 # include <md_object.h>
  50 #endif /* HAVE_SERVER_SUPPORT */
  51 #include <uapi/linux/lustre/lustre_ioctl.h>
  52 #include "llog_internal.h"
  53 #include <lustre_ioctl_old.h>
  54
  55 static __u64 obd_max_alloc;
  56
  57 static DEFINE_SPINLOCK(obd_updatemax_lock);
  58
  59 /* The following are visible and mutable through /proc/sys/lustre/. */
  60 unsigned int obd_debug_peer_on_timeout;
  61 EXPORT_SYMBOL(obd_debug_peer_on_timeout);
  62 unsigned int obd_dump_on_timeout;
  63 EXPORT_SYMBOL(obd_dump_on_timeout);
  64 unsigned int obd_dump_on_eviction;
  65 EXPORT_SYMBOL(obd_dump_on_eviction);
  66 unsigned int obd_lbug_on_eviction;
  67 EXPORT_SYMBOL(obd_lbug_on_eviction);
  68 unsigned long obd_max_dirty_pages;
  69 EXPORT_SYMBOL(obd_max_dirty_pages);
  70 atomic_long_t obd_dirty_pages;
  71 EXPORT_SYMBOL(obd_dirty_pages);
  72 unsigned int obd_timeout = OBD_TIMEOUT_DEFAULT;   /* seconds */
  73 EXPORT_SYMBOL(obd_timeout);
  74 unsigned int ldlm_timeout = LDLM_TIMEOUT_DEFAULT; /* seconds */
  75 EXPORT_SYMBOL(ldlm_timeout);
  76 unsigned int ping_interval = (OBD_TIMEOUT_DEFAULT > 4) ?
  77                              (OBD_TIMEOUT_DEFAULT / 4) : 1;
  78 EXPORT_SYMBOL(ping_interval);
  79 unsigned int ping_evict_timeout_multiplier = 6;
  80 EXPORT_SYMBOL(ping_evict_timeout_multiplier);
  81 unsigned int obd_timeout_set;
  82 EXPORT_SYMBOL(obd_timeout_set);
  83 unsigned int ldlm_timeout_set;
  84 EXPORT_SYMBOL(ldlm_timeout_set);
  85 /* bulk transfer timeout, give up after 100s by default */
  86 unsigned int bulk_timeout = 100; /* seconds */
  87 EXPORT_SYMBOL(bulk_timeout);
  88
  89 /* Adaptive timeout defs here instead of ptlrpc module for /proc/sys/ access */
  90 unsigned int at_min = 5;
  91 EXPORT_SYMBOL(at_min);
  92 unsigned int at_max = 600;
  93 EXPORT_SYMBOL(at_max);
  94 unsigned int at_history = 600;
  95 EXPORT_SYMBOL(at_history);
  96 int at_early_margin = 5;
  97 EXPORT_SYMBOL(at_early_margin);
  98 int at_extra = 30;
  99 EXPORT_SYMBOL(at_extra);
 100
 101 struct percpu_counter obd_memory;
 102 EXPORT_SYMBOL(obd_memory);
 103
 104 static int obdclass_oom_handler(struct notifier_block *self,
 105                                 unsigned long notused, void *nfreed)
 106 {
 107         /* in bytes */
 108         pr_info("obd_memory max: %llu, obd_memory current: %llu\n",
 109                 obd_memory_max(), obd_memory_sum());
 110
 111         return NOTIFY_OK;
 112 }
 113
 114 static struct notifier_block obdclass_oom = {
 115         .notifier_call = obdclass_oom_handler
 116 };
 117
 118 int obd_ioctl_msg(const char *file, const char *func, int line, int level,
 119                   const char *name, unsigned int cmd, const char *msg, int rc)
 120 {
 121         static struct cfs_debug_limit_state cdls;
 122         char *dirs[] = {
 123                 [_IOC_NONE]             = "_IO",
 124                 [_IOC_READ]             = "_IOR",
 125                 [_IOC_WRITE]            = "_IOW",
 126                 [_IOC_READ|_IOC_WRITE]  = "_IOWR",
 127         };
 128         char type;
 129
 130         type = _IOC_TYPE(cmd);
 131         __CDEBUG_WITH_LOC(file, func, line, level, &cdls,
 132                           "%s: iocontrol from '%s' cmd=%x %s('%c', %u, %u) %s: rc = %d\n",
 133                           name, current->comm, cmd,
 134                           dirs[_IOC_DIR(cmd)] ?: "_IO?",
 135                           isprint(type) ? type : '?', _IOC_NR(cmd),
 136                           _IOC_SIZE(cmd), msg, rc);
 137         return rc;
 138 }
 139 EXPORT_SYMBOL(obd_ioctl_msg);
 140
 141 static int class_resolve_dev_name(__u32 len, const char *name)
 142 {
 143         int rc;
 144         int dev;
 145
 146         ENTRY;
 147         if (!len || !name) {
 148                 CERROR("No name passed,!\n");
 149                 GOTO(out, rc = -EINVAL);
 150         }
 151         if (name[len - 1] != 0) {
 152                 CERROR("Name not nul terminated!\n");
 153                 GOTO(out, rc = -EINVAL);
 154         }
 155
 156         CDEBUG(D_IOCTL, "device name %s\n", name);
 157         dev = class_name2dev(name);
 158         if (dev == -1) {
 159                 CDEBUG(D_IOCTL, "No device for name %s!\n", name);
 160                 GOTO(out, rc = -EINVAL);
 161         }
 162
 163         CDEBUG(D_IOCTL, "device name %s, dev %d\n", name, dev);
 164         rc = dev;
 165
 166 out:
 167         RETURN(rc);
 168 }
 169
 170 #define OBD_MAX_IOCTL_BUFFER    8192
 171
 172 static int obd_ioctl_is_invalid(struct obd_ioctl_data *data)
 173 {
 174         const int maxlen = 1 << 30;
 175         if (data->ioc_len > maxlen) {
 176                 CERROR("OBD ioctl: ioc_len larger than 1<<30\n");
 177                 return 1;
 178         }
 179
 180         if (data->ioc_inllen1 > maxlen) {
 181                 CERROR("OBD ioctl: ioc_inllen1 larger than 1<<30\n");
 182                 return 1;
 183         }
 184
 185         if (data->ioc_inllen2 > maxlen) {
 186                 CERROR("OBD ioctl: ioc_inllen2 larger than 1<<30\n");
 187                 return 1;
 188         }
 189
 190         if (data->ioc_inllen3 > maxlen) {
 191                 CERROR("OBD ioctl: ioc_inllen3 larger than 1<<30\n");
 192                 return 1;
 193         }
 194
 195         if (data->ioc_inllen4 > maxlen) {
 196                 CERROR("OBD ioctl: ioc_inllen4 larger than 1<<30\n");
 197                 return 1;
 198         }
 199
 200         if (data->ioc_inlbuf1 && data->ioc_inllen1 == 0) {
 201                 CERROR("OBD ioctl: inlbuf1 pointer but 0 length\n");
 202                 return 1;
 203         }
 204
 205         if (data->ioc_inlbuf2 && data->ioc_inllen2 == 0) {
 206                 CERROR("OBD ioctl: inlbuf2 pointer but 0 length\n");
 207                 return 1;
 208         }
 209
 210         if (data->ioc_inlbuf3 && data->ioc_inllen3 == 0) {
 211                 CERROR("OBD ioctl: inlbuf3 pointer but 0 length\n");
 212                 return 1;
 213         }
 214
 215         if (data->ioc_inlbuf4 && data->ioc_inllen4 == 0) {
 216                 CERROR("OBD ioctl: inlbuf4 pointer but 0 length\n");
 217                 return 1;
 218         }
 219
 220         if (data->ioc_pbuf1 && data->ioc_plen1 == 0) {
 221                 CERROR("OBD ioctl: pbuf1 pointer but 0 length\n");
 222                 return 1;
 223         }
 224
 225         if (data->ioc_pbuf2 && data->ioc_plen2 == 0) {
 226                 CERROR("OBD ioctl: pbuf2 pointer but 0 length\n");
 227                 return 1;
 228         }
 229
 230         if (!data->ioc_pbuf1 && data->ioc_plen1 != 0) {
 231                 CERROR("OBD ioctl: plen1 set but NULL pointer\n");
 232                 return 1;
 233         }
 234
 235         if (!data->ioc_pbuf2 && data->ioc_plen2 != 0) {
 236                 CERROR("OBD ioctl: plen2 set but NULL pointer\n");
 237                 return 1;
 238         }
 239
 240         if (obd_ioctl_packlen(data) > data->ioc_len) {
 241                 CERROR("OBD ioctl: packlen exceeds ioc_len (%d > %d)\n",
 242                        obd_ioctl_packlen(data), data->ioc_len);
 243                 return 1;
 244         }
 245
 246         return 0;
 247 }
 248
 249 /* buffer MUST be at least the size of obd_ioctl_hdr */
 250 int obd_ioctl_getdata(struct obd_ioctl_data **datap, int *len, void __user *arg)
 251 {
 252         struct obd_ioctl_hdr hdr;
 253         struct obd_ioctl_data *data;
 254         int offset = 0;
 255         int rc = -EINVAL;
 256
 257         ENTRY;
 258         if (copy_from_user(&hdr, arg, sizeof(hdr)))
 259                 RETURN(-EFAULT);
 260
 261         if (hdr.ioc_version != OBD_IOCTL_VERSION) {
 262                 CERROR("%s: kernel/user version mismatch (%x != %x): rc = %d\n",
 263                        current->comm, OBD_IOCTL_VERSION, hdr.ioc_version, rc);
 264                 RETURN(rc);
 265         }
 266
 267         if (hdr.ioc_len > OBD_MAX_IOCTL_BUFFER) {
 268                 CERROR("%s: user buffer len %d exceeds %d max: rc = %d\n",
 269                        current->comm, hdr.ioc_len, OBD_MAX_IOCTL_BUFFER, rc);
 270                 RETURN(rc);
 271         }
 272
 273         if (hdr.ioc_len < sizeof(*data)) {
 274                 CERROR("%s: user buffer %d too small for ioctl %zu: rc = %d\n",
 275                        current->comm, hdr.ioc_len, sizeof(*data), rc);
 276                 RETURN(rc);
 277         }
 278
 279         /* When there are lots of processes calling vmalloc on multi-core
 280          * system, the high lock contention will hurt performance badly,
 281          * obdfilter-survey is an example, which relies on ioctl. So we'd
 282          * better avoid vmalloc on ioctl path. LU-66
 283          */
 284         OBD_ALLOC_LARGE(data, hdr.ioc_len);
 285         if (!data) {
 286                 rc = -ENOMEM;
 287                 CERROR("%s: cannot allocate control buffer len %d: rc = %d\n",
 288                        current->comm, hdr.ioc_len, rc);
 289                 RETURN(rc);
 290         }
 291         *len = hdr.ioc_len;
 292
 293         if (copy_from_user(data, arg, hdr.ioc_len))
 294                 GOTO(out_free, rc = -EFAULT);
 295
 296         if (obd_ioctl_is_invalid(data))
 297                 GOTO(out_free, rc = -EINVAL);
 298
 299         if (data->ioc_inllen1) {
 300                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
 301                 offset += round_up(data->ioc_inllen1, 8);
 302         }
 303
 304         if (data->ioc_inllen2) {
 305                 data->ioc_inlbuf2 = &data->ioc_bulk[0] + offset;
 306                 offset += round_up(data->ioc_inllen2, 8);
 307         }
 308
 309         if (data->ioc_inllen3) {
 310                 data->ioc_inlbuf3 = &data->ioc_bulk[0] + offset;
 311                 offset += round_up(data->ioc_inllen3, 8);
 312         }
 313
 314         if (data->ioc_inllen4)
 315                 data->ioc_inlbuf4 = &data->ioc_bulk[0] + offset;
 316
 317         *datap = data;
 318
 319         RETURN(0);
 320
 321 out_free:
 322         OBD_FREE_LARGE(data, *len);
 323         RETURN(rc);
 324 }
 325 EXPORT_SYMBOL(obd_ioctl_getdata);
 326
 327 int class_handle_ioctl(unsigned int cmd, void __user *uarg)
 328 {
 329         struct obd_ioctl_data *data;
 330         struct obd_device *obd = NULL;
 331         int rc, len = 0;
 332
 333         ENTRY;
 334         CDEBUG(D_IOCTL, "obdclass: cmd=%x len=%u uarg=%pK\n", cmd, len, uarg);
 335         if (unlikely(_IOC_TYPE(cmd) != 'f' && !OBD_IOC_BARRIER_ALLOW(cmd) &&
 336                      !IOC_OSC_SET_ACTIVE_ALLOW(cmd)))
 337                 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "unknown", -ENOTTY));
 338
 339         rc = obd_ioctl_getdata(&data, &len, uarg);
 340         if (rc) {
 341                 CERROR("%s: ioctl data error: rc = %d\n", current->comm, rc);
 342                 RETURN(rc);
 343         }
 344
 345         switch (cmd) {
 346         case OBD_IOC_PROCESS_CFG: {
 347                 struct lustre_cfg *lcfg;
 348
 349                 if (!data->ioc_plen1 || !data->ioc_pbuf1) {
 350                         rc = OBD_IOC_ERROR("obdclass", cmd, "no config buffer",
 351                                            -EINVAL);
 352                         GOTO(out, rc);
 353                 }
 354                 OBD_ALLOC(lcfg, data->ioc_plen1);
 355                 if (lcfg == NULL)
 356                         GOTO(out, rc = -ENOMEM);
 357                 rc = copy_from_user(lcfg, data->ioc_pbuf1, data->ioc_plen1);
 358                 if (!rc)
 359                         rc = lustre_cfg_sanity_check(lcfg, data->ioc_plen1);
 360                 if (!rc)
 361                         rc = class_process_config(lcfg);
 362
 363                 OBD_FREE(lcfg, data->ioc_plen1);
 364                 GOTO(out, rc);
 365         }
 366
 367 #ifdef OBD_GET_VERSION
 368         case_OBD_IOC_DEPRECATED(OBD_GET_VERSION, "obdclass", 2, 15)
 369                 if (!data->ioc_inlbuf1) {
 370                         rc = OBD_IOC_ERROR("obdclass", cmd, "no buffer passed",
 371                                            -EINVAL);
 372                         GOTO(out, rc);
 373                 }
 374
 375                 if (strlen(LUSTRE_VERSION_STRING) + 1 > data->ioc_inllen1) {
 376                         rc = OBD_IOC_ERROR("obdclass", cmd, "buffer too small",
 377                                            -EINVAL);
 378                         GOTO(out, rc);
 379                 }
 380
 381                 memcpy(data->ioc_bulk, LUSTRE_VERSION_STRING,
 382                        strlen(LUSTRE_VERSION_STRING) + 1);
 383
 384                 if (copy_to_user(uarg, data, len))
 385                         rc = -EFAULT;
 386                 GOTO(out, rc);
 387 #endif
 388         case OBD_IOC_NAME2DEV: {
 389                 /* Resolve device name, does not change current selected dev */
 390                 int dev;
 391
 392                 dev = class_resolve_dev_name(data->ioc_inllen1,
 393                                              data->ioc_inlbuf1);
 394                 data->ioc_dev = dev;
 395                 if (dev < 0)
 396                         GOTO(out, rc = -EINVAL);
 397
 398                 if (copy_to_user(uarg, data, sizeof(*data)))
 399                         rc = -EFAULT;
 400                 GOTO(out, rc);
 401         }
 402
 403         case OBD_IOC_UUID2DEV: {
 404                 /* Resolve device uuid, does not change current selected dev */
 405                 struct obd_uuid uuid;
 406                 int dev;
 407
 408                 if (!data->ioc_inllen1 || !data->ioc_inlbuf1) {
 409                         rc = OBD_IOC_ERROR("obdclass", cmd, "no UUID passed",
 410                                            -EINVAL);
 411                         GOTO(out, rc);
 412                 }
 413                 if (data->ioc_inlbuf1[data->ioc_inllen1 - 1] != 0) {
 414                         rc = OBD_IOC_ERROR("obdclass", cmd, "unterminated UUID",
 415                                            -EINVAL);
 416                         GOTO(out, rc);
 417                 }
 418
 419                 CDEBUG(D_IOCTL, "device name %s\n", data->ioc_inlbuf1);
 420                 obd_str2uuid(&uuid, data->ioc_inlbuf1);
 421                 dev = class_uuid2dev(&uuid);
 422                 data->ioc_dev = dev;
 423                 if (dev == -1) {
 424                         CDEBUG(D_IOCTL, "No device for UUID %s!\n",
 425                                data->ioc_inlbuf1);
 426                         GOTO(out, rc = -EINVAL);
 427                 }
 428
 429                 CDEBUG(D_IOCTL, "device name %s, dev %d\n", data->ioc_inlbuf1,
 430                        dev);
 431                 if (copy_to_user(uarg, data, sizeof(*data)))
 432                         rc = -EFAULT;
 433                 GOTO(out, rc);
 434         }
 435
 436         case OBD_IOC_GETDEVICE: {
 437                 int index = data->ioc_count;
 438                 char *status, *str;
 439
 440                 if (!data->ioc_inlbuf1) {
 441                         rc = OBD_IOC_ERROR("obdclass", cmd, "no buffer passed",
 442                                            -EINVAL);
 443                         GOTO(out, rc);
 444                 }
 445                 if (data->ioc_inllen1 < 128) {
 446                         rc = OBD_IOC_ERROR("obdclass", cmd, "too small version",
 447                                            -EINVAL);
 448                         GOTO(out, rc);
 449                 }
 450
 451                 obd = class_num2obd(index);
 452                 if (!obd)
 453                         GOTO(out, rc = -ENOENT);
 454
 455                 if (obd->obd_stopping)
 456                         status = "ST";
 457                 else if (obd->obd_inactive)
 458                         status = "IN";
 459                 else if (obd->obd_set_up)
 460                         status = "UP";
 461                 else if (obd->obd_attached)
 462                         status = "AT";
 463                 else
 464                         status = "--";
 465
 466                 str = data->ioc_bulk;
 467                 snprintf(str, len - sizeof(*data), "%3d %s %s %s %s %d",
 468                          index, status, obd->obd_type->typ_name,
 469                          obd->obd_name, obd->obd_uuid.uuid,
 470                          kref_read(&obd->obd_refcount));
 471
 472                 if (copy_to_user(uarg, data, len))
 473                         rc = -EFAULT;
 474
 475                 GOTO(out, rc);
 476         }
 477         }
 478
 479         if (data->ioc_dev == OBD_DEV_BY_DEVNAME) {
 480                 if (data->ioc_inllen4 <= 0 || data->ioc_inlbuf4 == NULL)
 481                         GOTO(out, rc = -EINVAL);
 482                 if (strnlen(data->ioc_inlbuf4, MAX_OBD_NAME) >= MAX_OBD_NAME)
 483                         GOTO(out, rc = -EINVAL);
 484                 obd = class_name2obd(data->ioc_inlbuf4);
 485         } else {
 486                 obd = class_num2obd(data->ioc_dev);
 487         }
 488
 489         if (obd == NULL) {
 490                 rc = OBD_IOC_ERROR(data->ioc_inlbuf4, cmd, "no device found",
 491                                    -EINVAL);
 492                 GOTO(out, rc);
 493         }
 494         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
 495
 496         if (!obd->obd_set_up || obd->obd_stopping) {
 497                 rc = -EINVAL;
 498                 CERROR("obdclass: device %d not set up: rc = %d\n",
 499                        data->ioc_dev, rc);
 500                 GOTO(out, rc);
 501         }
 502
 503         rc = obd_iocontrol(cmd, obd->obd_self_export, len, data, NULL);
 504         if (rc)
 505                 GOTO(out, rc);
 506
 507         if (copy_to_user(uarg, data, len))
 508                 rc = -EFAULT;
 509 out:
 510         OBD_FREE_LARGE(data, len);
 511         RETURN(rc);
 512 } /* class_handle_ioctl */
 513
 514 /* to control /dev/obd */
 515 static long obd_class_ioctl(struct file *filp, unsigned int cmd,
 516                             unsigned long arg)
 517 {
 518         int err = 0;
 519
 520         ENTRY;
 521         /* Allow non-root access for some limited ioctls */
 522         if (!capable(CAP_SYS_ADMIN))
 523                 RETURN(-EACCES);
 524
 525         if ((cmd & 0xffffff00) == ((int)'T') << 8) /* ignore all tty ioctls */
 526                 RETURN(-ENOTTY);
 527
 528         err = class_handle_ioctl(cmd, (void __user *)arg);
 529
 530         RETURN(err);
 531 }
 532
 533 /* declare character device */
 534 static const struct file_operations obd_psdev_fops = {
 535         .owner          = THIS_MODULE,
 536         .unlocked_ioctl = obd_class_ioctl,      /* unlocked_ioctl */
 537 };
 538
 539 /* modules setup */
 540 static struct miscdevice obd_psdev = {
 541         .minor  = MISC_DYNAMIC_MINOR,
 542         .name   = OBD_DEV_NAME,
 543         .fops   = &obd_psdev_fops,
 544 };
 545
 546 #define test_string_to_size_err(value, expect, def_unit, __rc)                 \
 547 ({                                                                             \
 548         u64 __size;                                                            \
 549         int __ret;                                                             \
 550                                                                                \
 551         BUILD_BUG_ON(sizeof(value) >= 23);                                     \
 552         __ret = sysfs_memparse(value, sizeof(value) - 1, &__size, def_unit);   \
 553         if (__ret != __rc)                                                     \
 554                 CERROR("string_helper: parsing '%s' expect rc %d != got %d\n", \
 555                        value, __rc, __ret);                                    \
 556         else if (!__ret && (u64)expect != __size)                              \
 557                 CERROR("string_helper: parsing '%s' expect %llu != got %llu\n",\
 558                        value, (u64)expect, __size);                            \
 559         __ret;                                                                 \
 560 })
 561 #define test_string_to_size_one(value, expect, def_unit)                       \
 562         test_string_to_size_err(value, expect, def_unit, 0)
 563
 564 static int __init obd_init_checks(void)
 565 {
 566         __u64 u64val, div64val;
 567         char buf[64];
 568         int len, ret = 0;
 569
 570         CDEBUG(D_INFO, "OBD_OBJECT_EOF = %#llx\n", (__u64)OBD_OBJECT_EOF);
 571
 572         u64val = OBD_OBJECT_EOF;
 573         CDEBUG(D_INFO, "u64val OBD_OBJECT_EOF = %#llx\n", u64val);
 574         if (u64val != OBD_OBJECT_EOF) {
 575                 CERROR("__u64 %#llx(%d) != 0xffffffffffffffff\n",
 576                        u64val, (int)sizeof(u64val));
 577                 ret = -EINVAL;
 578         }
 579         len = snprintf(buf, sizeof(buf), "%#llx", u64val);
 580         if (len != 18) {
 581                 CERROR("u64 hex wrong length, strlen(%s)=%d != 18\n", buf, len);
 582                 ret = -EINVAL;
 583         }
 584
 585         div64val = OBD_OBJECT_EOF;
 586         CDEBUG(D_INFO, "u64val OBD_OBJECT_EOF = %#llx\n", u64val);
 587         if (u64val != OBD_OBJECT_EOF) {
 588                 CERROR("__u64 %#llx(%d) != 0xffffffffffffffff\n",
 589                        u64val, (int)sizeof(u64val));
 590                 ret = -EOVERFLOW;
 591         }
 592         if (u64val >> 8 != OBD_OBJECT_EOF >> 8) {
 593                 CERROR("__u64 %#llx(%d) != 0xffffffffffffffff\n",
 594                        u64val, (int)sizeof(u64val));
 595                 ret = -EOVERFLOW;
 596         }
 597         if (do_div(div64val, 256) != (u64val & 255)) {
 598                 CERROR("do_div(%#llx,256) != %llu\n", u64val, u64val & 255);
 599                 ret = -EOVERFLOW;
 600         }
 601         if (u64val >> 8 != div64val) {
 602                 CERROR("do_div(%#llx,256) %llu != %llu\n",
 603                        u64val, div64val, u64val >> 8);
 604                 ret = -EOVERFLOW;
 605         }
 606         len = snprintf(buf, sizeof(buf), "%#llx", u64val);
 607         if (len != 18) {
 608                 CERROR("u64 hex wrong length! strlen(%s)=%d != 18\n", buf, len);
 609                 ret = -EINVAL;
 610         }
 611         len = snprintf(buf, sizeof(buf), "%llu", u64val);
 612         if (len != 20) {
 613                 CERROR("u64 wrong length! strlen(%s)=%d != 20\n", buf, len);
 614                 ret = -EINVAL;
 615         }
 616         len = snprintf(buf, sizeof(buf), "%lld", u64val);
 617         if (len != 2) {
 618                 CERROR("s64 wrong length! strlen(%s)=%d != 2\n", buf, len);
 619                 ret = -EINVAL;
 620         }
 621         if ((u64val & ~PAGE_MASK) >= PAGE_SIZE) {
 622                 CERROR("mask failed: u64val %llu >= %llu\n", u64val,
 623                        (__u64)PAGE_SIZE);
 624                 ret = -EINVAL;
 625         }
 626         if (ret)
 627                 RETURN(ret);
 628
 629         /* invalid string */
 630         if (!test_string_to_size_err("256B34", 256, "B", -EINVAL)) {
 631                 CERROR("string_helpers: format should be number then units\n");
 632                 ret = -EINVAL;
 633         }
 634         if (!test_string_to_size_err("132OpQ", 132, "B", -EINVAL)) {
 635                 CERROR("string_helpers: invalid units should be rejected\n");
 636                 ret = -EINVAL;
 637         }
 638         if (!test_string_to_size_err("1.82B", 1, "B", -EINVAL)) {
 639                 CERROR("string_helpers: 'B' with '.' should be invalid\n");
 640                 ret = -EINVAL;
 641         }
 642         if (test_string_to_size_one("343\n", 343, "B")) {
 643                 CERROR("string_helpers: should ignore newline\n");
 644                 ret = -EINVAL;
 645         }
 646         if (ret)
 647                 RETURN(ret);
 648
 649         /* memparse unit handling */
 650         ret = 0;
 651         ret += test_string_to_size_one("0B", 0, "B");
 652         ret += test_string_to_size_one("512B", 512, "B");
 653         ret += test_string_to_size_one("1.067kB", 1067, "B");
 654         ret += test_string_to_size_one("1.042KiB", 1067, "B");
 655         ret += test_string_to_size_one("8", 8388608, "M");
 656         ret += test_string_to_size_one("65536", 65536, "B");
 657         ret += test_string_to_size_one("128", 131072, "K");
 658         ret += test_string_to_size_one("1M", 1048576, "B");
 659         ret += test_string_to_size_one("0.5T", 549755813888ULL, "T");
 660         ret += test_string_to_size_one("256.5G", 275414777856ULL, "G");
 661         if (ret)
 662                 RETURN(ret);
 663
 664         /* string helper values */
 665         ret += test_string_to_size_one("16", 16777216, "MiB");
 666         ret += test_string_to_size_one("8.39MB", 8390000, "MiB");
 667         ret += test_string_to_size_one("8.00MiB", 8388608, "MiB");
 668         ret += test_string_to_size_one("256GB", 256000000000ULL, "GiB");
 669         ret += test_string_to_size_one("238.731GiB", 256335459385ULL, "GiB");
 670         if (ret)
 671                 RETURN(ret);
 672
 673         /* huge values */
 674         ret += test_string_to_size_one("0.4TB", 400000000000ULL, "TiB");
 675         ret += test_string_to_size_one("12.5TiB", 13743895347200ULL, "TiB");
 676         ret += test_string_to_size_one("2PB", 2000000000000000ULL, "PiB");
 677         ret += test_string_to_size_one("16PiB", 18014398509481984ULL, "PiB");
 678         if (ret)
 679                 RETURN(ret);
 680
 681         /* huge values should overflow */
 682         if (!test_string_to_size_err("1000EiB", 0, "EiB", -EOVERFLOW)) {
 683                 CERROR("string_helpers: failed to detect binary overflow\n");
 684                 ret = -EINVAL;
 685         }
 686         if (!test_string_to_size_err("1000EB", 0, "EiB", -EOVERFLOW)) {
 687                 CERROR("string_helpers: failed to detect decimal overflow\n");
 688                 ret = -EINVAL;
 689         }
 690
 691         return ret;
 692 }
 693
 694 static int __init obdclass_init(void)
 695 {
 696         int err;
 697
 698         LCONSOLE_INFO("Lustre: Build Version: "LUSTRE_VERSION_STRING"\n");
 699
 700         register_oom_notifier(&obdclass_oom);
 701
 702         err = libcfs_setup();
 703         if (err)
 704                 return err;
 705
 706         err = obd_init_checks();
 707         if (err)
 708                 return err;
 709
 710         err = percpu_counter_init(&obd_memory, 0, GFP_KERNEL);
 711         if (err < 0) {
 712                 CERROR("obdclass: initializing 'obd_memory' failed: rc = %d\n",
 713                        err);
 714                 return err;
 715         }
 716
 717         err = libcfs_kkuc_init();
 718         if (err)
 719                 goto cleanup_obd_memory;
 720
 721         err = obd_zombie_impexp_init();
 722         if (err)
 723                 goto cleanup_kkuc;
 724
 725         err = class_handle_init();
 726         if (err)
 727                 goto cleanup_zombie_impexp;
 728
 729         err = misc_register(&obd_psdev);
 730         if (err) {
 731                 CERROR("cannot register OBD miscdevice: rc = %d\n", err);
 732                 goto cleanup_class_handle;
 733         }
 734
 735         /* Default the dirty page cache cap to 1/2 of system memory.
 736          * For clients with less memory, a larger fraction is needed
 737          * for other purposes (mostly for BGL). */
 738         if (cfs_totalram_pages() <= 512 << (20 - PAGE_SHIFT))
 739                 obd_max_dirty_pages = cfs_totalram_pages() / 4;
 740         else
 741                 obd_max_dirty_pages = cfs_totalram_pages() / 2;
 742
 743         err = obd_init_caches();
 744         if (err)
 745                 goto cleanup_deregister;
 746
 747         err = class_procfs_init();
 748         if (err)
 749                 goto cleanup_caches;
 750
 751         err = lu_global_init();
 752         if (err)
 753                 goto cleanup_class_procfs;
 754
 755         err = cl_global_init();
 756         if (err != 0)
 757                 goto cleanup_lu_global;
 758
 759         err = llog_info_init();
 760         if (err)
 761                 goto cleanup_cl_global;
 762
 763         err = obd_pool_init();
 764         if (err)
 765                 goto cleanup_llog_info;
 766
 767 #ifdef HAVE_SERVER_SUPPORT
 768         err = dt_global_init();
 769         if (err != 0)
 770                 goto cleanup_obd_pool;
 771
 772         err = lu_ucred_global_init();
 773         if (err != 0)
 774                 goto cleanup_dt_global;
 775 #endif /* HAVE_SERVER_SUPPORT */
 776
 777         /* simulate a late OOM situation now to require all
 778          * alloc'ed/initialized resources to be freed
 779          */
 780         if (CFS_FAIL_CHECK(OBD_FAIL_OBDCLASS_MODULE_LOAD)) {
 781                 /* force error to ensure module will be unloaded/cleaned */
 782                 err = -ENOMEM;
 783                 goto cleanup_all;
 784         }
 785         return 0;
 786
 787 cleanup_all:
 788 #ifdef HAVE_SERVER_SUPPORT
 789         lu_ucred_global_fini();
 790
 791 cleanup_dt_global:
 792         dt_global_fini();
 793
 794 cleanup_obd_pool:
 795 #endif /* HAVE_SERVER_SUPPORT */
 796         obd_pool_fini();
 797
 798 cleanup_llog_info:
 799         llog_info_fini();
 800
 801 cleanup_cl_global:
 802         cl_global_fini();
 803
 804 cleanup_lu_global:
 805         lu_global_fini();
 806
 807 cleanup_class_procfs:
 808         class_procfs_clean();
 809
 810 cleanup_caches:
 811         obd_cleanup_caches();
 812
 813 cleanup_deregister:
 814         misc_deregister(&obd_psdev);
 815
 816 cleanup_class_handle:
 817         class_handle_cleanup();
 818
 819 cleanup_zombie_impexp:
 820         obd_zombie_impexp_stop();
 821
 822 cleanup_kkuc:
 823         libcfs_kkuc_fini();
 824
 825 cleanup_obd_memory:
 826         percpu_counter_destroy(&obd_memory);
 827
 828         unregister_oom_notifier(&obdclass_oom);
 829         return err;
 830 }
 831
 832 void obd_update_maxusage(void)
 833 {
 834         __u64 max;
 835
 836         max = obd_memory_sum();
 837
 838         spin_lock(&obd_updatemax_lock);
 839         if (max > obd_max_alloc)
 840                 obd_max_alloc = max;
 841         spin_unlock(&obd_updatemax_lock);
 842 }
 843 EXPORT_SYMBOL(obd_update_maxusage);
 844
 845 __u64 obd_memory_max(void)
 846 {
 847         __u64 ret;
 848
 849         obd_update_maxusage();
 850         spin_lock(&obd_updatemax_lock);
 851         ret = obd_max_alloc;
 852         spin_unlock(&obd_updatemax_lock);
 853
 854         return ret;
 855 }
 856 EXPORT_SYMBOL(obd_memory_max);
 857
 858 static void __exit obdclass_exit(void)
 859 {
 860         __u64 memory_leaked;
 861         __u64 memory_max;
 862         ENTRY;
 863
 864         misc_deregister(&obd_psdev);
 865 #ifdef HAVE_SERVER_SUPPORT
 866         lu_ucred_global_fini();
 867         dt_global_fini();
 868 #endif /* HAVE_SERVER_SUPPORT */
 869         obd_pool_fini();
 870         llog_info_fini();
 871         cl_global_fini();
 872         lu_global_fini();
 873
 874         obd_cleanup_caches();
 875
 876         class_procfs_clean();
 877
 878         class_handle_cleanup();
 879         class_del_uuid(NULL); /* Delete all UUIDs. */
 880         obd_zombie_impexp_stop();
 881         libcfs_kkuc_fini();
 882
 883         memory_leaked = obd_memory_sum();
 884         memory_max = obd_memory_max();
 885
 886         percpu_counter_destroy(&obd_memory);
 887         /* the below message is checked in test-framework.sh check_mem_leak() */
 888         CDEBUG((memory_leaked) ? D_ERROR : D_INFO,
 889                "obd_memory max: %llu, leaked: %llu\n",
 890                memory_max, memory_leaked);
 891
 892         unregister_oom_notifier(&obdclass_oom);
 893
 894         EXIT;
 895 }
 896
 897 void obd_heat_clear(struct obd_heat_instance *instance, int count)
 898 {
 899         ENTRY;
 900
 901         memset(instance, 0, sizeof(*instance) * count);
 902         RETURN_EXIT;
 903 }
 904 EXPORT_SYMBOL(obd_heat_clear);
 905
 906 /*
 907  * The file heat is calculated for every time interval period I. The access
 908  * frequency during each period is counted. The file heat is only recalculated
 909  * at the end of a time period.  And a percentage of the former file heat is
 910  * lost when recalculated. The recursion formula to calculate the heat of the
 911  * file f is as follow:
 912  *
 913  * Hi+1(f) = (1-P)*Hi(f)+ P*Ci
 914  *
 915  * Where Hi is the heat value in the period between time points i*I and
 916  * (i+1)*I; Ci is the access count in the period; the symbol P refers to the
 917  * weight of Ci. The larger the value the value of P is, the more influence Ci
 918  * has on the file heat.
 919  */
 920 void obd_heat_decay(struct obd_heat_instance *instance,  __u64 time_second,
 921                     unsigned int weight, unsigned int period_second)
 922 {
 923         u64 second;
 924
 925         ENTRY;
 926
 927         if (instance->ohi_time_second > time_second) {
 928                 obd_heat_clear(instance, 1);
 929                 RETURN_EXIT;
 930         }
 931
 932         if (instance->ohi_time_second == 0)
 933                 RETURN_EXIT;
 934
 935         for (second = instance->ohi_time_second + period_second;
 936              second < time_second;
 937              second += period_second) {
 938                 instance->ohi_heat = instance->ohi_heat *
 939                                 (256 - weight) / 256 +
 940                                 instance->ohi_count * weight / 256;
 941                 instance->ohi_count = 0;
 942                 instance->ohi_time_second = second;
 943         }
 944         RETURN_EXIT;
 945 }
 946 EXPORT_SYMBOL(obd_heat_decay);
 947
 948 __u64 obd_heat_get(struct obd_heat_instance *instance, unsigned int time_second,
 949                    unsigned int weight, unsigned int period_second)
 950 {
 951         ENTRY;
 952
 953         obd_heat_decay(instance, time_second, weight, period_second);
 954
 955         if (instance->ohi_count == 0)
 956                 RETURN(instance->ohi_heat);
 957
 958         RETURN(instance->ohi_heat * (256 - weight) / 256 +
 959                instance->ohi_count * weight / 256);
 960 }
 961 EXPORT_SYMBOL(obd_heat_get);
 962
 963 void obd_heat_add(struct obd_heat_instance *instance,
 964                   unsigned int time_second,  __u64 count,
 965                   unsigned int weight, unsigned int period_second)
 966 {
 967         ENTRY;
 968
 969         obd_heat_decay(instance, time_second, weight, period_second);
 970         if (instance->ohi_time_second == 0) {
 971                 instance->ohi_time_second = time_second;
 972                 instance->ohi_heat = 0;
 973                 instance->ohi_count = count;
 974         } else {
 975                 instance->ohi_count += count;
 976         }
 977         RETURN_EXIT;
 978 }
 979 EXPORT_SYMBOL(obd_heat_add);
 980
 981 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
 982 MODULE_DESCRIPTION("Lustre Class Driver");
 983 MODULE_VERSION(LUSTRE_VERSION_STRING);
 984 MODULE_LICENSE("GPL");
 985
 986 module_init(obdclass_init);
 987 module_exit(obdclass_exit);