lustre/obdclass/linux/linux-sysctl.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #include <linux/module.h>
  38 #include <linux/autoconf.h>
  39 #include <linux/sysctl.h>
  40 #include <linux/sched.h>
  41 #include <linux/mm.h>
  42 #include <linux/sysctl.h>
  43 #include <linux/version.h>
  44 #include <linux/proc_fs.h>
  45 #include <linux/slab.h>
  46 #include <linux/stat.h>
  47 #include <linux/ctype.h>
  48 #include <asm/bitops.h>
  49 #include <asm/uaccess.h>
  50 #include <linux/utsname.h>
  51
  52 #define DEBUG_SUBSYSTEM S_CLASS
  53
  54 #include <obd_support.h>
  55 #include <lprocfs_status.h>
  56
  57 cfs_sysctl_table_header_t *obd_table_header = NULL;
  58
  59 #ifndef HAVE_SYSCTL_UNNUMBERED
  60
  61 #define OBD_SYSCTL 300
  62
  63 enum {
  64         OBD_FAIL_LOC = 1,       /* control test failures instrumentation */
  65         OBD_FAIL_VAL,           /* userdata for fail loc */
  66         OBD_TIMEOUT,            /* RPC timeout before recovery/intr */
  67         OBD_DUMP_ON_TIMEOUT,    /* dump kernel debug log upon eviction */
  68         OBD_MEMUSED,            /* bytes currently OBD_ALLOCated */
  69         OBD_PAGESUSED,          /* pages currently OBD_PAGE_ALLOCated */
  70         OBD_MAXMEMUSED,         /* maximum bytes OBD_ALLOCated concurrently */
  71         OBD_MAXPAGESUSED,       /* maximum pages OBD_PAGE_ALLOCated concurrently */
  72         OBD_SYNCFILTER,         /* XXX temporary, as we play with sync osts.. */
  73         OBD_LDLM_TIMEOUT,       /* LDLM timeout for ASTs before client eviction */
  74         OBD_DUMP_ON_EVICTION,   /* dump kernel debug log upon eviction */
  75         OBD_DEBUG_PEER_ON_TIMEOUT, /* dump peer debug when RPC times out */
  76         OBD_ALLOC_FAIL_RATE,    /* memory allocation random failure rate */
  77         OBD_MAX_DIRTY_PAGES,    /* maximum dirty pages */
  78         OBD_AT_MIN,             /* Adaptive timeouts params */
  79         OBD_AT_MAX,
  80         OBD_AT_EXTRA,
  81         OBD_AT_EARLY_MARGIN,
  82         OBD_AT_HISTORY,
  83 };
  84
  85 #else
  86
  87 #define OBD_SYSCTL              CTL_UNNUMBERED
  88
  89 #define OBD_FAIL_LOC            CTL_UNNUMBERED
  90 #define OBD_FAIL_VAL            CTL_UNNUMBERED
  91 #define OBD_TIMEOUT             CTL_UNNUMBERED
  92 #define OBD_DUMP_ON_TIMEOUT     CTL_UNNUMBERED
  93 #define OBD_MEMUSED             CTL_UNNUMBERED
  94 #define OBD_PAGESUSED           CTL_UNNUMBERED
  95 #define OBD_MAXMEMUSED          CTL_UNNUMBERED
  96 #define OBD_MAXPAGESUSED        CTL_UNNUMBERED
  97 #define OBD_SYNCFILTER          CTL_UNNUMBERED
  98 #define OBD_LDLM_TIMEOUT        CTL_UNNUMBERED
  99 #define OBD_DUMP_ON_EVICTION    CTL_UNNUMBERED
 100 #define OBD_DEBUG_PEER_ON_TIMEOUT CTL_UNNUMBERED
 101 #define OBD_ALLOC_FAIL_RATE     CTL_UNNUMBERED
 102 #define OBD_MAX_DIRTY_PAGES     CTL_UNNUMBERED
 103 #define OBD_AT_MIN              CTL_UNNUMBERED
 104 #define OBD_AT_MAX              CTL_UNNUMBERED
 105 #define OBD_AT_EXTRA            CTL_UNNUMBERED
 106 #define OBD_AT_EARLY_MARGIN     CTL_UNNUMBERED
 107 #define OBD_AT_HISTORY          CTL_UNNUMBERED
 108
 109 #endif
 110
 111
 112 int LL_PROC_PROTO(proc_fail_loc)
 113 {
 114         int rc;
 115         long old_fail_loc = obd_fail_loc;
 116
 117         rc = ll_proc_dolongvec(table, write, filp, buffer, lenp, ppos);
 118         if (old_fail_loc != obd_fail_loc)
 119                 wake_up(&obd_race_waitq);
 120         return rc;
 121 }
 122
 123 int LL_PROC_PROTO(proc_set_timeout)
 124 {
 125         int rc;
 126
 127         rc = ll_proc_dointvec(table, write, filp, buffer, lenp, ppos);
 128         if (ldlm_timeout >= obd_timeout)
 129                 ldlm_timeout = max(obd_timeout / 3, 1U);
 130         return rc;
 131 }
 132
 133 int LL_PROC_PROTO(proc_memory_alloc)
 134 {
 135         char buf[22];
 136         int len;
 137         DECLARE_LL_PROC_PPOS_DECL;
 138
 139         if (!*lenp || (*ppos && !write)) {
 140                 *lenp = 0;
 141                 return 0;
 142         }
 143         if (write)
 144                 return -EINVAL;
 145
 146         len = snprintf(buf, sizeof(buf), LPU64"\n", obd_memory_sum());
 147         if (len > *lenp)
 148                 len = *lenp;
 149         buf[len] = '\0';
 150         if (copy_to_user(buffer, buf, len))
 151                 return -EFAULT;
 152         *lenp = len;
 153         *ppos += *lenp;
 154         return 0;
 155 }
 156
 157 int LL_PROC_PROTO(proc_pages_alloc)
 158 {
 159         char buf[22];
 160         int len;
 161         DECLARE_LL_PROC_PPOS_DECL;
 162
 163         if (!*lenp || (*ppos && !write)) {
 164                 *lenp = 0;
 165                 return 0;
 166         }
 167         if (write)
 168                 return -EINVAL;
 169
 170         len = snprintf(buf, sizeof(buf), LPU64"\n", obd_pages_sum());
 171         if (len > *lenp)
 172                 len = *lenp;
 173         buf[len] = '\0';
 174         if (copy_to_user(buffer, buf, len))
 175                 return -EFAULT;
 176         *lenp = len;
 177         *ppos += *lenp;
 178         return 0;
 179 }
 180
 181 int LL_PROC_PROTO(proc_mem_max)
 182 {
 183         char buf[22];
 184         int len;
 185         DECLARE_LL_PROC_PPOS_DECL;
 186
 187         if (!*lenp || (*ppos && !write)) {
 188                 *lenp = 0;
 189                 return 0;
 190         }
 191         if (write)
 192                 return -EINVAL;
 193
 194         len = snprintf(buf, sizeof(buf), LPU64"\n", obd_memory_max());
 195         if (len > *lenp)
 196                 len = *lenp;
 197         buf[len] = '\0';
 198         if (copy_to_user(buffer, buf, len))
 199                 return -EFAULT;
 200         *lenp = len;
 201         *ppos += *lenp;
 202         return 0;
 203 }
 204
 205 int LL_PROC_PROTO(proc_pages_max)
 206 {
 207         char buf[22];
 208         int len;
 209         DECLARE_LL_PROC_PPOS_DECL;
 210
 211         if (!*lenp || (*ppos && !write)) {
 212                 *lenp = 0;
 213                 return 0;
 214         }
 215         if (write)
 216                 return -EINVAL;
 217
 218         len = snprintf(buf, sizeof(buf), LPU64"\n", obd_pages_max());
 219         if (len > *lenp)
 220                 len = *lenp;
 221         buf[len] = '\0';
 222         if (copy_to_user(buffer, buf, len))
 223                 return -EFAULT;
 224         *lenp = len;
 225         *ppos += *lenp;
 226         return 0;
 227 }
 228
 229 int LL_PROC_PROTO(proc_max_dirty_pages_in_mb)
 230 {
 231         int rc = 0;
 232         DECLARE_LL_PROC_PPOS_DECL;
 233
 234         if (!table->data || !table->maxlen || !*lenp || (*ppos && !write)) {
 235                 *lenp = 0;
 236                 return 0;
 237         }
 238         if (write) {
 239                 rc = lprocfs_write_frac_helper(buffer, *lenp,
 240                                                (unsigned int*)table->data,
 241                                                1 << (20 - CFS_PAGE_SHIFT));
 242                 /* Don't allow them to let dirty pages exceed 90% of system memory,
 243                  * and set a hard minimum of 4MB. */
 244                 if (obd_max_dirty_pages > ((num_physpages / 10) * 9)) {
 245                         CERROR("Refusing to set max dirty pages to %u, which "
 246                                "is more than 90%% of available RAM; setting to %lu\n",
 247                                obd_max_dirty_pages, ((num_physpages / 10) * 9));
 248                         obd_max_dirty_pages = ((num_physpages / 10) * 9);
 249                 } else if (obd_max_dirty_pages < 4 << (20 - CFS_PAGE_SHIFT)) {
 250                         obd_max_dirty_pages = 4 << (20 - CFS_PAGE_SHIFT);
 251                 }
 252         } else {
 253                 char buf[21];
 254                 int len;
 255
 256                 len = lprocfs_read_frac_helper(buf, sizeof(buf),
 257                                                *(unsigned int*)table->data,
 258                                                1 << (20 - CFS_PAGE_SHIFT));
 259                 if (len > *lenp)
 260                         len = *lenp;
 261                 buf[len] = '\0';
 262                 if (copy_to_user(buffer, buf, len))
 263                         return -EFAULT;
 264                 *lenp = len;
 265         }
 266         *ppos += *lenp;
 267         return rc;
 268 }
 269
 270 #ifdef RANDOM_FAIL_ALLOC
 271 int LL_PROC_PROTO(proc_alloc_fail_rate)
 272 {
 273         int rc          = 0;
 274         DECLARE_LL_PROC_PPOS_DECL;
 275
 276         if (!table->data || !table->maxlen || !*lenp || (*ppos && !write)) {
 277                 *lenp = 0;
 278                 return 0;
 279         }
 280         if (write) {
 281                 rc = lprocfs_write_frac_helper(buffer, *lenp,
 282                                                (unsigned int*)table->data,
 283                                                OBD_ALLOC_FAIL_MULT);
 284         } else {
 285                 char buf[21];
 286                 int  len;
 287
 288                 len = lprocfs_read_frac_helper(buf, 21,
 289                                                *(unsigned int*)table->data,
 290                                                OBD_ALLOC_FAIL_MULT);
 291                 if (len > *lenp)
 292                         len = *lenp;
 293                 buf[len] = '\0';
 294                 if (copy_to_user(buffer, buf, len))
 295                         return -EFAULT;
 296                 *lenp = len;
 297         }
 298         *ppos += *lenp;
 299         return rc;
 300 }
 301 #endif
 302
 303 int LL_PROC_PROTO(proc_at_min)
 304 {
 305         return ll_proc_dointvec(table, write, filp, buffer, lenp, ppos);
 306 }
 307 int LL_PROC_PROTO(proc_at_max)
 308 {
 309         return ll_proc_dointvec(table, write, filp, buffer, lenp, ppos);
 310 }
 311 int LL_PROC_PROTO(proc_at_extra)
 312 {
 313         return ll_proc_dointvec(table, write, filp, buffer, lenp, ppos);
 314 }
 315 int LL_PROC_PROTO(proc_at_early_margin)
 316 {
 317         return ll_proc_dointvec(table, write, filp, buffer, lenp, ppos);
 318 }
 319 int LL_PROC_PROTO(proc_at_history)
 320 {
 321         return ll_proc_dointvec(table, write, filp, buffer, lenp, ppos);
 322 }
 323
 324 static cfs_sysctl_table_t obd_table[] = {
 325         {
 326                 .ctl_name = OBD_FAIL_LOC,
 327                 .procname = "fail_loc",
 328                 .data     = &obd_fail_loc,
 329                 .maxlen   = sizeof(obd_fail_loc),
 330                 .mode     = 0644,
 331                 .proc_handler = &proc_fail_loc
 332         },
 333         {
 334                 .ctl_name = OBD_FAIL_VAL,
 335                 .procname = "fail_val",
 336                 .data     = &obd_fail_val,
 337                 .maxlen   = sizeof(int),
 338                 .mode     = 0644,
 339                 .proc_handler = &proc_dointvec
 340         },
 341         {
 342                 .ctl_name = OBD_TIMEOUT,
 343                 .procname = "timeout",
 344                 .data     = &obd_timeout,
 345                 .maxlen   = sizeof(int),
 346                 .mode     = 0644,
 347                 .proc_handler = &proc_set_timeout
 348         },
 349         {
 350                 .ctl_name = OBD_DEBUG_PEER_ON_TIMEOUT,
 351                 .procname = "debug_peer_on_timeout",
 352                 .data     = &obd_debug_peer_on_timeout,
 353                 .maxlen   = sizeof(int),
 354                 .mode     = 0644,
 355                 .proc_handler = &proc_dointvec
 356         },
 357         {
 358                 .ctl_name = OBD_DUMP_ON_TIMEOUT,
 359                 .procname = "dump_on_timeout",
 360                 .data     = &obd_dump_on_timeout,
 361                 .maxlen   = sizeof(int),
 362                 .mode     = 0644,
 363                 .proc_handler = &proc_dointvec
 364         },
 365         {
 366                 .ctl_name = OBD_DUMP_ON_EVICTION,
 367                 .procname = "dump_on_eviction",
 368                 .data     = &obd_dump_on_eviction,
 369                 .maxlen   = sizeof(int),
 370                 .mode     = 0644,
 371                 .proc_handler = &proc_dointvec
 372         },
 373         {
 374                 .ctl_name = OBD_MEMUSED,
 375                 .procname = "memused",
 376                 .data     = NULL,
 377                 .maxlen   = 0,
 378                 .mode     = 0444,
 379                 .proc_handler = &proc_memory_alloc
 380         },
 381         {
 382                 .ctl_name = OBD_PAGESUSED,
 383                 .procname = "pagesused",
 384                 .data     = NULL,
 385                 .maxlen   = 0,
 386                 .mode     = 0444,
 387                 .proc_handler = &proc_pages_alloc
 388         },
 389         {
 390                 .ctl_name = OBD_MAXMEMUSED,
 391                 .procname = "memused_max",
 392                 .data     = NULL,
 393                 .maxlen   = 0,
 394                 .mode     = 0444,
 395                 .proc_handler = &proc_mem_max
 396         },
 397         {
 398                 .ctl_name = OBD_MAXPAGESUSED,
 399                 .procname = "pagesused_max",
 400                 .data     = NULL,
 401                 .maxlen   = 0,
 402                 .mode     = 0444,
 403                 .proc_handler = &proc_pages_max
 404         },
 405         {
 406                 .ctl_name = OBD_LDLM_TIMEOUT,
 407                 .procname = "ldlm_timeout",
 408                 .data     = &ldlm_timeout,
 409                 .maxlen   = sizeof(int),
 410                 .mode     = 0644,
 411                 .proc_handler = &proc_set_timeout
 412         },
 413 #ifdef RANDOM_FAIL_ALLOC
 414         {
 415                 .ctl_name = OBD_ALLOC_FAIL_RATE,
 416                 .procname = "alloc_fail_rate",
 417                 .data     = &obd_alloc_fail_rate,
 418                 .maxlen   = sizeof(int),
 419                 .mode     = 0644,
 420                 .proc_handler = &proc_alloc_fail_rate
 421         },
 422 #endif
 423         {
 424                 .ctl_name = OBD_MAX_DIRTY_PAGES,
 425                 .procname = "max_dirty_mb",
 426                 .data     = &obd_max_dirty_pages,
 427                 .maxlen   = sizeof(int),
 428                 .mode     = 0644,
 429                 .proc_handler = &proc_max_dirty_pages_in_mb
 430         },
 431         {
 432                 .ctl_name = OBD_AT_MIN,
 433                 .procname = "at_min",
 434                 .data     = &at_min,
 435                 .maxlen   = sizeof(int),
 436                 .mode     = 0644,
 437                 .proc_handler = &proc_at_min
 438         },
 439         {
 440                 .ctl_name = OBD_AT_MAX,
 441                 .procname = "at_max",
 442                 .data     = &at_max,
 443                 .maxlen   = sizeof(int),
 444                 .mode     = 0644,
 445                 .proc_handler = &proc_at_max
 446         },
 447         {
 448                 .ctl_name = OBD_AT_EXTRA,
 449                 .procname = "at_extra",
 450                 .data     = &at_extra,
 451                 .maxlen   = sizeof(int),
 452                 .mode     = 0644,
 453                 .proc_handler = &proc_at_extra
 454         },
 455         {
 456                 .ctl_name = OBD_AT_EARLY_MARGIN,
 457                 .procname = "at_early_margin",
 458                 .data     = &at_early_margin,
 459                 .maxlen   = sizeof(int),
 460                 .mode     = 0644,
 461                 .proc_handler = &proc_at_early_margin
 462         },
 463         {
 464                 .ctl_name = OBD_AT_HISTORY,
 465                 .procname = "at_history",
 466                 .data     = &at_history,
 467                 .maxlen   = sizeof(int),
 468                 .mode     = 0644,
 469                 .proc_handler = &proc_at_history
 470         },
 471         { 0 }
 472 };
 473
 474 static cfs_sysctl_table_t parent_table[] = {
 475         {
 476                 .ctl_name = OBD_SYSCTL,
 477                 .procname = "lustre",
 478                 .data     = NULL,
 479                 .maxlen   = 0,
 480                 .mode     = 0555,
 481                 .child    = obd_table
 482         },
 483         {0}
 484 };
 485
 486 void obd_sysctl_init (void)
 487 {
 488 #ifdef CONFIG_SYSCTL
 489         if ( !obd_table_header )
 490                 obd_table_header = cfs_register_sysctl_table(parent_table, 0);
 491 #endif
 492 }
 493
 494 void obd_sysctl_clean (void)
 495 {
 496 #ifdef CONFIG_SYSCTL
 497         if ( obd_table_header )
 498                 cfs_unregister_sysctl_table(obd_table_header);
 499         obd_table_header = NULL;
 500 #endif
 501 }