lib/libdevstat/devstat.c

   1 /*
   2  * Copyright (c) 1997, 1998 Kenneth D. Merry.
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. The name of the author may not be used to endorse or promote products
  14  *    derived from this software without specific prior written permission.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/types.h>
  33 #include <sys/sysctl.h>
  34 #include <sys/errno.h>
  35 #include <sys/resource.h>
  36 #include <sys/queue.h>
  37
  38 #include <ctype.h>
  39 #include <err.h>
  40 #include <fcntl.h>
  41 #include <limits.h>
  42 #include <stdio.h>
  43 #include <stdlib.h>
  44 #include <string.h>
  45 #include <stdarg.h>
  46 #include <kvm.h>
  47 #include <nlist.h>
  48
  49 #include "devstat.h"
  50
  51 int
  52 compute_stats(struct devstat *current, struct devstat *previous,
  53               long double etime, u_int64_t *total_bytes,
  54               u_int64_t *total_transfers, u_int64_t *total_blocks,
  55               long double *kb_per_transfer, long double *transfers_per_second,
  56               long double *mb_per_second, long double *blocks_per_second,
  57               long double *ms_per_transaction);
  58
  59 typedef enum {
  60         DEVSTAT_ARG_NOTYPE,
  61         DEVSTAT_ARG_UINT64,
  62         DEVSTAT_ARG_LD,
  63         DEVSTAT_ARG_SKIP
  64 } devstat_arg_type;
  65
  66 char devstat_errbuf[DEVSTAT_ERRBUF_SIZE];
  67
  68 /*
  69  * Table to match descriptive strings with device types.  These are in
  70  * order from most common to least common to speed search time.
  71  */
  72 struct devstat_match_table match_table[] = {
  73         {"da",          DEVSTAT_TYPE_DIRECT,    DEVSTAT_MATCH_TYPE},
  74         {"cd",          DEVSTAT_TYPE_CDROM,     DEVSTAT_MATCH_TYPE},
  75         {"scsi",        DEVSTAT_TYPE_IF_SCSI,   DEVSTAT_MATCH_IF},
  76         {"ide",         DEVSTAT_TYPE_IF_IDE,    DEVSTAT_MATCH_IF},
  77         {"other",       DEVSTAT_TYPE_IF_OTHER,  DEVSTAT_MATCH_IF},
  78         {"worm",        DEVSTAT_TYPE_WORM,      DEVSTAT_MATCH_TYPE},
  79         {"sa",          DEVSTAT_TYPE_SEQUENTIAL,DEVSTAT_MATCH_TYPE},
  80         {"pass",        DEVSTAT_TYPE_PASS,      DEVSTAT_MATCH_PASS},
  81         {"optical",     DEVSTAT_TYPE_OPTICAL,   DEVSTAT_MATCH_TYPE},
  82         {"array",       DEVSTAT_TYPE_STORARRAY, DEVSTAT_MATCH_TYPE},
  83         {"changer",     DEVSTAT_TYPE_CHANGER,   DEVSTAT_MATCH_TYPE},
  84         {"scanner",     DEVSTAT_TYPE_SCANNER,   DEVSTAT_MATCH_TYPE},
  85         {"printer",     DEVSTAT_TYPE_PRINTER,   DEVSTAT_MATCH_TYPE},
  86         {"floppy",      DEVSTAT_TYPE_FLOPPY,    DEVSTAT_MATCH_TYPE},
  87         {"proc",        DEVSTAT_TYPE_PROCESSOR, DEVSTAT_MATCH_TYPE},
  88         {"comm",        DEVSTAT_TYPE_COMM,      DEVSTAT_MATCH_TYPE},
  89         {"enclosure",   DEVSTAT_TYPE_ENCLOSURE, DEVSTAT_MATCH_TYPE},
  90         {NULL,          0,                      0}
  91 };
  92
  93 struct devstat_args {
  94         devstat_metric          metric;
  95         devstat_arg_type        argtype;
  96 } devstat_arg_list[] = {
  97         { DSM_NONE, DEVSTAT_ARG_NOTYPE },
  98         { DSM_TOTAL_BYTES, DEVSTAT_ARG_UINT64 },
  99         { DSM_TOTAL_BYTES_READ, DEVSTAT_ARG_UINT64 },
 100         { DSM_TOTAL_BYTES_WRITE, DEVSTAT_ARG_UINT64 },
 101         { DSM_TOTAL_TRANSFERS, DEVSTAT_ARG_UINT64 },
 102         { DSM_TOTAL_TRANSFERS_READ, DEVSTAT_ARG_UINT64 },
 103         { DSM_TOTAL_TRANSFERS_WRITE, DEVSTAT_ARG_UINT64 },
 104         { DSM_TOTAL_TRANSFERS_OTHER, DEVSTAT_ARG_UINT64 },
 105         { DSM_TOTAL_BLOCKS, DEVSTAT_ARG_UINT64 },
 106         { DSM_TOTAL_BLOCKS_READ, DEVSTAT_ARG_UINT64 },
 107         { DSM_TOTAL_BLOCKS_WRITE, DEVSTAT_ARG_UINT64 },
 108         { DSM_KB_PER_TRANSFER, DEVSTAT_ARG_LD },
 109         { DSM_KB_PER_TRANSFER_READ, DEVSTAT_ARG_LD },
 110         { DSM_KB_PER_TRANSFER_WRITE, DEVSTAT_ARG_LD },
 111         { DSM_TRANSFERS_PER_SECOND, DEVSTAT_ARG_LD },
 112         { DSM_TRANSFERS_PER_SECOND_READ, DEVSTAT_ARG_LD },
 113         { DSM_TRANSFERS_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
 114         { DSM_TRANSFERS_PER_SECOND_OTHER, DEVSTAT_ARG_LD },
 115         { DSM_MB_PER_SECOND, DEVSTAT_ARG_LD },
 116         { DSM_MB_PER_SECOND_READ, DEVSTAT_ARG_LD },
 117         { DSM_MB_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
 118         { DSM_BLOCKS_PER_SECOND, DEVSTAT_ARG_LD },
 119         { DSM_BLOCKS_PER_SECOND_READ, DEVSTAT_ARG_LD },
 120         { DSM_BLOCKS_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
 121         { DSM_MS_PER_TRANSACTION, DEVSTAT_ARG_LD },
 122         { DSM_MS_PER_TRANSACTION_READ, DEVSTAT_ARG_LD },
 123         { DSM_MS_PER_TRANSACTION_WRITE, DEVSTAT_ARG_LD },
 124         { DSM_SKIP, DEVSTAT_ARG_SKIP },
 125         { DSM_TOTAL_BYTES_FREE, DEVSTAT_ARG_UINT64 },
 126         { DSM_TOTAL_TRANSFERS_FREE, DEVSTAT_ARG_UINT64 },
 127         { DSM_TOTAL_BLOCKS_FREE, DEVSTAT_ARG_UINT64 },
 128         { DSM_KB_PER_TRANSFER_FREE, DEVSTAT_ARG_LD },
 129         { DSM_MB_PER_SECOND_FREE, DEVSTAT_ARG_LD },
 130         { DSM_TRANSFERS_PER_SECOND_FREE, DEVSTAT_ARG_LD },
 131         { DSM_BLOCKS_PER_SECOND_FREE, DEVSTAT_ARG_LD },
 132         { DSM_MS_PER_TRANSACTION_OTHER, DEVSTAT_ARG_LD },
 133         { DSM_MS_PER_TRANSACTION_FREE, DEVSTAT_ARG_LD },
 134         { DSM_BUSY_PCT, DEVSTAT_ARG_LD },
 135         { DSM_QUEUE_LENGTH, DEVSTAT_ARG_UINT64 },
 136         { DSM_TOTAL_DURATION, DEVSTAT_ARG_LD },
 137         { DSM_TOTAL_DURATION_READ, DEVSTAT_ARG_LD },
 138         { DSM_TOTAL_DURATION_WRITE, DEVSTAT_ARG_LD },
 139         { DSM_TOTAL_DURATION_FREE, DEVSTAT_ARG_LD },
 140         { DSM_TOTAL_DURATION_OTHER, DEVSTAT_ARG_LD },
 141         { DSM_TOTAL_BUSY_TIME, DEVSTAT_ARG_LD },
 142 };
 143
 144 static const char *namelist[] = {
 145 #define X_NUMDEVS       0
 146         "_devstat_num_devs",
 147 #define X_GENERATION    1
 148         "_devstat_generation",
 149 #define X_VERSION       2
 150         "_devstat_version",
 151 #define X_DEVICE_STATQ  3
 152         "_device_statq",
 153 #define X_TIME_UPTIME   4
 154         "_time_uptime",
 155 #define X_END           5
 156 };
 157
 158 /*
 159  * Local function declarations.
 160  */
 161 static int compare_select(const void *arg1, const void *arg2);
 162 static int readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes);
 163 static int readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes);
 164 static char *get_devstat_kvm(kvm_t *kd);
 165
 166 #define KREADNL(kd, var, val) \
 167         readkmem_nl(kd, namelist[var], &val, sizeof(val))
 168
 169 int
 170 devstat_getnumdevs(kvm_t *kd)
 171 {
 172         size_t numdevsize;
 173         int numdevs;
 174
 175         numdevsize = sizeof(int);
 176
 177         /*
 178          * Find out how many devices we have in the system.
 179          */
 180         if (kd == NULL) {
 181                 if (sysctlbyname("kern.devstat.numdevs", &numdevs,
 182                                  &numdevsize, NULL, 0) == -1) {
 183                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 184                                  "%s: error getting number of devices\n"
 185                                  "%s: %s", __func__, __func__,
 186                                  strerror(errno));
 187                         return(-1);
 188                 } else
 189                         return(numdevs);
 190         } else {
 191
 192                 if (KREADNL(kd, X_NUMDEVS, numdevs) == -1)
 193                         return(-1);
 194                 else
 195                         return(numdevs);
 196         }
 197 }
 198
 199 /*
 200  * This is an easy way to get the generation number, but the generation is
 201  * supplied in a more atmoic manner by the kern.devstat.all sysctl.
 202  * Because this generation sysctl is separate from the statistics sysctl,
 203  * the device list and the generation could change between the time that
 204  * this function is called and the device list is retrieved.
 205  */
 206 long
 207 devstat_getgeneration(kvm_t *kd)
 208 {
 209         size_t gensize;
 210         long generation;
 211
 212         gensize = sizeof(long);
 213
 214         /*
 215          * Get the current generation number.
 216          */
 217         if (kd == NULL) {
 218                 if (sysctlbyname("kern.devstat.generation", &generation,
 219                                  &gensize, NULL, 0) == -1) {
 220                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 221                                  "%s: error getting devstat generation\n%s: %s",
 222                                  __func__, __func__, strerror(errno));
 223                         return(-1);
 224                 } else
 225                         return(generation);
 226         } else {
 227                 if (KREADNL(kd, X_GENERATION, generation) == -1)
 228                         return(-1);
 229                 else
 230                         return(generation);
 231         }
 232 }
 233
 234 /*
 235  * Get the current devstat version.  The return value of this function
 236  * should be compared with DEVSTAT_VERSION, which is defined in
 237  * sys/devicestat.h.  This will enable userland programs to determine
 238  * whether they are out of sync with the kernel.
 239  */
 240 int
 241 devstat_getversion(kvm_t *kd)
 242 {
 243         size_t versize;
 244         int version;
 245
 246         versize = sizeof(int);
 247
 248         /*
 249          * Get the current devstat version.
 250          */
 251         if (kd == NULL) {
 252                 if (sysctlbyname("kern.devstat.version", &version, &versize,
 253                                  NULL, 0) == -1) {
 254                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 255                                  "%s: error getting devstat version\n%s: %s",
 256                                  __func__, __func__, strerror(errno));
 257                         return(-1);
 258                 } else
 259                         return(version);
 260         } else {
 261                 if (KREADNL(kd, X_VERSION, version) == -1)
 262                         return(-1);
 263                 else
 264                         return(version);
 265         }
 266 }
 267
 268 /*
 269  * Check the devstat version we know about against the devstat version the
 270  * kernel knows about.  If they don't match, print an error into the
 271  * devstat error buffer, and return -1.  If they match, return 0.
 272  */
 273 int
 274 devstat_checkversion(kvm_t *kd)
 275 {
 276         int buflen, res, retval = 0, version;
 277
 278         version = devstat_getversion(kd);
 279
 280         if (version != DEVSTAT_VERSION) {
 281                 /*
 282                  * If getversion() returns an error (i.e. -1), then it
 283                  * has printed an error message in the buffer.  Therefore,
 284                  * we need to add a \n to the end of that message before we
 285                  * print our own message in the buffer.
 286                  */
 287                 if (version == -1)
 288                         buflen = strlen(devstat_errbuf);
 289                 else
 290                         buflen = 0;
 291
 292                 res = snprintf(devstat_errbuf + buflen,
 293                                DEVSTAT_ERRBUF_SIZE - buflen,
 294                                "%s%s: userland devstat version %d is not "
 295                                "the same as the kernel\n%s: devstat "
 296                                "version %d\n", version == -1 ? "\n" : "",
 297                                __func__, DEVSTAT_VERSION, __func__, version);
 298
 299                 if (res < 0)
 300                         devstat_errbuf[buflen] = '\0';
 301
 302                 buflen = strlen(devstat_errbuf);
 303                 if (version < DEVSTAT_VERSION)
 304                         res = snprintf(devstat_errbuf + buflen,
 305                                        DEVSTAT_ERRBUF_SIZE - buflen,
 306                                        "%s: libdevstat newer than kernel\n",
 307                                        __func__);
 308                 else
 309                         res = snprintf(devstat_errbuf + buflen,
 310                                        DEVSTAT_ERRBUF_SIZE - buflen,
 311                                        "%s: kernel newer than libdevstat\n",
 312                                        __func__);
 313
 314                 if (res < 0)
 315                         devstat_errbuf[buflen] = '\0';
 316
 317                 retval = -1;
 318         }
 319
 320         return(retval);
 321 }
 322
 323 /*
 324  * Get the current list of devices and statistics, and the current
 325  * generation number.
 326  *
 327  * Return values:
 328  * -1  -- error
 329  *  0  -- device list is unchanged
 330  *  1  -- device list has changed
 331  */
 332 int
 333 devstat_getdevs(kvm_t *kd, struct statinfo *stats)
 334 {
 335         int error;
 336         size_t dssize;
 337         long oldgeneration;
 338         int retval = 0;
 339         struct devinfo *dinfo;
 340         struct timespec ts;
 341
 342         dinfo = stats->dinfo;
 343
 344         if (dinfo == NULL) {
 345                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 346                          "%s: stats->dinfo was NULL", __func__);
 347                 return(-1);
 348         }
 349
 350         oldgeneration = dinfo->generation;
 351
 352         if (kd == NULL) {
 353                 clock_gettime(CLOCK_MONOTONIC, &ts);
 354                 stats->snap_time = ts.tv_sec + ts.tv_nsec * 1e-9;
 355
 356                 /* If this is our first time through, mem_ptr will be null. */
 357                 if (dinfo->mem_ptr == NULL) {
 358                         /*
 359                          * Get the number of devices.  If it's negative, it's an
 360                          * error.  Don't bother setting the error string, since
 361                          * getnumdevs() has already done that for us.
 362                          */
 363                         if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 364                                 return(-1);
 365
 366                         /*
 367                          * The kern.devstat.all sysctl returns the current
 368                          * generation number, as well as all the devices.
 369                          * So we need four bytes more.
 370                          */
 371                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 372                                  sizeof(long);
 373                         dinfo->mem_ptr = (u_int8_t *)malloc(dssize);
 374                         if (dinfo->mem_ptr == NULL) {
 375                                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 376                                          "%s: Cannot allocate memory for mem_ptr element",
 377                                          __func__);
 378                                 return(-1);
 379                         }
 380                 } else
 381                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 382                                  sizeof(long);
 383
 384                 /*
 385                  * Request all of the devices.  We only really allow for one
 386                  * ENOMEM failure.  It would, of course, be possible to just go
 387                  * in a loop and keep reallocing the device structure until we
 388                  * don't get ENOMEM back.  I'm not sure it's worth it, though.
 389                  * If devices are being added to the system that quickly, maybe
 390                  * the user can just wait until all devices are added.
 391                  */
 392                 for (;;) {
 393                         error = sysctlbyname("kern.devstat.all",
 394                                              dinfo->mem_ptr,
 395                                              &dssize, NULL, 0);
 396                         if (error != -1 || errno != EBUSY)
 397                                 break;
 398                 }
 399                 if (error == -1) {
 400                         /*
 401                          * If we get ENOMEM back, that means that there are
 402                          * more devices now, so we need to allocate more
 403                          * space for the device array.
 404                          */
 405                         if (errno == ENOMEM) {
 406                                 /*
 407                                  * No need to set the error string here,
 408                                  * devstat_getnumdevs() will do that if it fails.
 409                                  */
 410                                 if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 411                                         return(-1);
 412
 413                                 dssize = (dinfo->numdevs *
 414                                         sizeof(struct devstat)) + sizeof(long);
 415                                 dinfo->mem_ptr = (u_int8_t *)
 416                                         realloc(dinfo->mem_ptr, dssize);
 417                                 if ((error = sysctlbyname("kern.devstat.all",
 418                                     dinfo->mem_ptr, &dssize, NULL, 0)) == -1) {
 419                                         snprintf(devstat_errbuf,
 420                                                  sizeof(devstat_errbuf),
 421                                                  "%s: error getting device "
 422                                                  "stats\n%s: %s", __func__,
 423                                                  __func__, strerror(errno));
 424                                         return(-1);
 425                                 }
 426                         } else {
 427                                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 428                                          "%s: error getting device stats\n"
 429                                          "%s: %s", __func__, __func__,
 430                                          strerror(errno));
 431                                 return(-1);
 432                         }
 433                 }
 434
 435         } else {
 436                 if (KREADNL(kd, X_TIME_UPTIME, ts.tv_sec) == -1)
 437                         return(-1);
 438                 else
 439                         stats->snap_time = ts.tv_sec;
 440
 441                 /*
 442                  * This is of course non-atomic, but since we are working
 443                  * on a core dump, the generation is unlikely to change
 444                  */
 445                 if ((dinfo->numdevs = devstat_getnumdevs(kd)) == -1)
 446                         return(-1);
 447                 if ((dinfo->mem_ptr = (u_int8_t *)get_devstat_kvm(kd)) == NULL)
 448                         return(-1);
 449         }
 450         /*
 451          * The sysctl spits out the generation as the first four bytes,
 452          * then all of the device statistics structures.
 453          */
 454         dinfo->generation = *(long *)dinfo->mem_ptr;
 455
 456         /*
 457          * If the generation has changed, and if the current number of
 458          * devices is not the same as the number of devices recorded in the
 459          * devinfo structure, it is likely that the device list has shrunk.
 460          * The reason that it is likely that the device list has shrunk in
 461          * this case is that if the device list has grown, the sysctl above
 462          * will return an ENOMEM error, and we will reset the number of
 463          * devices and reallocate the device array.  If the second sysctl
 464          * fails, we will return an error and therefore never get to this
 465          * point.  If the device list has shrunk, the sysctl will not
 466          * return an error since we have more space allocated than is
 467          * necessary.  So, in the shrinkage case, we catch it here and
 468          * reallocate the array so that we don't use any more space than is
 469          * necessary.
 470          */
 471         if (oldgeneration != dinfo->generation) {
 472                 if (devstat_getnumdevs(kd) != dinfo->numdevs) {
 473                         if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 474                                 return(-1);
 475                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 476                                 sizeof(long);
 477                         dinfo->mem_ptr = (u_int8_t *)realloc(dinfo->mem_ptr,
 478                                                              dssize);
 479                 }
 480                 retval = 1;
 481         }
 482
 483         dinfo->devices = (struct devstat *)(dinfo->mem_ptr + sizeof(long));
 484
 485         return(retval);
 486 }
 487
 488 /*
 489  * selectdevs():
 490  *
 491  * Devices are selected/deselected based upon the following criteria:
 492  * - devices specified by the user on the command line
 493  * - devices matching any device type expressions given on the command line
 494  * - devices with the highest I/O, if 'top' mode is enabled
 495  * - the first n unselected devices in the device list, if maxshowdevs
 496  *   devices haven't already been selected and if the user has not
 497  *   specified any devices on the command line and if we're in "add" mode.
 498  *
 499  * Input parameters:
 500  * - device selection list (dev_select)
 501  * - current number of devices selected (num_selected)
 502  * - total number of devices in the selection list (num_selections)
 503  * - devstat generation as of the last time selectdevs() was called
 504  *   (select_generation)
 505  * - current devstat generation (current_generation)
 506  * - current list of devices and statistics (devices)
 507  * - number of devices in the current device list (numdevs)
 508  * - compiled version of the command line device type arguments (matches)
 509  *   - This is optional.  If the number of devices is 0, this will be ignored.
 510  *   - The matching code pays attention to the current selection mode.  So
 511  *     if you pass in a matching expression, it will be evaluated based
 512  *     upon the selection mode that is passed in.  See below for details.
 513  * - number of device type matching expressions (num_matches)
 514  *   - Set to 0 to disable the matching code.
 515  * - list of devices specified on the command line by the user (dev_selections)
 516  * - number of devices selected on the command line by the user
 517  *   (num_dev_selections)
 518  * - Our selection mode.  There are four different selection modes:
 519  *      - add mode.  (DS_SELECT_ADD) Any devices matching devices explicitly
 520  *        selected by the user or devices matching a pattern given by the
 521  *        user will be selected in addition to devices that are already
 522  *        selected.  Additional devices will be selected, up to maxshowdevs
 523  *        number of devices.
 524  *      - only mode. (DS_SELECT_ONLY)  Only devices matching devices
 525  *        explicitly given by the user or devices matching a pattern
 526  *        given by the user will be selected.  No other devices will be
 527  *        selected.
 528  *      - addonly mode.  (DS_SELECT_ADDONLY)  This is similar to add and
 529  *        only.  Basically, this will not de-select any devices that are
 530  *        current selected, as only mode would, but it will also not
 531  *        gratuitously select up to maxshowdevs devices as add mode would.
 532  *      - remove mode.  (DS_SELECT_REMOVE)  Any devices matching devices
 533  *        explicitly selected by the user or devices matching a pattern
 534  *        given by the user will be de-selected.
 535  * - maximum number of devices we can select (maxshowdevs)
 536  * - flag indicating whether or not we're in 'top' mode (perf_select)
 537  *
 538  * Output data:
 539  * - the device selection list may be modified and passed back out
 540  * - the number of devices selected and the total number of items in the
 541  *   device selection list may be changed
 542  * - the selection generation may be changed to match the current generation
 543  *
 544  * Return values:
 545  * -1  -- error
 546  *  0  -- selected devices are unchanged
 547  *  1  -- selected devices changed
 548  */
 549 int
 550 devstat_selectdevs(struct device_selection **dev_select, int *num_selected,
 551                    int *num_selections, long *select_generation,
 552                    long current_generation, struct devstat *devices,
 553                    int numdevs, struct devstat_match *matches, int num_matches,
 554                    char **dev_selections, int num_dev_selections,
 555                    devstat_select_mode select_mode, int maxshowdevs,
 556                    int perf_select)
 557 {
 558         int i, j, k;
 559         int init_selections = 0, init_selected_var = 0;
 560         struct device_selection *old_dev_select = NULL;
 561         int old_num_selections = 0, old_num_selected;
 562         int selection_number = 0;
 563         int changed = 0, found = 0;
 564
 565         if ((dev_select == NULL) || (devices == NULL) || (numdevs < 0))
 566                 return(-1);
 567
 568         /*
 569          * We always want to make sure that we have as many dev_select
 570          * entries as there are devices.
 571          */
 572         /*
 573          * In this case, we haven't selected devices before.
 574          */
 575         if (*dev_select == NULL) {
 576                 *dev_select = (struct device_selection *)malloc(numdevs *
 577                         sizeof(struct device_selection));
 578                 *select_generation = current_generation;
 579                 init_selections = 1;
 580                 changed = 1;
 581         /*
 582          * In this case, we have selected devices before, but the device
 583          * list has changed since we last selected devices, so we need to
 584          * either enlarge or reduce the size of the device selection list.
 585          */
 586         } else if (*num_selections != numdevs) {
 587                 *dev_select = (struct device_selection *)reallocf(*dev_select,
 588                         numdevs * sizeof(struct device_selection));
 589                 *select_generation = current_generation;
 590                 init_selections = 1;
 591         /*
 592          * In this case, we've selected devices before, and the selection
 593          * list is the same size as it was the last time, but the device
 594          * list has changed.
 595          */
 596         } else if (*select_generation < current_generation) {
 597                 *select_generation = current_generation;
 598                 init_selections = 1;
 599         }
 600
 601         if (*dev_select == NULL) {
 602                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 603                          "%s: Cannot (re)allocate memory for dev_select argument",
 604                          __func__);
 605                 return(-1);
 606         }
 607
 608         /*
 609          * If we're in "only" mode, we want to clear out the selected
 610          * variable since we're going to select exactly what the user wants
 611          * this time through.
 612          */
 613         if (select_mode == DS_SELECT_ONLY)
 614                 init_selected_var = 1;
 615
 616         /*
 617          * In all cases, we want to back up the number of selected devices.
 618          * It is a quick and accurate way to determine whether the selected
 619          * devices have changed.
 620          */
 621         old_num_selected = *num_selected;
 622
 623         /*
 624          * We want to make a backup of the current selection list if
 625          * the list of devices has changed, or if we're in performance
 626          * selection mode.  In both cases, we don't want to make a backup
 627          * if we already know for sure that the list will be different.
 628          * This is certainly the case if this is our first time through the
 629          * selection code.
 630          */
 631         if (((init_selected_var != 0) || (init_selections != 0)
 632          || (perf_select != 0)) && (changed == 0)){
 633                 old_dev_select = (struct device_selection *)malloc(
 634                     *num_selections * sizeof(struct device_selection));
 635                 if (old_dev_select == NULL) {
 636                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 637                                  "%s: Cannot allocate memory for selection list backup",
 638                                  __func__);
 639                         return(-1);
 640                 }
 641                 old_num_selections = *num_selections;
 642                 bcopy(*dev_select, old_dev_select,
 643                     sizeof(struct device_selection) * *num_selections);
 644         }
 645
 646         if (init_selections != 0) {
 647                 bzero(*dev_select, sizeof(struct device_selection) * numdevs);
 648
 649                 for (i = 0; i < numdevs; i++) {
 650                         (*dev_select)[i].device_number =
 651                                 devices[i].device_number;
 652                         strncpy((*dev_select)[i].device_name,
 653                                 devices[i].device_name,
 654                                 DEVSTAT_NAME_LEN);
 655                         (*dev_select)[i].device_name[DEVSTAT_NAME_LEN - 1]='\0';
 656                         (*dev_select)[i].unit_number = devices[i].unit_number;
 657                         (*dev_select)[i].position = i;
 658                 }
 659                 *num_selections = numdevs;
 660         } else if (init_selected_var != 0) {
 661                 for (i = 0; i < numdevs; i++)
 662                         (*dev_select)[i].selected = 0;
 663         }
 664
 665         /* we haven't gotten around to selecting anything yet.. */
 666         if ((select_mode == DS_SELECT_ONLY) || (init_selections != 0)
 667          || (init_selected_var != 0))
 668                 *num_selected = 0;
 669
 670         /*
 671          * Look through any devices the user specified on the command line
 672          * and see if they match known devices.  If so, select them.
 673          */
 674         for (i = 0; (i < *num_selections) && (num_dev_selections > 0); i++) {
 675                 char tmpstr[80];
 676
 677                 snprintf(tmpstr, sizeof(tmpstr), "%s%d",
 678                          (*dev_select)[i].device_name,
 679                          (*dev_select)[i].unit_number);
 680                 for (j = 0; j < num_dev_selections; j++) {
 681                         if (strcmp(tmpstr, dev_selections[j]) == 0) {
 682                                 /*
 683                                  * Here we do different things based on the
 684                                  * mode we're in.  If we're in add or
 685                                  * addonly mode, we only select this device
 686                                  * if it hasn't already been selected.
 687                                  * Otherwise, we would be unnecessarily
 688                                  * changing the selection order and
 689                                  * incrementing the selection count.  If
 690                                  * we're in only mode, we unconditionally
 691                                  * select this device, since in only mode
 692                                  * any previous selections are erased and
 693                                  * manually specified devices are the first
 694                                  * ones to be selected.  If we're in remove
 695                                  * mode, we de-select the specified device and
 696                                  * decrement the selection count.
 697                                  */
 698                                 switch(select_mode) {
 699                                 case DS_SELECT_ADD:
 700                                 case DS_SELECT_ADDONLY:
 701                                         if ((*dev_select)[i].selected)
 702                                                 break;
 703                                         /* FALLTHROUGH */
 704                                 case DS_SELECT_ONLY:
 705                                         (*dev_select)[i].selected =
 706                                                 ++selection_number;
 707                                         (*num_selected)++;
 708                                         break;
 709                                 case DS_SELECT_REMOVE:
 710                                         (*dev_select)[i].selected = 0;
 711                                         (*num_selected)--;
 712                                         /*
 713                                          * This isn't passed back out, we
 714                                          * just use it to keep track of
 715                                          * how many devices we've removed.
 716                                          */
 717                                         num_dev_selections--;
 718                                         break;
 719                                 }
 720                                 break;
 721                         }
 722                 }
 723         }
 724
 725         /*
 726          * Go through the user's device type expressions and select devices
 727          * accordingly.  We only do this if the number of devices already
 728          * selected is less than the maximum number we can show.
 729          */
 730         for (i = 0; (i < num_matches) && (*num_selected < maxshowdevs); i++) {
 731                 /* We should probably indicate some error here */
 732                 if ((matches[i].match_fields == DEVSTAT_MATCH_NONE)
 733                  || (matches[i].num_match_categories <= 0))
 734                         continue;
 735
 736                 for (j = 0; j < numdevs; j++) {
 737                         int num_match_categories;
 738
 739                         num_match_categories = matches[i].num_match_categories;
 740
 741                         /*
 742                          * Determine whether or not the current device
 743                          * matches the given matching expression.  This if
 744                          * statement consists of three components:
 745                          *   - the device type check
 746                          *   - the device interface check
 747                          *   - the passthrough check
 748                          * If a the matching test is successful, it
 749                          * decrements the number of matching categories,
 750                          * and if we've reached the last element that
 751                          * needed to be matched, the if statement succeeds.
 752                          *
 753                          */
 754                         if ((((matches[i].match_fields & DEVSTAT_MATCH_TYPE)!=0)
 755                           && ((devices[j].device_type & DEVSTAT_TYPE_MASK) ==
 756                                 (matches[i].device_type & DEVSTAT_TYPE_MASK))
 757                           &&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 758                            || (((matches[i].match_fields &
 759                                 DEVSTAT_MATCH_PASS) == 0)
 760                             && ((devices[j].device_type &
 761                                 DEVSTAT_TYPE_PASS) == 0)))
 762                           && (--num_match_categories == 0))
 763                          || (((matches[i].match_fields & DEVSTAT_MATCH_IF) != 0)
 764                           && ((devices[j].device_type & DEVSTAT_TYPE_IF_MASK) ==
 765                                 (matches[i].device_type & DEVSTAT_TYPE_IF_MASK))
 766                           &&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 767                            || (((matches[i].match_fields &
 768                                 DEVSTAT_MATCH_PASS) == 0)
 769                             && ((devices[j].device_type &
 770                                 DEVSTAT_TYPE_PASS) == 0)))
 771                           && (--num_match_categories == 0))
 772                          || (((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 773                           && ((devices[j].device_type & DEVSTAT_TYPE_PASS) != 0)
 774                           && (--num_match_categories == 0))) {
 775
 776                                 /*
 777                                  * This is probably a non-optimal solution
 778                                  * to the problem that the devices in the
 779                                  * device list will not be in the same
 780                                  * order as the devices in the selection
 781                                  * array.
 782                                  */
 783                                 for (k = 0; k < numdevs; k++) {
 784                                         if ((*dev_select)[k].position == j) {
 785                                                 found = 1;
 786                                                 break;
 787                                         }
 788                                 }
 789
 790                                 /*
 791                                  * There shouldn't be a case where a device
 792                                  * in the device list is not in the
 793                                  * selection list...but it could happen.
 794                                  */
 795                                 if (found != 1) {
 796                                         fprintf(stderr, "selectdevs: couldn't"
 797                                                 " find %s%d in selection "
 798                                                 "list\n",
 799                                                 devices[j].device_name,
 800                                                 devices[j].unit_number);
 801                                         break;
 802                                 }
 803
 804                                 /*
 805                                  * We do different things based upon the
 806                                  * mode we're in.  If we're in add or only
 807                                  * mode, we go ahead and select this device
 808                                  * if it hasn't already been selected.  If
 809                                  * it has already been selected, we leave
 810                                  * it alone so we don't mess up the
 811                                  * selection ordering.  Manually specified
 812                                  * devices have already been selected, and
 813                                  * they have higher priority than pattern
 814                                  * matched devices.  If we're in remove
 815                                  * mode, we de-select the given device and
 816                                  * decrement the selected count.
 817                                  */
 818                                 switch(select_mode) {
 819                                 case DS_SELECT_ADD:
 820                                 case DS_SELECT_ADDONLY:
 821                                 case DS_SELECT_ONLY:
 822                                         if ((*dev_select)[k].selected != 0)
 823                                                 break;
 824                                         (*dev_select)[k].selected =
 825                                                 ++selection_number;
 826                                         (*num_selected)++;
 827                                         break;
 828                                 case DS_SELECT_REMOVE:
 829                                         (*dev_select)[k].selected = 0;
 830                                         (*num_selected)--;
 831                                         break;
 832                                 }
 833                         }
 834                 }
 835         }
 836
 837         /*
 838          * Here we implement "top" mode.  Devices are sorted in the
 839          * selection array based on two criteria:  whether or not they are
 840          * selected (not selection number, just the fact that they are
 841          * selected!) and the number of bytes in the "bytes" field of the
 842          * selection structure.  The bytes field generally must be kept up
 843          * by the user.  In the future, it may be maintained by library
 844          * functions, but for now the user has to do the work.
 845          *
 846          * At first glance, it may seem wrong that we don't go through and
 847          * select every device in the case where the user hasn't specified
 848          * any devices or patterns.  In fact, though, it won't make any
 849          * difference in the device sorting.  In that particular case (i.e.
 850          * when we're in "add" or "only" mode, and the user hasn't
 851          * specified anything) the first time through no devices will be
 852          * selected, so the only criterion used to sort them will be their
 853          * performance.  The second time through, and every time thereafter,
 854          * all devices will be selected, so again selection won't matter.
 855          */
 856         if (perf_select != 0) {
 857
 858                 /* Sort the device array by throughput  */
 859                 qsort(*dev_select, *num_selections,
 860                       sizeof(struct device_selection),
 861                       compare_select);
 862
 863                 if (*num_selected == 0) {
 864                         /*
 865                          * Here we select every device in the array, if it
 866                          * isn't already selected.  Because the 'selected'
 867                          * variable in the selection array entries contains
 868                          * the selection order, the devstats routine can show
 869                          * the devices that were selected first.
 870                          */
 871                         for (i = 0; i < *num_selections; i++) {
 872                                 if ((*dev_select)[i].selected == 0) {
 873                                         (*dev_select)[i].selected =
 874                                                 ++selection_number;
 875                                         (*num_selected)++;
 876                                 }
 877                         }
 878                 } else {
 879                         selection_number = 0;
 880                         for (i = 0; i < *num_selections; i++) {
 881                                 if ((*dev_select)[i].selected != 0) {
 882                                         (*dev_select)[i].selected =
 883                                                 ++selection_number;
 884                                 }
 885                         }
 886                 }
 887         }
 888
 889         /*
 890          * If we're in the "add" selection mode and if we haven't already
 891          * selected maxshowdevs number of devices, go through the array and
 892          * select any unselected devices.  If we're in "only" mode, we
 893          * obviously don't want to select anything other than what the user
 894          * specifies.  If we're in "remove" mode, it probably isn't a good
 895          * idea to go through and select any more devices, since we might
 896          * end up selecting something that the user wants removed.  Through
 897          * more complicated logic, we could actually figure this out, but
 898          * that would probably require combining this loop with the various
 899          * selections loops above.
 900          */
 901         if ((select_mode == DS_SELECT_ADD) && (*num_selected < maxshowdevs)) {
 902                 for (i = 0; i < *num_selections; i++)
 903                         if ((*dev_select)[i].selected == 0) {
 904                                 (*dev_select)[i].selected = ++selection_number;
 905                                 (*num_selected)++;
 906                         }
 907         }
 908
 909         /*
 910          * Look at the number of devices that have been selected.  If it
 911          * has changed, set the changed variable.  Otherwise, if we've
 912          * made a backup of the selection list, compare it to the current
 913          * selection list to see if the selected devices have changed.
 914          */
 915         if ((changed == 0) && (old_num_selected != *num_selected))
 916                 changed = 1;
 917         else if ((changed == 0) && (old_dev_select != NULL)) {
 918                 /*
 919                  * Now we go through the selection list and we look at
 920                  * it three different ways.
 921                  */
 922                 for (i = 0; (i < *num_selections) && (changed == 0) &&
 923                      (i < old_num_selections); i++) {
 924                         /*
 925                          * If the device at index i in both the new and old
 926                          * selection arrays has the same device number and
 927                          * selection status, it hasn't changed.  We
 928                          * continue on to the next index.
 929                          */
 930                         if (((*dev_select)[i].device_number ==
 931                              old_dev_select[i].device_number)
 932                          && ((*dev_select)[i].selected ==
 933                              old_dev_select[i].selected))
 934                                 continue;
 935
 936                         /*
 937                          * Now, if we're still going through the if
 938                          * statement, the above test wasn't true.  So we
 939                          * check here to see if the device at index i in
 940                          * the current array is the same as the device at
 941                          * index i in the old array.  If it is, that means
 942                          * that its selection number has changed.  Set
 943                          * changed to 1 and exit the loop.
 944                          */
 945                         else if ((*dev_select)[i].device_number ==
 946                                   old_dev_select[i].device_number) {
 947                                 changed = 1;
 948                                 break;
 949                         }
 950                         /*
 951                          * If we get here, then the device at index i in
 952                          * the current array isn't the same device as the
 953                          * device at index i in the old array.
 954                          */
 955                         else {
 956                                 found = 0;
 957
 958                                 /*
 959                                  * Search through the old selection array
 960                                  * looking for a device with the same
 961                                  * device number as the device at index i
 962                                  * in the current array.  If the selection
 963                                  * status is the same, then we mark it as
 964                                  * found.  If the selection status isn't
 965                                  * the same, we break out of the loop.
 966                                  * Since found isn't set, changed will be
 967                                  * set to 1 below.
 968                                  */
 969                                 for (j = 0; j < old_num_selections; j++) {
 970                                         if (((*dev_select)[i].device_number ==
 971                                               old_dev_select[j].device_number)
 972                                          && ((*dev_select)[i].selected ==
 973                                               old_dev_select[j].selected)){
 974                                                 found = 1;
 975                                                 break;
 976                                         }
 977                                         else if ((*dev_select)[i].device_number
 978                                             == old_dev_select[j].device_number)
 979                                                 break;
 980                                 }
 981                                 if (found == 0)
 982                                         changed = 1;
 983                         }
 984                 }
 985         }
 986         if (old_dev_select != NULL)
 987                 free(old_dev_select);
 988
 989         return(changed);
 990 }
 991
 992 /*
 993  * Comparison routine for qsort() above.  Note that the comparison here is
 994  * backwards -- generally, it should return a value to indicate whether
 995  * arg1 is <, =, or > arg2.  Instead, it returns the opposite.  The reason
 996  * it returns the opposite is so that the selection array will be sorted in
 997  * order of decreasing performance.  We sort on two parameters.  The first
 998  * sort key is whether or not one or the other of the devices in question
 999  * has been selected.  If one of them has, and the other one has not, the
1000  * selected device is automatically more important than the unselected
1001  * device.  If neither device is selected, we judge the devices based upon
1002  * performance.
1003  */
1004 static int
1005 compare_select(const void *arg1, const void *arg2)
1006 {
1007         if ((((const struct device_selection *)arg1)->selected)
1008          && (((const struct device_selection *)arg2)->selected == 0))
1009                 return(-1);
1010         else if ((((const struct device_selection *)arg1)->selected == 0)
1011               && (((const struct device_selection *)arg2)->selected))
1012                 return(1);
1013         else if (((const struct device_selection *)arg2)->bytes <
1014                  ((const struct device_selection *)arg1)->bytes)
1015                 return(-1);
1016         else if (((const struct device_selection *)arg2)->bytes >
1017                  ((const struct device_selection *)arg1)->bytes)
1018                 return(1);
1019         else
1020                 return(0);
1021 }
1022
1023 /*
1024  * Take a string with the general format "arg1,arg2,arg3", and build a
1025  * device matching expression from it.
1026  */
1027 int
1028 devstat_buildmatch(char *match_str, struct devstat_match **matches,
1029                    int *num_matches)
1030 {
1031         char *tstr[5];
1032         char **tempstr;
1033         int num_args;
1034         int i, j;
1035
1036         /* We can't do much without a string to parse */
1037         if (match_str == NULL) {
1038                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1039                          "%s: no match expression", __func__);
1040                 return(-1);
1041         }
1042
1043         /*
1044          * Break the (comma delimited) input string out into separate strings.
1045          */
1046         for (tempstr = tstr, num_args  = 0;
1047              (*tempstr = strsep(&match_str, ",")) != NULL && (num_args < 5);)
1048                 if (**tempstr != '\0') {
1049                         num_args++;
1050                         if (++tempstr >= &tstr[5])
1051                                 break;
1052                 }
1053
1054         /* The user gave us too many type arguments */
1055         if (num_args > 3) {
1056                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1057                          "%s: too many type arguments", __func__);
1058                 return(-1);
1059         }
1060
1061         if (*num_matches == 0)
1062                 *matches = NULL;
1063
1064         *matches = (struct devstat_match *)reallocf(*matches,
1065                   sizeof(struct devstat_match) * (*num_matches + 1));
1066
1067         if (*matches == NULL) {
1068                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1069                          "%s: Cannot allocate memory for matches list", __func__);
1070                 return(-1);
1071         }
1072
1073         /* Make sure the current entry is clear */
1074         bzero(&matches[0][*num_matches], sizeof(struct devstat_match));
1075
1076         /*
1077          * Step through the arguments the user gave us and build a device
1078          * matching expression from them.
1079          */
1080         for (i = 0; i < num_args; i++) {
1081                 char *tempstr2, *tempstr3;
1082
1083                 /*
1084                  * Get rid of leading white space.
1085                  */
1086                 tempstr2 = tstr[i];
1087                 while (isspace(*tempstr2) && (*tempstr2 != '\0'))
1088                         tempstr2++;
1089
1090                 /*
1091                  * Get rid of trailing white space.
1092                  */
1093                 tempstr3 = &tempstr2[strlen(tempstr2) - 1];
1094
1095                 while ((*tempstr3 != '\0') && (tempstr3 > tempstr2)
1096                     && (isspace(*tempstr3))) {
1097                         *tempstr3 = '\0';
1098                         tempstr3--;
1099                 }
1100
1101                 /*
1102                  * Go through the match table comparing the user's
1103                  * arguments to known device types, interfaces, etc.
1104                  */
1105                 for (j = 0; match_table[j].match_str != NULL; j++) {
1106                         /*
1107                          * We do case-insensitive matching, in case someone
1108                          * wants to enter "SCSI" instead of "scsi" or
1109                          * something like that.  Only compare as many
1110                          * characters as are in the string in the match
1111                          * table.  This should help if someone tries to use
1112                          * a super-long match expression.
1113                          */
1114                         if (strncasecmp(tempstr2, match_table[j].match_str,
1115                             strlen(match_table[j].match_str)) == 0) {
1116                                 /*
1117                                  * Make sure the user hasn't specified two
1118                                  * items of the same type, like "da" and
1119                                  * "cd".  One device cannot be both.
1120                                  */
1121                                 if (((*matches)[*num_matches].match_fields &
1122                                     match_table[j].match_field) != 0) {
1123                                         snprintf(devstat_errbuf,
1124                                                  sizeof(devstat_errbuf),
1125                                                  "%s: cannot have more than "
1126                                                  "one match item in a single "
1127                                                  "category", __func__);
1128                                         return(-1);
1129                                 }
1130                                 /*
1131                                  * If we've gotten this far, we have a
1132                                  * winner.  Set the appropriate fields in
1133                                  * the match entry.
1134                                  */
1135                                 (*matches)[*num_matches].match_fields |=
1136                                         match_table[j].match_field;
1137                                 (*matches)[*num_matches].device_type |=
1138                                         match_table[j].type;
1139                                 (*matches)[*num_matches].num_match_categories++;
1140                                 break;
1141                         }
1142                 }
1143                 /*
1144                  * We should have found a match in the above for loop.  If
1145                  * not, that means the user entered an invalid device type
1146                  * or interface.
1147                  */
1148                 if ((*matches)[*num_matches].num_match_categories != (i + 1)) {
1149                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1150                                  "%s: unknown match item \"%s\"", __func__,
1151                                  tstr[i]);
1152                         return(-1);
1153                 }
1154         }
1155
1156         (*num_matches)++;
1157
1158         return(0);
1159 }
1160
1161 /*
1162  * Compute a number of device statistics.  Only one field is mandatory, and
1163  * that is "current".  Everything else is optional.  The caller passes in
1164  * pointers to variables to hold the various statistics he desires.  If he
1165  * doesn't want a particular staistic, he should pass in a NULL pointer.
1166  * Return values:
1167  * 0   -- success
1168  * -1  -- failure
1169  */
1170 int
1171 compute_stats(struct devstat *current, struct devstat *previous,
1172               long double etime, u_int64_t *total_bytes,
1173               u_int64_t *total_transfers, u_int64_t *total_blocks,
1174               long double *kb_per_transfer, long double *transfers_per_second,
1175               long double *mb_per_second, long double *blocks_per_second,
1176               long double *ms_per_transaction)
1177 {
1178         return(devstat_compute_statistics(current, previous, etime,
1179                total_bytes ? DSM_TOTAL_BYTES : DSM_SKIP,
1180                total_bytes,
1181                total_transfers ? DSM_TOTAL_TRANSFERS : DSM_SKIP,
1182                total_transfers,
1183                total_blocks ? DSM_TOTAL_BLOCKS : DSM_SKIP,
1184                total_blocks,
1185                kb_per_transfer ? DSM_KB_PER_TRANSFER : DSM_SKIP,
1186                kb_per_transfer,
1187                transfers_per_second ? DSM_TRANSFERS_PER_SECOND : DSM_SKIP,
1188                transfers_per_second,
1189                mb_per_second ? DSM_MB_PER_SECOND : DSM_SKIP,
1190                mb_per_second,
1191                blocks_per_second ? DSM_BLOCKS_PER_SECOND : DSM_SKIP,
1192                blocks_per_second,
1193                ms_per_transaction ? DSM_MS_PER_TRANSACTION : DSM_SKIP,
1194                ms_per_transaction,
1195                DSM_NONE));
1196 }
1197
1198
1199 /* This is 1/2^64 */
1200 #define BINTIME_SCALE 5.42101086242752217003726400434970855712890625e-20
1201
1202 long double
1203 devstat_compute_etime(struct bintime *cur_time, struct bintime *prev_time)
1204 {
1205         long double etime;
1206
1207         etime = cur_time->sec;
1208         etime += cur_time->frac * BINTIME_SCALE;
1209         if (prev_time != NULL) {
1210                 etime -= prev_time->sec;
1211                 etime -= prev_time->frac * BINTIME_SCALE;
1212         }
1213         return(etime);
1214 }
1215
1216 #define DELTA(field, index)                             \
1217         (current->field[(index)] - (previous ? previous->field[(index)] : 0))
1218
1219 #define DELTA_T(field)                                  \
1220         devstat_compute_etime(&current->field,          \
1221         (previous ? &previous->field : NULL))
1222
1223 int
1224 devstat_compute_statistics(struct devstat *current, struct devstat *previous,
1225                            long double etime, ...)
1226 {
1227         u_int64_t totalbytes, totalbytesread, totalbyteswrite, totalbytesfree;
1228         u_int64_t totaltransfers, totaltransfersread, totaltransferswrite;
1229         u_int64_t totaltransfersother, totalblocks, totalblocksread;
1230         u_int64_t totalblockswrite, totaltransfersfree, totalblocksfree;
1231         long double totalduration, totaldurationread, totaldurationwrite;
1232         long double totaldurationfree, totaldurationother;
1233         va_list ap;
1234         devstat_metric metric;
1235         u_int64_t *destu64;
1236         long double *destld;
1237         int retval;
1238
1239         retval = 0;
1240
1241         /*
1242          * current is the only mandatory field.
1243          */
1244         if (current == NULL) {
1245                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1246                          "%s: current stats structure was NULL", __func__);
1247                 return(-1);
1248         }
1249
1250         totalbytesread = DELTA(bytes, DEVSTAT_READ);
1251         totalbyteswrite = DELTA(bytes, DEVSTAT_WRITE);
1252         totalbytesfree = DELTA(bytes, DEVSTAT_FREE);
1253         totalbytes = totalbytesread + totalbyteswrite + totalbytesfree;
1254
1255         totaltransfersread = DELTA(operations, DEVSTAT_READ);
1256         totaltransferswrite = DELTA(operations, DEVSTAT_WRITE);
1257         totaltransfersother = DELTA(operations, DEVSTAT_NO_DATA);
1258         totaltransfersfree = DELTA(operations, DEVSTAT_FREE);
1259         totaltransfers = totaltransfersread + totaltransferswrite +
1260                          totaltransfersother + totaltransfersfree;
1261
1262         totalblocks = totalbytes;
1263         totalblocksread = totalbytesread;
1264         totalblockswrite = totalbyteswrite;
1265         totalblocksfree = totalbytesfree;
1266
1267         if (current->block_size > 0) {
1268                 totalblocks /= current->block_size;
1269                 totalblocksread /= current->block_size;
1270                 totalblockswrite /= current->block_size;
1271                 totalblocksfree /= current->block_size;
1272         } else {
1273                 totalblocks /= 512;
1274                 totalblocksread /= 512;
1275                 totalblockswrite /= 512;
1276                 totalblocksfree /= 512;
1277         }
1278
1279         totaldurationread = DELTA_T(duration[DEVSTAT_READ]);
1280         totaldurationwrite = DELTA_T(duration[DEVSTAT_WRITE]);
1281         totaldurationfree = DELTA_T(duration[DEVSTAT_FREE]);
1282         totaldurationother = DELTA_T(duration[DEVSTAT_NO_DATA]);
1283         totalduration = totaldurationread + totaldurationwrite +
1284             totaldurationfree + totaldurationother;
1285
1286         va_start(ap, etime);
1287
1288         while ((metric = (devstat_metric)va_arg(ap, devstat_metric)) != 0) {
1289
1290                 if (metric == DSM_NONE)
1291                         break;
1292
1293                 if (metric >= DSM_MAX) {
1294                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1295                                  "%s: metric %d is out of range", __func__,
1296                                  metric);
1297                         retval = -1;
1298                         goto bailout;
1299                 }
1300
1301                 switch (devstat_arg_list[metric].argtype) {
1302                 case DEVSTAT_ARG_UINT64:
1303                         destu64 = (u_int64_t *)va_arg(ap, u_int64_t *);
1304                         break;
1305                 case DEVSTAT_ARG_LD:
1306                         destld = (long double *)va_arg(ap, long double *);
1307                         break;
1308                 case DEVSTAT_ARG_SKIP:
1309                         destld = (long double *)va_arg(ap, long double *);
1310                         break;
1311                 default:
1312                         retval = -1;
1313                         goto bailout;
1314                         break; /* NOTREACHED */
1315                 }
1316
1317                 if (devstat_arg_list[metric].argtype == DEVSTAT_ARG_SKIP)
1318                         continue;
1319
1320                 switch (metric) {
1321                 case DSM_TOTAL_BYTES:
1322                         *destu64 = totalbytes;
1323                         break;
1324                 case DSM_TOTAL_BYTES_READ:
1325                         *destu64 = totalbytesread;
1326                         break;
1327                 case DSM_TOTAL_BYTES_WRITE:
1328                         *destu64 = totalbyteswrite;
1329                         break;
1330                 case DSM_TOTAL_BYTES_FREE:
1331                         *destu64 = totalbytesfree;
1332                         break;
1333                 case DSM_TOTAL_TRANSFERS:
1334                         *destu64 = totaltransfers;
1335                         break;
1336                 case DSM_TOTAL_TRANSFERS_READ:
1337                         *destu64 = totaltransfersread;
1338                         break;
1339                 case DSM_TOTAL_TRANSFERS_WRITE:
1340                         *destu64 = totaltransferswrite;
1341                         break;
1342                 case DSM_TOTAL_TRANSFERS_FREE:
1343                         *destu64 = totaltransfersfree;
1344                         break;
1345                 case DSM_TOTAL_TRANSFERS_OTHER:
1346                         *destu64 = totaltransfersother;
1347                         break;
1348                 case DSM_TOTAL_BLOCKS:
1349                         *destu64 = totalblocks;
1350                         break;
1351                 case DSM_TOTAL_BLOCKS_READ:
1352                         *destu64 = totalblocksread;
1353                         break;
1354                 case DSM_TOTAL_BLOCKS_WRITE:
1355                         *destu64 = totalblockswrite;
1356                         break;
1357                 case DSM_TOTAL_BLOCKS_FREE:
1358                         *destu64 = totalblocksfree;
1359                         break;
1360                 case DSM_KB_PER_TRANSFER:
1361                         *destld = totalbytes;
1362                         *destld /= 1024;
1363                         if (totaltransfers > 0)
1364                                 *destld /= totaltransfers;
1365                         else
1366                                 *destld = 0.0;
1367                         break;
1368                 case DSM_KB_PER_TRANSFER_READ:
1369                         *destld = totalbytesread;
1370                         *destld /= 1024;
1371                         if (totaltransfersread > 0)
1372                                 *destld /= totaltransfersread;
1373                         else
1374                                 *destld = 0.0;
1375                         break;
1376                 case DSM_KB_PER_TRANSFER_WRITE:
1377                         *destld = totalbyteswrite;
1378                         *destld /= 1024;
1379                         if (totaltransferswrite > 0)
1380                                 *destld /= totaltransferswrite;
1381                         else
1382                                 *destld = 0.0;
1383                         break;
1384                 case DSM_KB_PER_TRANSFER_FREE:
1385                         *destld = totalbytesfree;
1386                         *destld /= 1024;
1387                         if (totaltransfersfree > 0)
1388                                 *destld /= totaltransfersfree;
1389                         else
1390                                 *destld = 0.0;
1391                         break;
1392                 case DSM_TRANSFERS_PER_SECOND:
1393                         if (etime > 0.0) {
1394                                 *destld = totaltransfers;
1395                                 *destld /= etime;
1396                         } else
1397                                 *destld = 0.0;
1398                         break;
1399                 case DSM_TRANSFERS_PER_SECOND_READ:
1400                         if (etime > 0.0) {
1401                                 *destld = totaltransfersread;
1402                                 *destld /= etime;
1403                         } else
1404                                 *destld = 0.0;
1405                         break;
1406                 case DSM_TRANSFERS_PER_SECOND_WRITE:
1407                         if (etime > 0.0) {
1408                                 *destld = totaltransferswrite;
1409                                 *destld /= etime;
1410                         } else
1411                                 *destld = 0.0;
1412                         break;
1413                 case DSM_TRANSFERS_PER_SECOND_FREE:
1414                         if (etime > 0.0) {
1415                                 *destld = totaltransfersfree;
1416                                 *destld /= etime;
1417                         } else
1418                                 *destld = 0.0;
1419                         break;
1420                 case DSM_TRANSFERS_PER_SECOND_OTHER:
1421                         if (etime > 0.0) {
1422                                 *destld = totaltransfersother;
1423                                 *destld /= etime;
1424                         } else
1425                                 *destld = 0.0;
1426                         break;
1427                 case DSM_MB_PER_SECOND:
1428                         *destld = totalbytes;
1429                         *destld /= 1024 * 1024;
1430                         if (etime > 0.0)
1431                                 *destld /= etime;
1432                         else
1433                                 *destld = 0.0;
1434                         break;
1435                 case DSM_MB_PER_SECOND_READ:
1436                         *destld = totalbytesread;
1437                         *destld /= 1024 * 1024;
1438                         if (etime > 0.0)
1439                                 *destld /= etime;
1440                         else
1441                                 *destld = 0.0;
1442                         break;
1443                 case DSM_MB_PER_SECOND_WRITE:
1444                         *destld = totalbyteswrite;
1445                         *destld /= 1024 * 1024;
1446                         if (etime > 0.0)
1447                                 *destld /= etime;
1448                         else
1449                                 *destld = 0.0;
1450                         break;
1451                 case DSM_MB_PER_SECOND_FREE:
1452                         *destld = totalbytesfree;
1453                         *destld /= 1024 * 1024;
1454                         if (etime > 0.0)
1455                                 *destld /= etime;
1456                         else
1457                                 *destld = 0.0;
1458                         break;
1459                 case DSM_BLOCKS_PER_SECOND:
1460                         *destld = totalblocks;
1461                         if (etime > 0.0)
1462                                 *destld /= etime;
1463                         else
1464                                 *destld = 0.0;
1465                         break;
1466                 case DSM_BLOCKS_PER_SECOND_READ:
1467                         *destld = totalblocksread;
1468                         if (etime > 0.0)
1469                                 *destld /= etime;
1470                         else
1471                                 *destld = 0.0;
1472                         break;
1473                 case DSM_BLOCKS_PER_SECOND_WRITE:
1474                         *destld = totalblockswrite;
1475                         if (etime > 0.0)
1476                                 *destld /= etime;
1477                         else
1478                                 *destld = 0.0;
1479                         break;
1480                 case DSM_BLOCKS_PER_SECOND_FREE:
1481                         *destld = totalblocksfree;
1482                         if (etime > 0.0)
1483                                 *destld /= etime;
1484                         else
1485                                 *destld = 0.0;
1486                         break;
1487                 /*
1488                  * Some devstat callers update the duration and some don't.
1489                  * So this will only be accurate if they provide the
1490                  * duration.
1491                  */
1492                 case DSM_MS_PER_TRANSACTION:
1493                         if (totaltransfers > 0) {
1494                                 *destld = totalduration;
1495                                 *destld /= totaltransfers;
1496                                 *destld *= 1000;
1497                         } else
1498                                 *destld = 0.0;
1499                         break;
1500                 case DSM_MS_PER_TRANSACTION_READ:
1501                         if (totaltransfersread > 0) {
1502                                 *destld = totaldurationread;
1503                                 *destld /= totaltransfersread;
1504                                 *destld *= 1000;
1505                         } else
1506                                 *destld = 0.0;
1507                         break;
1508                 case DSM_MS_PER_TRANSACTION_WRITE:
1509                         if (totaltransferswrite > 0) {
1510                                 *destld = totaldurationwrite;
1511                                 *destld /= totaltransferswrite;
1512                                 *destld *= 1000;
1513                         } else
1514                                 *destld = 0.0;
1515                         break;
1516                 case DSM_MS_PER_TRANSACTION_FREE:
1517                         if (totaltransfersfree > 0) {
1518                                 *destld = totaldurationfree;
1519                                 *destld /= totaltransfersfree;
1520                                 *destld *= 1000;
1521                         } else
1522                                 *destld = 0.0;
1523                         break;
1524                 case DSM_MS_PER_TRANSACTION_OTHER:
1525                         if (totaltransfersother > 0) {
1526                                 *destld = totaldurationother;
1527                                 *destld /= totaltransfersother;
1528                                 *destld *= 1000;
1529                         } else
1530                                 *destld = 0.0;
1531                         break;
1532                 case DSM_BUSY_PCT:
1533                         *destld = DELTA_T(busy_time);
1534                         if (*destld < 0)
1535                                 *destld = 0;
1536                         *destld /= etime;
1537                         *destld *= 100;
1538                         if (*destld < 0)
1539                                 *destld = 0;
1540                         break;
1541                 case DSM_QUEUE_LENGTH:
1542                         *destu64 = current->start_count - current->end_count;
1543                         break;
1544                 case DSM_TOTAL_DURATION:
1545                         *destld = totalduration;
1546                         break;
1547                 case DSM_TOTAL_DURATION_READ:
1548                         *destld = totaldurationread;
1549                         break;
1550                 case DSM_TOTAL_DURATION_WRITE:
1551                         *destld = totaldurationwrite;
1552                         break;
1553                 case DSM_TOTAL_DURATION_FREE:
1554                         *destld = totaldurationfree;
1555                         break;
1556                 case DSM_TOTAL_DURATION_OTHER:
1557                         *destld = totaldurationother;
1558                         break;
1559                 case DSM_TOTAL_BUSY_TIME:
1560                         *destld = DELTA_T(busy_time);
1561                         break;
1562 /*
1563  * XXX: comment out the default block to see if any case's are missing.
1564  */
1565 #if 1
1566                 default:
1567                         /*
1568                          * This shouldn't happen, since we should have
1569                          * caught any out of range metrics at the top of
1570                          * the loop.
1571                          */
1572                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1573                                  "%s: unknown metric %d", __func__, metric);
1574                         retval = -1;
1575                         goto bailout;
1576                         break; /* NOTREACHED */
1577 #endif
1578                 }
1579         }
1580
1581 bailout:
1582
1583         va_end(ap);
1584         return(retval);
1585 }
1586
1587 static int
1588 readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes)
1589 {
1590
1591         if (kvm_read(kd, addr, buf, nbytes) == -1) {
1592                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1593                          "%s: error reading value (kvm_read): %s", __func__,
1594                          kvm_geterr(kd));
1595                 return(-1);
1596         }
1597         return(0);
1598 }
1599
1600 static int
1601 readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes)
1602 {
1603         struct nlist nl[2];
1604
1605         nl[0].n_name = (char *)name;
1606         nl[1].n_name = NULL;
1607
1608         if (kvm_nlist(kd, nl) == -1) {
1609                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1610                          "%s: error getting name list (kvm_nlist): %s",
1611                          __func__, kvm_geterr(kd));
1612                 return(-1);
1613         }
1614         return(readkmem(kd, nl[0].n_value, buf, nbytes));
1615 }
1616
1617 /*
1618  * This duplicates the functionality of the kernel sysctl handler for poking
1619  * through crash dumps.
1620  */
1621 static char *
1622 get_devstat_kvm(kvm_t *kd)
1623 {
1624         int i, wp;
1625         long gen;
1626         struct devstat *nds;
1627         struct devstat ds;
1628         struct devstatlist dhead;
1629         int num_devs;
1630         char *rv = NULL;
1631
1632         if ((num_devs = devstat_getnumdevs(kd)) <= 0)
1633                 return(NULL);
1634         if (KREADNL(kd, X_DEVICE_STATQ, dhead) == -1)
1635                 return(NULL);
1636
1637         nds = STAILQ_FIRST(&dhead);
1638
1639         if ((rv = malloc(sizeof(gen))) == NULL) {
1640                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1641                          "%s: out of memory (initial malloc failed)",
1642                          __func__);
1643                 return(NULL);
1644         }
1645         gen = devstat_getgeneration(kd);
1646         memcpy(rv, &gen, sizeof(gen));
1647         wp = sizeof(gen);
1648         /*
1649          * Now push out all the devices.
1650          */
1651         for (i = 0; (nds != NULL) && (i < num_devs);
1652              nds = STAILQ_NEXT(nds, dev_links), i++) {
1653                 if (readkmem(kd, (long)nds, &ds, sizeof(ds)) == -1) {
1654                         free(rv);
1655                         return(NULL);
1656                 }
1657                 nds = &ds;
1658                 rv = (char *)reallocf(rv, sizeof(gen) +
1659                                       sizeof(ds) * (i + 1));
1660                 if (rv == NULL) {
1661                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1662                                  "%s: out of memory (malloc failed)",
1663                                  __func__);
1664                         return(NULL);
1665                 }
1666                 memcpy(rv + wp, &ds, sizeof(ds));
1667                 wp += sizeof(ds);
1668         }
1669         return(rv);
1670 }