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_END           4
 154 };
 155
 156 /*
 157  * Local function declarations.
 158  */
 159 static int compare_select(const void *arg1, const void *arg2);
 160 static int readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes);
 161 static int readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes);
 162 static char *get_devstat_kvm(kvm_t *kd);
 163
 164 #define KREADNL(kd, var, val) \
 165         readkmem_nl(kd, namelist[var], &val, sizeof(val))
 166
 167 int
 168 devstat_getnumdevs(kvm_t *kd)
 169 {
 170         size_t numdevsize;
 171         int numdevs;
 172
 173         numdevsize = sizeof(int);
 174
 175         /*
 176          * Find out how many devices we have in the system.
 177          */
 178         if (kd == NULL) {
 179                 if (sysctlbyname("kern.devstat.numdevs", &numdevs,
 180                                  &numdevsize, NULL, 0) == -1) {
 181                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 182                                  "%s: error getting number of devices\n"
 183                                  "%s: %s", __func__, __func__,
 184                                  strerror(errno));
 185                         return(-1);
 186                 } else
 187                         return(numdevs);
 188         } else {
 189
 190                 if (KREADNL(kd, X_NUMDEVS, numdevs) == -1)
 191                         return(-1);
 192                 else
 193                         return(numdevs);
 194         }
 195 }
 196
 197 /*
 198  * This is an easy way to get the generation number, but the generation is
 199  * supplied in a more atmoic manner by the kern.devstat.all sysctl.
 200  * Because this generation sysctl is separate from the statistics sysctl,
 201  * the device list and the generation could change between the time that
 202  * this function is called and the device list is retreived.
 203  */
 204 long
 205 devstat_getgeneration(kvm_t *kd)
 206 {
 207         size_t gensize;
 208         long generation;
 209
 210         gensize = sizeof(long);
 211
 212         /*
 213          * Get the current generation number.
 214          */
 215         if (kd == NULL) {
 216                 if (sysctlbyname("kern.devstat.generation", &generation,
 217                                  &gensize, NULL, 0) == -1) {
 218                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 219                                  "%s: error getting devstat generation\n%s: %s",
 220                                  __func__, __func__, strerror(errno));
 221                         return(-1);
 222                 } else
 223                         return(generation);
 224         } else {
 225                 if (KREADNL(kd, X_GENERATION, generation) == -1)
 226                         return(-1);
 227                 else
 228                         return(generation);
 229         }
 230 }
 231
 232 /*
 233  * Get the current devstat version.  The return value of this function
 234  * should be compared with DEVSTAT_VERSION, which is defined in
 235  * sys/devicestat.h.  This will enable userland programs to determine
 236  * whether they are out of sync with the kernel.
 237  */
 238 int
 239 devstat_getversion(kvm_t *kd)
 240 {
 241         size_t versize;
 242         int version;
 243
 244         versize = sizeof(int);
 245
 246         /*
 247          * Get the current devstat version.
 248          */
 249         if (kd == NULL) {
 250                 if (sysctlbyname("kern.devstat.version", &version, &versize,
 251                                  NULL, 0) == -1) {
 252                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 253                                  "%s: error getting devstat version\n%s: %s",
 254                                  __func__, __func__, strerror(errno));
 255                         return(-1);
 256                 } else
 257                         return(version);
 258         } else {
 259                 if (KREADNL(kd, X_VERSION, version) == -1)
 260                         return(-1);
 261                 else
 262                         return(version);
 263         }
 264 }
 265
 266 /*
 267  * Check the devstat version we know about against the devstat version the
 268  * kernel knows about.  If they don't match, print an error into the
 269  * devstat error buffer, and return -1.  If they match, return 0.
 270  */
 271 int
 272 devstat_checkversion(kvm_t *kd)
 273 {
 274         int buflen, res, retval = 0, version;
 275
 276         version = devstat_getversion(kd);
 277
 278         if (version != DEVSTAT_VERSION) {
 279                 /*
 280                  * If getversion() returns an error (i.e. -1), then it
 281                  * has printed an error message in the buffer.  Therefore,
 282                  * we need to add a \n to the end of that message before we
 283                  * print our own message in the buffer.
 284                  */
 285                 if (version == -1)
 286                         buflen = strlen(devstat_errbuf);
 287                 else
 288                         buflen = 0;
 289
 290                 res = snprintf(devstat_errbuf + buflen,
 291                                DEVSTAT_ERRBUF_SIZE - buflen,
 292                                "%s%s: userland devstat version %d is not "
 293                                "the same as the kernel\n%s: devstat "
 294                                "version %d\n", version == -1 ? "\n" : "",
 295                                __func__, DEVSTAT_VERSION, __func__, version);
 296
 297                 if (res < 0)
 298                         devstat_errbuf[buflen] = '\0';
 299
 300                 buflen = strlen(devstat_errbuf);
 301                 if (version < DEVSTAT_VERSION)
 302                         res = snprintf(devstat_errbuf + buflen,
 303                                        DEVSTAT_ERRBUF_SIZE - buflen,
 304                                        "%s: libdevstat newer than kernel\n",
 305                                        __func__);
 306                 else
 307                         res = snprintf(devstat_errbuf + buflen,
 308                                        DEVSTAT_ERRBUF_SIZE - buflen,
 309                                        "%s: kernel newer than libdevstat\n",
 310                                        __func__);
 311
 312                 if (res < 0)
 313                         devstat_errbuf[buflen] = '\0';
 314
 315                 retval = -1;
 316         }
 317
 318         return(retval);
 319 }
 320
 321 /*
 322  * Get the current list of devices and statistics, and the current
 323  * generation number.
 324  *
 325  * Return values:
 326  * -1  -- error
 327  *  0  -- device list is unchanged
 328  *  1  -- device list has changed
 329  */
 330 int
 331 devstat_getdevs(kvm_t *kd, struct statinfo *stats)
 332 {
 333         int error;
 334         size_t dssize;
 335         int oldnumdevs;
 336         long oldgeneration;
 337         int retval = 0;
 338         struct devinfo *dinfo;
 339         struct timespec ts;
 340
 341         dinfo = stats->dinfo;
 342
 343         if (dinfo == NULL) {
 344                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 345                          "%s: stats->dinfo was NULL", __func__);
 346                 return(-1);
 347         }
 348
 349         oldnumdevs = dinfo->numdevs;
 350         oldgeneration = dinfo->generation;
 351
 352         clock_gettime(CLOCK_MONOTONIC, &ts);
 353         stats->snap_time = ts.tv_sec + ts.tv_nsec * 1e-9;
 354
 355         if (kd == NULL) {
 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                 /*
 437                  * This is of course non-atomic, but since we are working
 438                  * on a core dump, the generation is unlikely to change
 439                  */
 440                 if ((dinfo->numdevs = devstat_getnumdevs(kd)) == -1)
 441                         return(-1);
 442                 if ((dinfo->mem_ptr = (u_int8_t *)get_devstat_kvm(kd)) == NULL)
 443                         return(-1);
 444         }
 445         /*
 446          * The sysctl spits out the generation as the first four bytes,
 447          * then all of the device statistics structures.
 448          */
 449         dinfo->generation = *(long *)dinfo->mem_ptr;
 450
 451         /*
 452          * If the generation has changed, and if the current number of
 453          * devices is not the same as the number of devices recorded in the
 454          * devinfo structure, it is likely that the device list has shrunk.
 455          * The reason that it is likely that the device list has shrunk in
 456          * this case is that if the device list has grown, the sysctl above
 457          * will return an ENOMEM error, and we will reset the number of
 458          * devices and reallocate the device array.  If the second sysctl
 459          * fails, we will return an error and therefore never get to this
 460          * point.  If the device list has shrunk, the sysctl will not
 461          * return an error since we have more space allocated than is
 462          * necessary.  So, in the shrinkage case, we catch it here and
 463          * reallocate the array so that we don't use any more space than is
 464          * necessary.
 465          */
 466         if (oldgeneration != dinfo->generation) {
 467                 if (devstat_getnumdevs(kd) != dinfo->numdevs) {
 468                         if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 469                                 return(-1);
 470                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 471                                 sizeof(long);
 472                         dinfo->mem_ptr = (u_int8_t *)realloc(dinfo->mem_ptr,
 473                                                              dssize);
 474                 }
 475                 retval = 1;
 476         }
 477
 478         dinfo->devices = (struct devstat *)(dinfo->mem_ptr + sizeof(long));
 479
 480         return(retval);
 481 }
 482
 483 /*
 484  * selectdevs():
 485  *
 486  * Devices are selected/deselected based upon the following criteria:
 487  * - devices specified by the user on the command line
 488  * - devices matching any device type expressions given on the command line
 489  * - devices with the highest I/O, if 'top' mode is enabled
 490  * - the first n unselected devices in the device list, if maxshowdevs
 491  *   devices haven't already been selected and if the user has not
 492  *   specified any devices on the command line and if we're in "add" mode.
 493  *
 494  * Input parameters:
 495  * - device selection list (dev_select)
 496  * - current number of devices selected (num_selected)
 497  * - total number of devices in the selection list (num_selections)
 498  * - devstat generation as of the last time selectdevs() was called
 499  *   (select_generation)
 500  * - current devstat generation (current_generation)
 501  * - current list of devices and statistics (devices)
 502  * - number of devices in the current device list (numdevs)
 503  * - compiled version of the command line device type arguments (matches)
 504  *   - This is optional.  If the number of devices is 0, this will be ignored.
 505  *   - The matching code pays attention to the current selection mode.  So
 506  *     if you pass in a matching expression, it will be evaluated based
 507  *     upon the selection mode that is passed in.  See below for details.
 508  * - number of device type matching expressions (num_matches)
 509  *   - Set to 0 to disable the matching code.
 510  * - list of devices specified on the command line by the user (dev_selections)
 511  * - number of devices selected on the command line by the user
 512  *   (num_dev_selections)
 513  * - Our selection mode.  There are four different selection modes:
 514  *      - add mode.  (DS_SELECT_ADD) Any devices matching devices explicitly
 515  *        selected by the user or devices matching a pattern given by the
 516  *        user will be selected in addition to devices that are already
 517  *        selected.  Additional devices will be selected, up to maxshowdevs
 518  *        number of devices.
 519  *      - only mode. (DS_SELECT_ONLY)  Only devices matching devices
 520  *        explicitly given by the user or devices matching a pattern
 521  *        given by the user will be selected.  No other devices will be
 522  *        selected.
 523  *      - addonly mode.  (DS_SELECT_ADDONLY)  This is similar to add and
 524  *        only.  Basically, this will not de-select any devices that are
 525  *        current selected, as only mode would, but it will also not
 526  *        gratuitously select up to maxshowdevs devices as add mode would.
 527  *      - remove mode.  (DS_SELECT_REMOVE)  Any devices matching devices
 528  *        explicitly selected by the user or devices matching a pattern
 529  *        given by the user will be de-selected.
 530  * - maximum number of devices we can select (maxshowdevs)
 531  * - flag indicating whether or not we're in 'top' mode (perf_select)
 532  *
 533  * Output data:
 534  * - the device selection list may be modified and passed back out
 535  * - the number of devices selected and the total number of items in the
 536  *   device selection list may be changed
 537  * - the selection generation may be changed to match the current generation
 538  *
 539  * Return values:
 540  * -1  -- error
 541  *  0  -- selected devices are unchanged
 542  *  1  -- selected devices changed
 543  */
 544 int
 545 devstat_selectdevs(struct device_selection **dev_select, int *num_selected,
 546                    int *num_selections, long *select_generation,
 547                    long current_generation, struct devstat *devices,
 548                    int numdevs, struct devstat_match *matches, int num_matches,
 549                    char **dev_selections, int num_dev_selections,
 550                    devstat_select_mode select_mode, int maxshowdevs,
 551                    int perf_select)
 552 {
 553         int i, j, k;
 554         int init_selections = 0, init_selected_var = 0;
 555         struct device_selection *old_dev_select = NULL;
 556         int old_num_selections = 0, old_num_selected;
 557         int selection_number = 0;
 558         int changed = 0, found = 0;
 559
 560         if ((dev_select == NULL) || (devices == NULL) || (numdevs < 0))
 561                 return(-1);
 562
 563         /*
 564          * We always want to make sure that we have as many dev_select
 565          * entries as there are devices.
 566          */
 567         /*
 568          * In this case, we haven't selected devices before.
 569          */
 570         if (*dev_select == NULL) {
 571                 *dev_select = (struct device_selection *)malloc(numdevs *
 572                         sizeof(struct device_selection));
 573                 *select_generation = current_generation;
 574                 init_selections = 1;
 575                 changed = 1;
 576         /*
 577          * In this case, we have selected devices before, but the device
 578          * list has changed since we last selected devices, so we need to
 579          * either enlarge or reduce the size of the device selection list.
 580          */
 581         } else if (*num_selections != numdevs) {
 582                 *dev_select = (struct device_selection *)reallocf(*dev_select,
 583                         numdevs * sizeof(struct device_selection));
 584                 *select_generation = current_generation;
 585                 init_selections = 1;
 586         /*
 587          * In this case, we've selected devices before, and the selection
 588          * list is the same size as it was the last time, but the device
 589          * list has changed.
 590          */
 591         } else if (*select_generation < current_generation) {
 592                 *select_generation = current_generation;
 593                 init_selections = 1;
 594         }
 595
 596         if (*dev_select == NULL) {
 597                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 598                          "%s: Cannot (re)allocate memory for dev_select argument",
 599                          __func__);
 600                 return(-1);
 601         }
 602
 603         /*
 604          * If we're in "only" mode, we want to clear out the selected
 605          * variable since we're going to select exactly what the user wants
 606          * this time through.
 607          */
 608         if (select_mode == DS_SELECT_ONLY)
 609                 init_selected_var = 1;
 610
 611         /*
 612          * In all cases, we want to back up the number of selected devices.
 613          * It is a quick and accurate way to determine whether the selected
 614          * devices have changed.
 615          */
 616         old_num_selected = *num_selected;
 617
 618         /*
 619          * We want to make a backup of the current selection list if
 620          * the list of devices has changed, or if we're in performance
 621          * selection mode.  In both cases, we don't want to make a backup
 622          * if we already know for sure that the list will be different.
 623          * This is certainly the case if this is our first time through the
 624          * selection code.
 625          */
 626         if (((init_selected_var != 0) || (init_selections != 0)
 627          || (perf_select != 0)) && (changed == 0)){
 628                 old_dev_select = (struct device_selection *)malloc(
 629                     *num_selections * sizeof(struct device_selection));
 630                 if (old_dev_select == NULL) {
 631                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 632                                  "%s: Cannot allocate memory for selection list backup",
 633                                  __func__);
 634                         return(-1);
 635                 }
 636                 old_num_selections = *num_selections;
 637                 bcopy(*dev_select, old_dev_select,
 638                     sizeof(struct device_selection) * *num_selections);
 639         }
 640
 641         if (init_selections != 0) {
 642                 bzero(*dev_select, sizeof(struct device_selection) * numdevs);
 643
 644                 for (i = 0; i < numdevs; i++) {
 645                         (*dev_select)[i].device_number =
 646                                 devices[i].device_number;
 647                         strncpy((*dev_select)[i].device_name,
 648                                 devices[i].device_name,
 649                                 DEVSTAT_NAME_LEN);
 650                         (*dev_select)[i].device_name[DEVSTAT_NAME_LEN - 1]='\0';
 651                         (*dev_select)[i].unit_number = devices[i].unit_number;
 652                         (*dev_select)[i].position = i;
 653                 }
 654                 *num_selections = numdevs;
 655         } else if (init_selected_var != 0) {
 656                 for (i = 0; i < numdevs; i++)
 657                         (*dev_select)[i].selected = 0;
 658         }
 659
 660         /* we haven't gotten around to selecting anything yet.. */
 661         if ((select_mode == DS_SELECT_ONLY) || (init_selections != 0)
 662          || (init_selected_var != 0))
 663                 *num_selected = 0;
 664
 665         /*
 666          * Look through any devices the user specified on the command line
 667          * and see if they match known devices.  If so, select them.
 668          */
 669         for (i = 0; (i < *num_selections) && (num_dev_selections > 0); i++) {
 670                 char tmpstr[80];
 671
 672                 snprintf(tmpstr, sizeof(tmpstr), "%s%d",
 673                          (*dev_select)[i].device_name,
 674                          (*dev_select)[i].unit_number);
 675                 for (j = 0; j < num_dev_selections; j++) {
 676                         if (strcmp(tmpstr, dev_selections[j]) == 0) {
 677                                 /*
 678                                  * Here we do different things based on the
 679                                  * mode we're in.  If we're in add or
 680                                  * addonly mode, we only select this device
 681                                  * if it hasn't already been selected.
 682                                  * Otherwise, we would be unnecessarily
 683                                  * changing the selection order and
 684                                  * incrementing the selection count.  If
 685                                  * we're in only mode, we unconditionally
 686                                  * select this device, since in only mode
 687                                  * any previous selections are erased and
 688                                  * manually specified devices are the first
 689                                  * ones to be selected.  If we're in remove
 690                                  * mode, we de-select the specified device and
 691                                  * decrement the selection count.
 692                                  */
 693                                 switch(select_mode) {
 694                                 case DS_SELECT_ADD:
 695                                 case DS_SELECT_ADDONLY:
 696                                         if ((*dev_select)[i].selected)
 697                                                 break;
 698                                         /* FALLTHROUGH */
 699                                 case DS_SELECT_ONLY:
 700                                         (*dev_select)[i].selected =
 701                                                 ++selection_number;
 702                                         (*num_selected)++;
 703                                         break;
 704                                 case DS_SELECT_REMOVE:
 705                                         (*dev_select)[i].selected = 0;
 706                                         (*num_selected)--;
 707                                         /*
 708                                          * This isn't passed back out, we
 709                                          * just use it to keep track of
 710                                          * how many devices we've removed.
 711                                          */
 712                                         num_dev_selections--;
 713                                         break;
 714                                 }
 715                                 break;
 716                         }
 717                 }
 718         }
 719
 720         /*
 721          * Go through the user's device type expressions and select devices
 722          * accordingly.  We only do this if the number of devices already
 723          * selected is less than the maximum number we can show.
 724          */
 725         for (i = 0; (i < num_matches) && (*num_selected < maxshowdevs); i++) {
 726                 /* We should probably indicate some error here */
 727                 if ((matches[i].match_fields == DEVSTAT_MATCH_NONE)
 728                  || (matches[i].num_match_categories <= 0))
 729                         continue;
 730
 731                 for (j = 0; j < numdevs; j++) {
 732                         int num_match_categories;
 733
 734                         num_match_categories = matches[i].num_match_categories;
 735
 736                         /*
 737                          * Determine whether or not the current device
 738                          * matches the given matching expression.  This if
 739                          * statement consists of three components:
 740                          *   - the device type check
 741                          *   - the device interface check
 742                          *   - the passthrough check
 743                          * If a the matching test is successful, it
 744                          * decrements the number of matching categories,
 745                          * and if we've reached the last element that
 746                          * needed to be matched, the if statement succeeds.
 747                          *
 748                          */
 749                         if ((((matches[i].match_fields & DEVSTAT_MATCH_TYPE)!=0)
 750                           && ((devices[j].device_type & DEVSTAT_TYPE_MASK) ==
 751                                 (matches[i].device_type & DEVSTAT_TYPE_MASK))
 752                           &&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 753                            || (((matches[i].match_fields &
 754                                 DEVSTAT_MATCH_PASS) == 0)
 755                             && ((devices[j].device_type &
 756                                 DEVSTAT_TYPE_PASS) == 0)))
 757                           && (--num_match_categories == 0))
 758                          || (((matches[i].match_fields & DEVSTAT_MATCH_IF) != 0)
 759                           && ((devices[j].device_type & DEVSTAT_TYPE_IF_MASK) ==
 760                                 (matches[i].device_type & DEVSTAT_TYPE_IF_MASK))
 761                           &&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 762                            || (((matches[i].match_fields &
 763                                 DEVSTAT_MATCH_PASS) == 0)
 764                             && ((devices[j].device_type &
 765                                 DEVSTAT_TYPE_PASS) == 0)))
 766                           && (--num_match_categories == 0))
 767                          || (((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 768                           && ((devices[j].device_type & DEVSTAT_TYPE_PASS) != 0)
 769                           && (--num_match_categories == 0))) {
 770
 771                                 /*
 772                                  * This is probably a non-optimal solution
 773                                  * to the problem that the devices in the
 774                                  * device list will not be in the same
 775                                  * order as the devices in the selection
 776                                  * array.
 777                                  */
 778                                 for (k = 0; k < numdevs; k++) {
 779                                         if ((*dev_select)[k].position == j) {
 780                                                 found = 1;
 781                                                 break;
 782                                         }
 783                                 }
 784
 785                                 /*
 786                                  * There shouldn't be a case where a device
 787                                  * in the device list is not in the
 788                                  * selection list...but it could happen.
 789                                  */
 790                                 if (found != 1) {
 791                                         fprintf(stderr, "selectdevs: couldn't"
 792                                                 " find %s%d in selection "
 793                                                 "list\n",
 794                                                 devices[j].device_name,
 795                                                 devices[j].unit_number);
 796                                         break;
 797                                 }
 798
 799                                 /*
 800                                  * We do different things based upon the
 801                                  * mode we're in.  If we're in add or only
 802                                  * mode, we go ahead and select this device
 803                                  * if it hasn't already been selected.  If
 804                                  * it has already been selected, we leave
 805                                  * it alone so we don't mess up the
 806                                  * selection ordering.  Manually specified
 807                                  * devices have already been selected, and
 808                                  * they have higher priority than pattern
 809                                  * matched devices.  If we're in remove
 810                                  * mode, we de-select the given device and
 811                                  * decrement the selected count.
 812                                  */
 813                                 switch(select_mode) {
 814                                 case DS_SELECT_ADD:
 815                                 case DS_SELECT_ADDONLY:
 816                                 case DS_SELECT_ONLY:
 817                                         if ((*dev_select)[k].selected != 0)
 818                                                 break;
 819                                         (*dev_select)[k].selected =
 820                                                 ++selection_number;
 821                                         (*num_selected)++;
 822                                         break;
 823                                 case DS_SELECT_REMOVE:
 824                                         (*dev_select)[k].selected = 0;
 825                                         (*num_selected)--;
 826                                         break;
 827                                 }
 828                         }
 829                 }
 830         }
 831
 832         /*
 833          * Here we implement "top" mode.  Devices are sorted in the
 834          * selection array based on two criteria:  whether or not they are
 835          * selected (not selection number, just the fact that they are
 836          * selected!) and the number of bytes in the "bytes" field of the
 837          * selection structure.  The bytes field generally must be kept up
 838          * by the user.  In the future, it may be maintained by library
 839          * functions, but for now the user has to do the work.
 840          *
 841          * At first glance, it may seem wrong that we don't go through and
 842          * select every device in the case where the user hasn't specified
 843          * any devices or patterns.  In fact, though, it won't make any
 844          * difference in the device sorting.  In that particular case (i.e.
 845          * when we're in "add" or "only" mode, and the user hasn't
 846          * specified anything) the first time through no devices will be
 847          * selected, so the only criterion used to sort them will be their
 848          * performance.  The second time through, and every time thereafter,
 849          * all devices will be selected, so again selection won't matter.
 850          */
 851         if (perf_select != 0) {
 852
 853                 /* Sort the device array by throughput  */
 854                 qsort(*dev_select, *num_selections,
 855                       sizeof(struct device_selection),
 856                       compare_select);
 857
 858                 if (*num_selected == 0) {
 859                         /*
 860                          * Here we select every device in the array, if it
 861                          * isn't already selected.  Because the 'selected'
 862                          * variable in the selection array entries contains
 863                          * the selection order, the devstats routine can show
 864                          * the devices that were selected first.
 865                          */
 866                         for (i = 0; i < *num_selections; i++) {
 867                                 if ((*dev_select)[i].selected == 0) {
 868                                         (*dev_select)[i].selected =
 869                                                 ++selection_number;
 870                                         (*num_selected)++;
 871                                 }
 872                         }
 873                 } else {
 874                         selection_number = 0;
 875                         for (i = 0; i < *num_selections; i++) {
 876                                 if ((*dev_select)[i].selected != 0) {
 877                                         (*dev_select)[i].selected =
 878                                                 ++selection_number;
 879                                 }
 880                         }
 881                 }
 882         }
 883
 884         /*
 885          * If we're in the "add" selection mode and if we haven't already
 886          * selected maxshowdevs number of devices, go through the array and
 887          * select any unselected devices.  If we're in "only" mode, we
 888          * obviously don't want to select anything other than what the user
 889          * specifies.  If we're in "remove" mode, it probably isn't a good
 890          * idea to go through and select any more devices, since we might
 891          * end up selecting something that the user wants removed.  Through
 892          * more complicated logic, we could actually figure this out, but
 893          * that would probably require combining this loop with the various
 894          * selections loops above.
 895          */
 896         if ((select_mode == DS_SELECT_ADD) && (*num_selected < maxshowdevs)) {
 897                 for (i = 0; i < *num_selections; i++)
 898                         if ((*dev_select)[i].selected == 0) {
 899                                 (*dev_select)[i].selected = ++selection_number;
 900                                 (*num_selected)++;
 901                         }
 902         }
 903
 904         /*
 905          * Look at the number of devices that have been selected.  If it
 906          * has changed, set the changed variable.  Otherwise, if we've
 907          * made a backup of the selection list, compare it to the current
 908          * selection list to see if the selected devices have changed.
 909          */
 910         if ((changed == 0) && (old_num_selected != *num_selected))
 911                 changed = 1;
 912         else if ((changed == 0) && (old_dev_select != NULL)) {
 913                 /*
 914                  * Now we go through the selection list and we look at
 915                  * it three different ways.
 916                  */
 917                 for (i = 0; (i < *num_selections) && (changed == 0) &&
 918                      (i < old_num_selections); i++) {
 919                         /*
 920                          * If the device at index i in both the new and old
 921                          * selection arrays has the same device number and
 922                          * selection status, it hasn't changed.  We
 923                          * continue on to the next index.
 924                          */
 925                         if (((*dev_select)[i].device_number ==
 926                              old_dev_select[i].device_number)
 927                          && ((*dev_select)[i].selected ==
 928                              old_dev_select[i].selected))
 929                                 continue;
 930
 931                         /*
 932                          * Now, if we're still going through the if
 933                          * statement, the above test wasn't true.  So we
 934                          * check here to see if the device at index i in
 935                          * the current array is the same as the device at
 936                          * index i in the old array.  If it is, that means
 937                          * that its selection number has changed.  Set
 938                          * changed to 1 and exit the loop.
 939                          */
 940                         else if ((*dev_select)[i].device_number ==
 941                                   old_dev_select[i].device_number) {
 942                                 changed = 1;
 943                                 break;
 944                         }
 945                         /*
 946                          * If we get here, then the device at index i in
 947                          * the current array isn't the same device as the
 948                          * device at index i in the old array.
 949                          */
 950                         else {
 951                                 found = 0;
 952
 953                                 /*
 954                                  * Search through the old selection array
 955                                  * looking for a device with the same
 956                                  * device number as the device at index i
 957                                  * in the current array.  If the selection
 958                                  * status is the same, then we mark it as
 959                                  * found.  If the selection status isn't
 960                                  * the same, we break out of the loop.
 961                                  * Since found isn't set, changed will be
 962                                  * set to 1 below.
 963                                  */
 964                                 for (j = 0; j < old_num_selections; j++) {
 965                                         if (((*dev_select)[i].device_number ==
 966                                               old_dev_select[j].device_number)
 967                                          && ((*dev_select)[i].selected ==
 968                                               old_dev_select[j].selected)){
 969                                                 found = 1;
 970                                                 break;
 971                                         }
 972                                         else if ((*dev_select)[i].device_number
 973                                             == old_dev_select[j].device_number)
 974                                                 break;
 975                                 }
 976                                 if (found == 0)
 977                                         changed = 1;
 978                         }
 979                 }
 980         }
 981         if (old_dev_select != NULL)
 982                 free(old_dev_select);
 983
 984         return(changed);
 985 }
 986
 987 /*
 988  * Comparison routine for qsort() above.  Note that the comparison here is
 989  * backwards -- generally, it should return a value to indicate whether
 990  * arg1 is <, =, or > arg2.  Instead, it returns the opposite.  The reason
 991  * it returns the opposite is so that the selection array will be sorted in
 992  * order of decreasing performance.  We sort on two parameters.  The first
 993  * sort key is whether or not one or the other of the devices in question
 994  * has been selected.  If one of them has, and the other one has not, the
 995  * selected device is automatically more important than the unselected
 996  * device.  If neither device is selected, we judge the devices based upon
 997  * performance.
 998  */
 999 static int
1000 compare_select(const void *arg1, const void *arg2)
1001 {
1002         if ((((const struct device_selection *)arg1)->selected)
1003          && (((const struct device_selection *)arg2)->selected == 0))
1004                 return(-1);
1005         else if ((((const struct device_selection *)arg1)->selected == 0)
1006               && (((const struct device_selection *)arg2)->selected))
1007                 return(1);
1008         else if (((const struct device_selection *)arg2)->bytes <
1009                  ((const struct device_selection *)arg1)->bytes)
1010                 return(-1);
1011         else if (((const struct device_selection *)arg2)->bytes >
1012                  ((const struct device_selection *)arg1)->bytes)
1013                 return(1);
1014         else
1015                 return(0);
1016 }
1017
1018 /*
1019  * Take a string with the general format "arg1,arg2,arg3", and build a
1020  * device matching expression from it.
1021  */
1022 int
1023 devstat_buildmatch(char *match_str, struct devstat_match **matches,
1024                    int *num_matches)
1025 {
1026         char *tstr[5];
1027         char **tempstr;
1028         int num_args;
1029         int i, j;
1030
1031         /* We can't do much without a string to parse */
1032         if (match_str == NULL) {
1033                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1034                          "%s: no match expression", __func__);
1035                 return(-1);
1036         }
1037
1038         /*
1039          * Break the (comma delimited) input string out into separate strings.
1040          */
1041         for (tempstr = tstr, num_args  = 0;
1042              (*tempstr = strsep(&match_str, ",")) != NULL && (num_args < 5);)
1043                 if (**tempstr != '\0') {
1044                         num_args++;
1045                         if (++tempstr >= &tstr[5])
1046                                 break;
1047                 }
1048
1049         /* The user gave us too many type arguments */
1050         if (num_args > 3) {
1051                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1052                          "%s: too many type arguments", __func__);
1053                 return(-1);
1054         }
1055
1056         if (*num_matches == 0)
1057                 *matches = NULL;
1058
1059         *matches = (struct devstat_match *)reallocf(*matches,
1060                   sizeof(struct devstat_match) * (*num_matches + 1));
1061
1062         if (*matches == NULL) {
1063                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1064                          "%s: Cannot allocate memory for matches list", __func__);
1065                 return(-1);
1066         }
1067
1068         /* Make sure the current entry is clear */
1069         bzero(&matches[0][*num_matches], sizeof(struct devstat_match));
1070
1071         /*
1072          * Step through the arguments the user gave us and build a device
1073          * matching expression from them.
1074          */
1075         for (i = 0; i < num_args; i++) {
1076                 char *tempstr2, *tempstr3;
1077
1078                 /*
1079                  * Get rid of leading white space.
1080                  */
1081                 tempstr2 = tstr[i];
1082                 while (isspace(*tempstr2) && (*tempstr2 != '\0'))
1083                         tempstr2++;
1084
1085                 /*
1086                  * Get rid of trailing white space.
1087                  */
1088                 tempstr3 = &tempstr2[strlen(tempstr2) - 1];
1089
1090                 while ((*tempstr3 != '\0') && (tempstr3 > tempstr2)
1091                     && (isspace(*tempstr3))) {
1092                         *tempstr3 = '\0';
1093                         tempstr3--;
1094                 }
1095
1096                 /*
1097                  * Go through the match table comparing the user's
1098                  * arguments to known device types, interfaces, etc.
1099                  */
1100                 for (j = 0; match_table[j].match_str != NULL; j++) {
1101                         /*
1102                          * We do case-insensitive matching, in case someone
1103                          * wants to enter "SCSI" instead of "scsi" or
1104                          * something like that.  Only compare as many
1105                          * characters as are in the string in the match
1106                          * table.  This should help if someone tries to use
1107                          * a super-long match expression.
1108                          */
1109                         if (strncasecmp(tempstr2, match_table[j].match_str,
1110                             strlen(match_table[j].match_str)) == 0) {
1111                                 /*
1112                                  * Make sure the user hasn't specified two
1113                                  * items of the same type, like "da" and
1114                                  * "cd".  One device cannot be both.
1115                                  */
1116                                 if (((*matches)[*num_matches].match_fields &
1117                                     match_table[j].match_field) != 0) {
1118                                         snprintf(devstat_errbuf,
1119                                                  sizeof(devstat_errbuf),
1120                                                  "%s: cannot have more than "
1121                                                  "one match item in a single "
1122                                                  "category", __func__);
1123                                         return(-1);
1124                                 }
1125                                 /*
1126                                  * If we've gotten this far, we have a
1127                                  * winner.  Set the appropriate fields in
1128                                  * the match entry.
1129                                  */
1130                                 (*matches)[*num_matches].match_fields |=
1131                                         match_table[j].match_field;
1132                                 (*matches)[*num_matches].device_type |=
1133                                         match_table[j].type;
1134                                 (*matches)[*num_matches].num_match_categories++;
1135                                 break;
1136                         }
1137                 }
1138                 /*
1139                  * We should have found a match in the above for loop.  If
1140                  * not, that means the user entered an invalid device type
1141                  * or interface.
1142                  */
1143                 if ((*matches)[*num_matches].num_match_categories != (i + 1)) {
1144                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1145                                  "%s: unknown match item \"%s\"", __func__,
1146                                  tstr[i]);
1147                         return(-1);
1148                 }
1149         }
1150
1151         (*num_matches)++;
1152
1153         return(0);
1154 }
1155
1156 /*
1157  * Compute a number of device statistics.  Only one field is mandatory, and
1158  * that is "current".  Everything else is optional.  The caller passes in
1159  * pointers to variables to hold the various statistics he desires.  If he
1160  * doesn't want a particular staistic, he should pass in a NULL pointer.
1161  * Return values:
1162  * 0   -- success
1163  * -1  -- failure
1164  */
1165 int
1166 compute_stats(struct devstat *current, struct devstat *previous,
1167               long double etime, u_int64_t *total_bytes,
1168               u_int64_t *total_transfers, u_int64_t *total_blocks,
1169               long double *kb_per_transfer, long double *transfers_per_second,
1170               long double *mb_per_second, long double *blocks_per_second,
1171               long double *ms_per_transaction)
1172 {
1173         return(devstat_compute_statistics(current, previous, etime,
1174                total_bytes ? DSM_TOTAL_BYTES : DSM_SKIP,
1175                total_bytes,
1176                total_transfers ? DSM_TOTAL_TRANSFERS : DSM_SKIP,
1177                total_transfers,
1178                total_blocks ? DSM_TOTAL_BLOCKS : DSM_SKIP,
1179                total_blocks,
1180                kb_per_transfer ? DSM_KB_PER_TRANSFER : DSM_SKIP,
1181                kb_per_transfer,
1182                transfers_per_second ? DSM_TRANSFERS_PER_SECOND : DSM_SKIP,
1183                transfers_per_second,
1184                mb_per_second ? DSM_MB_PER_SECOND : DSM_SKIP,
1185                mb_per_second,
1186                blocks_per_second ? DSM_BLOCKS_PER_SECOND : DSM_SKIP,
1187                blocks_per_second,
1188                ms_per_transaction ? DSM_MS_PER_TRANSACTION : DSM_SKIP,
1189                ms_per_transaction,
1190                DSM_NONE));
1191 }
1192
1193
1194 /* This is 1/2^64 */
1195 #define BINTIME_SCALE 5.42101086242752217003726400434970855712890625e-20
1196
1197 long double
1198 devstat_compute_etime(struct bintime *cur_time, struct bintime *prev_time)
1199 {
1200         long double etime;
1201
1202         etime = cur_time->sec;
1203         etime += cur_time->frac * BINTIME_SCALE;
1204         if (prev_time != NULL) {
1205                 etime -= prev_time->sec;
1206                 etime -= prev_time->frac * BINTIME_SCALE;
1207         }
1208         return(etime);
1209 }
1210
1211 #define DELTA(field, index)                             \
1212         (current->field[(index)] - (previous ? previous->field[(index)] : 0))
1213
1214 #define DELTA_T(field)                                  \
1215         devstat_compute_etime(&current->field,          \
1216         (previous ? &previous->field : NULL))
1217
1218 int
1219 devstat_compute_statistics(struct devstat *current, struct devstat *previous,
1220                            long double etime, ...)
1221 {
1222         u_int64_t totalbytes, totalbytesread, totalbyteswrite, totalbytesfree;
1223         u_int64_t totaltransfers, totaltransfersread, totaltransferswrite;
1224         u_int64_t totaltransfersother, totalblocks, totalblocksread;
1225         u_int64_t totalblockswrite, totaltransfersfree, totalblocksfree;
1226         long double totalduration, totaldurationread, totaldurationwrite;
1227         long double totaldurationfree, totaldurationother;
1228         va_list ap;
1229         devstat_metric metric;
1230         u_int64_t *destu64;
1231         long double *destld;
1232         int retval;
1233
1234         retval = 0;
1235
1236         /*
1237          * current is the only mandatory field.
1238          */
1239         if (current == NULL) {
1240                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1241                          "%s: current stats structure was NULL", __func__);
1242                 return(-1);
1243         }
1244
1245         totalbytesread = DELTA(bytes, DEVSTAT_READ);
1246         totalbyteswrite = DELTA(bytes, DEVSTAT_WRITE);
1247         totalbytesfree = DELTA(bytes, DEVSTAT_FREE);
1248         totalbytes = totalbytesread + totalbyteswrite + totalbytesfree;
1249
1250         totaltransfersread = DELTA(operations, DEVSTAT_READ);
1251         totaltransferswrite = DELTA(operations, DEVSTAT_WRITE);
1252         totaltransfersother = DELTA(operations, DEVSTAT_NO_DATA);
1253         totaltransfersfree = DELTA(operations, DEVSTAT_FREE);
1254         totaltransfers = totaltransfersread + totaltransferswrite +
1255                          totaltransfersother + totaltransfersfree;
1256
1257         totalblocks = totalbytes;
1258         totalblocksread = totalbytesread;
1259         totalblockswrite = totalbyteswrite;
1260         totalblocksfree = totalbytesfree;
1261
1262         if (current->block_size > 0) {
1263                 totalblocks /= current->block_size;
1264                 totalblocksread /= current->block_size;
1265                 totalblockswrite /= current->block_size;
1266                 totalblocksfree /= current->block_size;
1267         } else {
1268                 totalblocks /= 512;
1269                 totalblocksread /= 512;
1270                 totalblockswrite /= 512;
1271                 totalblocksfree /= 512;
1272         }
1273
1274         totaldurationread = DELTA_T(duration[DEVSTAT_READ]);
1275         totaldurationwrite = DELTA_T(duration[DEVSTAT_WRITE]);
1276         totaldurationfree = DELTA_T(duration[DEVSTAT_FREE]);
1277         totaldurationother = DELTA_T(duration[DEVSTAT_NO_DATA]);
1278         totalduration = totaldurationread + totaldurationwrite +
1279             totaldurationfree + totaldurationother;
1280
1281         va_start(ap, etime);
1282
1283         while ((metric = (devstat_metric)va_arg(ap, devstat_metric)) != 0) {
1284
1285                 if (metric == DSM_NONE)
1286                         break;
1287
1288                 if (metric >= DSM_MAX) {
1289                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1290                                  "%s: metric %d is out of range", __func__,
1291                                  metric);
1292                         retval = -1;
1293                         goto bailout;
1294                 }
1295
1296                 switch (devstat_arg_list[metric].argtype) {
1297                 case DEVSTAT_ARG_UINT64:
1298                         destu64 = (u_int64_t *)va_arg(ap, u_int64_t *);
1299                         break;
1300                 case DEVSTAT_ARG_LD:
1301                         destld = (long double *)va_arg(ap, long double *);
1302                         break;
1303                 case DEVSTAT_ARG_SKIP:
1304                         destld = (long double *)va_arg(ap, long double *);
1305                         break;
1306                 default:
1307                         retval = -1;
1308                         goto bailout;
1309                         break; /* NOTREACHED */
1310                 }
1311
1312                 if (devstat_arg_list[metric].argtype == DEVSTAT_ARG_SKIP)
1313                         continue;
1314
1315                 switch (metric) {
1316                 case DSM_TOTAL_BYTES:
1317                         *destu64 = totalbytes;
1318                         break;
1319                 case DSM_TOTAL_BYTES_READ:
1320                         *destu64 = totalbytesread;
1321                         break;
1322                 case DSM_TOTAL_BYTES_WRITE:
1323                         *destu64 = totalbyteswrite;
1324                         break;
1325                 case DSM_TOTAL_BYTES_FREE:
1326                         *destu64 = totalbytesfree;
1327                         break;
1328                 case DSM_TOTAL_TRANSFERS:
1329                         *destu64 = totaltransfers;
1330                         break;
1331                 case DSM_TOTAL_TRANSFERS_READ:
1332                         *destu64 = totaltransfersread;
1333                         break;
1334                 case DSM_TOTAL_TRANSFERS_WRITE:
1335                         *destu64 = totaltransferswrite;
1336                         break;
1337                 case DSM_TOTAL_TRANSFERS_FREE:
1338                         *destu64 = totaltransfersfree;
1339                         break;
1340                 case DSM_TOTAL_TRANSFERS_OTHER:
1341                         *destu64 = totaltransfersother;
1342                         break;
1343                 case DSM_TOTAL_BLOCKS:
1344                         *destu64 = totalblocks;
1345                         break;
1346                 case DSM_TOTAL_BLOCKS_READ:
1347                         *destu64 = totalblocksread;
1348                         break;
1349                 case DSM_TOTAL_BLOCKS_WRITE:
1350                         *destu64 = totalblockswrite;
1351                         break;
1352                 case DSM_TOTAL_BLOCKS_FREE:
1353                         *destu64 = totalblocksfree;
1354                         break;
1355                 case DSM_KB_PER_TRANSFER:
1356                         *destld = totalbytes;
1357                         *destld /= 1024;
1358                         if (totaltransfers > 0)
1359                                 *destld /= totaltransfers;
1360                         else
1361                                 *destld = 0.0;
1362                         break;
1363                 case DSM_KB_PER_TRANSFER_READ:
1364                         *destld = totalbytesread;
1365                         *destld /= 1024;
1366                         if (totaltransfersread > 0)
1367                                 *destld /= totaltransfersread;
1368                         else
1369                                 *destld = 0.0;
1370                         break;
1371                 case DSM_KB_PER_TRANSFER_WRITE:
1372                         *destld = totalbyteswrite;
1373                         *destld /= 1024;
1374                         if (totaltransferswrite > 0)
1375                                 *destld /= totaltransferswrite;
1376                         else
1377                                 *destld = 0.0;
1378                         break;
1379                 case DSM_KB_PER_TRANSFER_FREE:
1380                         *destld = totalbytesfree;
1381                         *destld /= 1024;
1382                         if (totaltransfersfree > 0)
1383                                 *destld /= totaltransfersfree;
1384                         else
1385                                 *destld = 0.0;
1386                         break;
1387                 case DSM_TRANSFERS_PER_SECOND:
1388                         if (etime > 0.0) {
1389                                 *destld = totaltransfers;
1390                                 *destld /= etime;
1391                         } else
1392                                 *destld = 0.0;
1393                         break;
1394                 case DSM_TRANSFERS_PER_SECOND_READ:
1395                         if (etime > 0.0) {
1396                                 *destld = totaltransfersread;
1397                                 *destld /= etime;
1398                         } else
1399                                 *destld = 0.0;
1400                         break;
1401                 case DSM_TRANSFERS_PER_SECOND_WRITE:
1402                         if (etime > 0.0) {
1403                                 *destld = totaltransferswrite;
1404                                 *destld /= etime;
1405                         } else
1406                                 *destld = 0.0;
1407                         break;
1408                 case DSM_TRANSFERS_PER_SECOND_FREE:
1409                         if (etime > 0.0) {
1410                                 *destld = totaltransfersfree;
1411                                 *destld /= etime;
1412                         } else
1413                                 *destld = 0.0;
1414                         break;
1415                 case DSM_TRANSFERS_PER_SECOND_OTHER:
1416                         if (etime > 0.0) {
1417                                 *destld = totaltransfersother;
1418                                 *destld /= etime;
1419                         } else
1420                                 *destld = 0.0;
1421                         break;
1422                 case DSM_MB_PER_SECOND:
1423                         *destld = totalbytes;
1424                         *destld /= 1024 * 1024;
1425                         if (etime > 0.0)
1426                                 *destld /= etime;
1427                         else
1428                                 *destld = 0.0;
1429                         break;
1430                 case DSM_MB_PER_SECOND_READ:
1431                         *destld = totalbytesread;
1432                         *destld /= 1024 * 1024;
1433                         if (etime > 0.0)
1434                                 *destld /= etime;
1435                         else
1436                                 *destld = 0.0;
1437                         break;
1438                 case DSM_MB_PER_SECOND_WRITE:
1439                         *destld = totalbyteswrite;
1440                         *destld /= 1024 * 1024;
1441                         if (etime > 0.0)
1442                                 *destld /= etime;
1443                         else
1444                                 *destld = 0.0;
1445                         break;
1446                 case DSM_MB_PER_SECOND_FREE:
1447                         *destld = totalbytesfree;
1448                         *destld /= 1024 * 1024;
1449                         if (etime > 0.0)
1450                                 *destld /= etime;
1451                         else
1452                                 *destld = 0.0;
1453                         break;
1454                 case DSM_BLOCKS_PER_SECOND:
1455                         *destld = totalblocks;
1456                         if (etime > 0.0)
1457                                 *destld /= etime;
1458                         else
1459                                 *destld = 0.0;
1460                         break;
1461                 case DSM_BLOCKS_PER_SECOND_READ:
1462                         *destld = totalblocksread;
1463                         if (etime > 0.0)
1464                                 *destld /= etime;
1465                         else
1466                                 *destld = 0.0;
1467                         break;
1468                 case DSM_BLOCKS_PER_SECOND_WRITE:
1469                         *destld = totalblockswrite;
1470                         if (etime > 0.0)
1471                                 *destld /= etime;
1472                         else
1473                                 *destld = 0.0;
1474                         break;
1475                 case DSM_BLOCKS_PER_SECOND_FREE:
1476                         *destld = totalblocksfree;
1477                         if (etime > 0.0)
1478                                 *destld /= etime;
1479                         else
1480                                 *destld = 0.0;
1481                         break;
1482                 /*
1483                  * This calculation is somewhat bogus.  It simply divides
1484                  * the elapsed time by the total number of transactions
1485                  * completed.  While that does give the caller a good
1486                  * picture of the average rate of transaction completion,
1487                  * it doesn't necessarily give the caller a good view of
1488                  * how long transactions took to complete on average.
1489                  * Those two numbers will be different for a device that
1490                  * can handle more than one transaction at a time.  e.g.
1491                  * SCSI disks doing tagged queueing.
1492                  *
1493                  * The only way to accurately determine the real average
1494                  * time per transaction would be to compute and store the
1495                  * time on a per-transaction basis.  That currently isn't
1496                  * done in the kernel, and would only be desireable if it
1497                  * could be implemented in a somewhat non-intrusive and high
1498                  * performance way.
1499                  */
1500                 case DSM_MS_PER_TRANSACTION:
1501                         if (totaltransfers > 0) {
1502                                 *destld = totalduration;
1503                                 *destld /= totaltransfers;
1504                                 *destld *= 1000;
1505                         } else
1506                                 *destld = 0.0;
1507                         break;
1508                 /*
1509                  * As above, these next two really only give the average
1510                  * rate of completion for read and write transactions, not
1511                  * the average time the transaction took to complete.
1512                  */
1513                 case DSM_MS_PER_TRANSACTION_READ:
1514                         if (totaltransfersread > 0) {
1515                                 *destld = totaldurationread;
1516                                 *destld /= totaltransfersread;
1517                                 *destld *= 1000;
1518                         } else
1519                                 *destld = 0.0;
1520                         break;
1521                 case DSM_MS_PER_TRANSACTION_WRITE:
1522                         if (totaltransferswrite > 0) {
1523                                 *destld = totaldurationwrite;
1524                                 *destld /= totaltransferswrite;
1525                                 *destld *= 1000;
1526                         } else
1527                                 *destld = 0.0;
1528                         break;
1529                 case DSM_MS_PER_TRANSACTION_FREE:
1530                         if (totaltransfersfree > 0) {
1531                                 *destld = totaldurationfree;
1532                                 *destld /= totaltransfersfree;
1533                                 *destld *= 1000;
1534                         } else
1535                                 *destld = 0.0;
1536                         break;
1537                 case DSM_MS_PER_TRANSACTION_OTHER:
1538                         if (totaltransfersother > 0) {
1539                                 *destld = totaldurationother;
1540                                 *destld /= totaltransfersother;
1541                                 *destld *= 1000;
1542                         } else
1543                                 *destld = 0.0;
1544                         break;
1545                 case DSM_BUSY_PCT:
1546                         *destld = DELTA_T(busy_time);
1547                         if (*destld < 0)
1548                                 *destld = 0;
1549                         *destld /= etime;
1550                         *destld *= 100;
1551                         if (*destld < 0)
1552                                 *destld = 0;
1553                         break;
1554                 case DSM_QUEUE_LENGTH:
1555                         *destu64 = current->start_count - current->end_count;
1556                         break;
1557                 case DSM_TOTAL_DURATION:
1558                         *destld = totalduration;
1559                         break;
1560                 case DSM_TOTAL_DURATION_READ:
1561                         *destld = totaldurationread;
1562                         break;
1563                 case DSM_TOTAL_DURATION_WRITE:
1564                         *destld = totaldurationwrite;
1565                         break;
1566                 case DSM_TOTAL_DURATION_FREE:
1567                         *destld = totaldurationfree;
1568                         break;
1569                 case DSM_TOTAL_DURATION_OTHER:
1570                         *destld = totaldurationother;
1571                         break;
1572                 case DSM_TOTAL_BUSY_TIME:
1573                         *destld = DELTA_T(busy_time);
1574                         break;
1575 /*
1576  * XXX: comment out the default block to see if any case's are missing.
1577  */
1578 #if 1
1579                 default:
1580                         /*
1581                          * This shouldn't happen, since we should have
1582                          * caught any out of range metrics at the top of
1583                          * the loop.
1584                          */
1585                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1586                                  "%s: unknown metric %d", __func__, metric);
1587                         retval = -1;
1588                         goto bailout;
1589                         break; /* NOTREACHED */
1590 #endif
1591                 }
1592         }
1593
1594 bailout:
1595
1596         va_end(ap);
1597         return(retval);
1598 }
1599
1600 static int
1601 readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes)
1602 {
1603
1604         if (kvm_read(kd, addr, buf, nbytes) == -1) {
1605                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1606                          "%s: error reading value (kvm_read): %s", __func__,
1607                          kvm_geterr(kd));
1608                 return(-1);
1609         }
1610         return(0);
1611 }
1612
1613 static int
1614 readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes)
1615 {
1616         struct nlist nl[2];
1617
1618         nl[0].n_name = (char *)name;
1619         nl[1].n_name = NULL;
1620
1621         if (kvm_nlist(kd, nl) == -1) {
1622                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1623                          "%s: error getting name list (kvm_nlist): %s",
1624                          __func__, kvm_geterr(kd));
1625                 return(-1);
1626         }
1627         return(readkmem(kd, nl[0].n_value, buf, nbytes));
1628 }
1629
1630 /*
1631  * This duplicates the functionality of the kernel sysctl handler for poking
1632  * through crash dumps.
1633  */
1634 static char *
1635 get_devstat_kvm(kvm_t *kd)
1636 {
1637         int i, wp;
1638         long gen;
1639         struct devstat *nds;
1640         struct devstat ds;
1641         struct devstatlist dhead;
1642         int num_devs;
1643         char *rv = NULL;
1644
1645         if ((num_devs = devstat_getnumdevs(kd)) <= 0)
1646                 return(NULL);
1647         if (KREADNL(kd, X_DEVICE_STATQ, dhead) == -1)
1648                 return(NULL);
1649
1650         nds = STAILQ_FIRST(&dhead);
1651
1652         if ((rv = malloc(sizeof(gen))) == NULL) {
1653                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1654                          "%s: out of memory (initial malloc failed)",
1655                          __func__);
1656                 return(NULL);
1657         }
1658         gen = devstat_getgeneration(kd);
1659         memcpy(rv, &gen, sizeof(gen));
1660         wp = sizeof(gen);
1661         /*
1662          * Now push out all the devices.
1663          */
1664         for (i = 0; (nds != NULL) && (i < num_devs);
1665              nds = STAILQ_NEXT(nds, dev_links), i++) {
1666                 if (readkmem(kd, (long)nds, &ds, sizeof(ds)) == -1) {
1667                         free(rv);
1668                         return(NULL);
1669                 }
1670                 nds = &ds;
1671                 rv = (char *)reallocf(rv, sizeof(gen) +
1672                                       sizeof(ds) * (i + 1));
1673                 if (rv == NULL) {
1674                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1675                                  "%s: out of memory (malloc failed)",
1676                                  __func__);
1677                         return(NULL);
1678                 }
1679                 memcpy(rv + wp, &ds, sizeof(ds));
1680                 wp += sizeof(ds);
1681         }
1682         return(rv);
1683 }