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