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                 } else
 369                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 370                                  sizeof(long);
 371
 372                 /*
 373                  * Request all of the devices.  We only really allow for one
 374                  * ENOMEM failure.  It would, of course, be possible to just go
 375                  * in a loop and keep reallocing the device structure until we
 376                  * don't get ENOMEM back.  I'm not sure it's worth it, though.
 377                  * If devices are being added to the system that quickly, maybe
 378                  * the user can just wait until all devices are added.
 379                  */
 380                 for (;;) {
 381                         error = sysctlbyname("kern.devstat.all",
 382                                              dinfo->mem_ptr,
 383                                              &dssize, NULL, 0);
 384                         if (error != -1 || errno != EBUSY)
 385                                 break;
 386                 }
 387                 if (error == -1) {
 388                         /*
 389                          * If we get ENOMEM back, that means that there are
 390                          * more devices now, so we need to allocate more
 391                          * space for the device array.
 392                          */
 393                         if (errno == ENOMEM) {
 394                                 /*
 395                                  * No need to set the error string here,
 396                                  * devstat_getnumdevs() will do that if it fails.
 397                                  */
 398                                 if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 399                                         return(-1);
 400
 401                                 dssize = (dinfo->numdevs *
 402                                         sizeof(struct devstat)) + sizeof(long);
 403                                 dinfo->mem_ptr = (u_int8_t *)
 404                                         realloc(dinfo->mem_ptr, dssize);
 405                                 if ((error = sysctlbyname("kern.devstat.all",
 406                                     dinfo->mem_ptr, &dssize, NULL, 0)) == -1) {
 407                                         snprintf(devstat_errbuf,
 408                                                  sizeof(devstat_errbuf),
 409                                                  "%s: error getting device "
 410                                                  "stats\n%s: %s", __func__,
 411                                                  __func__, strerror(errno));
 412                                         return(-1);
 413                                 }
 414                         } else {
 415                                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 416                                          "%s: error getting device stats\n"
 417                                          "%s: %s", __func__, __func__,
 418                                          strerror(errno));
 419                                 return(-1);
 420                         }
 421                 }
 422
 423         } else {
 424                 /*
 425                  * This is of course non-atomic, but since we are working
 426                  * on a core dump, the generation is unlikely to change
 427                  */
 428                 if ((dinfo->numdevs = devstat_getnumdevs(kd)) == -1)
 429                         return(-1);
 430                 if ((dinfo->mem_ptr = (u_int8_t *)get_devstat_kvm(kd)) == NULL)
 431                         return(-1);
 432         }
 433         /*
 434          * The sysctl spits out the generation as the first four bytes,
 435          * then all of the device statistics structures.
 436          */
 437         dinfo->generation = *(long *)dinfo->mem_ptr;
 438
 439         /*
 440          * If the generation has changed, and if the current number of
 441          * devices is not the same as the number of devices recorded in the
 442          * devinfo structure, it is likely that the device list has shrunk.
 443          * The reason that it is likely that the device list has shrunk in
 444          * this case is that if the device list has grown, the sysctl above
 445          * will return an ENOMEM error, and we will reset the number of
 446          * devices and reallocate the device array.  If the second sysctl
 447          * fails, we will return an error and therefore never get to this
 448          * point.  If the device list has shrunk, the sysctl will not
 449          * return an error since we have more space allocated than is
 450          * necessary.  So, in the shrinkage case, we catch it here and
 451          * reallocate the array so that we don't use any more space than is
 452          * necessary.
 453          */
 454         if (oldgeneration != dinfo->generation) {
 455                 if (devstat_getnumdevs(kd) != dinfo->numdevs) {
 456                         if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 457                                 return(-1);
 458                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 459                                 sizeof(long);
 460                         dinfo->mem_ptr = (u_int8_t *)realloc(dinfo->mem_ptr,
 461                                                              dssize);
 462                 }
 463                 retval = 1;
 464         }
 465
 466         dinfo->devices = (struct devstat *)(dinfo->mem_ptr + sizeof(long));
 467
 468         return(retval);
 469 }
 470
 471 /*
 472  * selectdevs():
 473  *
 474  * Devices are selected/deselected based upon the following criteria:
 475  * - devices specified by the user on the command line
 476  * - devices matching any device type expressions given on the command line
 477  * - devices with the highest I/O, if 'top' mode is enabled
 478  * - the first n unselected devices in the device list, if maxshowdevs
 479  *   devices haven't already been selected and if the user has not
 480  *   specified any devices on the command line and if we're in "add" mode.
 481  *
 482  * Input parameters:
 483  * - device selection list (dev_select)
 484  * - current number of devices selected (num_selected)
 485  * - total number of devices in the selection list (num_selections)
 486  * - devstat generation as of the last time selectdevs() was called
 487  *   (select_generation)
 488  * - current devstat generation (current_generation)
 489  * - current list of devices and statistics (devices)
 490  * - number of devices in the current device list (numdevs)
 491  * - compiled version of the command line device type arguments (matches)
 492  *   - This is optional.  If the number of devices is 0, this will be ignored.
 493  *   - The matching code pays attention to the current selection mode.  So
 494  *     if you pass in a matching expression, it will be evaluated based
 495  *     upon the selection mode that is passed in.  See below for details.
 496  * - number of device type matching expressions (num_matches)
 497  *   - Set to 0 to disable the matching code.
 498  * - list of devices specified on the command line by the user (dev_selections)
 499  * - number of devices selected on the command line by the user
 500  *   (num_dev_selections)
 501  * - Our selection mode.  There are four different selection modes:
 502  *      - add mode.  (DS_SELECT_ADD) Any devices matching devices explicitly
 503  *        selected by the user or devices matching a pattern given by the
 504  *        user will be selected in addition to devices that are already
 505  *        selected.  Additional devices will be selected, up to maxshowdevs
 506  *        number of devices.
 507  *      - only mode. (DS_SELECT_ONLY)  Only devices matching devices
 508  *        explicitly given by the user or devices matching a pattern
 509  *        given by the user will be selected.  No other devices will be
 510  *        selected.
 511  *      - addonly mode.  (DS_SELECT_ADDONLY)  This is similar to add and
 512  *        only.  Basically, this will not de-select any devices that are
 513  *        current selected, as only mode would, but it will also not
 514  *        gratuitously select up to maxshowdevs devices as add mode would.
 515  *      - remove mode.  (DS_SELECT_REMOVE)  Any devices matching devices
 516  *        explicitly selected by the user or devices matching a pattern
 517  *        given by the user will be de-selected.
 518  * - maximum number of devices we can select (maxshowdevs)
 519  * - flag indicating whether or not we're in 'top' mode (perf_select)
 520  *
 521  * Output data:
 522  * - the device selection list may be modified and passed back out
 523  * - the number of devices selected and the total number of items in the
 524  *   device selection list may be changed
 525  * - the selection generation may be changed to match the current generation
 526  *
 527  * Return values:
 528  * -1  -- error
 529  *  0  -- selected devices are unchanged
 530  *  1  -- selected devices changed
 531  */
 532 int
 533 devstat_selectdevs(struct device_selection **dev_select, int *num_selected,
 534                    int *num_selections, long *select_generation,
 535                    long current_generation, struct devstat *devices,
 536                    int numdevs, struct devstat_match *matches, int num_matches,
 537                    char **dev_selections, int num_dev_selections,
 538                    devstat_select_mode select_mode, int maxshowdevs,
 539                    int perf_select)
 540 {
 541         int i, j, k;
 542         int init_selections = 0, init_selected_var = 0;
 543         struct device_selection *old_dev_select = NULL;
 544         int old_num_selections = 0, old_num_selected;
 545         int selection_number = 0;
 546         int changed = 0, found = 0;
 547
 548         if ((dev_select == NULL) || (devices == NULL) || (numdevs < 0))
 549                 return(-1);
 550
 551         /*
 552          * We always want to make sure that we have as many dev_select
 553          * entries as there are devices.
 554          */
 555         /*
 556          * In this case, we haven't selected devices before.
 557          */
 558         if (*dev_select == NULL) {
 559                 *dev_select = (struct device_selection *)malloc(numdevs *
 560                         sizeof(struct device_selection));
 561                 *select_generation = current_generation;
 562                 init_selections = 1;
 563                 changed = 1;
 564         /*
 565          * In this case, we have selected devices before, but the device
 566          * list has changed since we last selected devices, so we need to
 567          * either enlarge or reduce the size of the device selection list.
 568          */
 569         } else if (*num_selections != numdevs) {
 570                 *dev_select = (struct device_selection *)realloc(*dev_select,
 571                         numdevs * sizeof(struct device_selection));
 572                 *select_generation = current_generation;
 573                 init_selections = 1;
 574         /*
 575          * In this case, we've selected devices before, and the selection
 576          * list is the same size as it was the last time, but the device
 577          * list has changed.
 578          */
 579         } else if (*select_generation < current_generation) {
 580                 *select_generation = current_generation;
 581                 init_selections = 1;
 582         }
 583
 584         /*
 585          * If we're in "only" mode, we want to clear out the selected
 586          * variable since we're going to select exactly what the user wants
 587          * this time through.
 588          */
 589         if (select_mode == DS_SELECT_ONLY)
 590                 init_selected_var = 1;
 591
 592         /*
 593          * In all cases, we want to back up the number of selected devices.
 594          * It is a quick and accurate way to determine whether the selected
 595          * devices have changed.
 596          */
 597         old_num_selected = *num_selected;
 598
 599         /*
 600          * We want to make a backup of the current selection list if
 601          * the list of devices has changed, or if we're in performance
 602          * selection mode.  In both cases, we don't want to make a backup
 603          * if we already know for sure that the list will be different.
 604          * This is certainly the case if this is our first time through the
 605          * selection code.
 606          */
 607         if (((init_selected_var != 0) || (init_selections != 0)
 608          || (perf_select != 0)) && (changed == 0)){
 609                 old_dev_select = (struct device_selection *)malloc(
 610                     *num_selections * sizeof(struct device_selection));
 611                 old_num_selections = *num_selections;
 612                 bcopy(*dev_select, old_dev_select,
 613                     sizeof(struct device_selection) * *num_selections);
 614         }
 615
 616         if (init_selections != 0) {
 617                 bzero(*dev_select, sizeof(struct device_selection) * numdevs);
 618
 619                 for (i = 0; i < numdevs; i++) {
 620                         (*dev_select)[i].device_number =
 621                                 devices[i].device_number;
 622                         strncpy((*dev_select)[i].device_name,
 623                                 devices[i].device_name,
 624                                 DEVSTAT_NAME_LEN);
 625                         (*dev_select)[i].device_name[DEVSTAT_NAME_LEN - 1]='\0';
 626                         (*dev_select)[i].unit_number = devices[i].unit_number;
 627                         (*dev_select)[i].position = i;
 628                 }
 629                 *num_selections = numdevs;
 630         } else if (init_selected_var != 0) {
 631                 for (i = 0; i < numdevs; i++)
 632                         (*dev_select)[i].selected = 0;
 633         }
 634
 635         /* we haven't gotten around to selecting anything yet.. */
 636         if ((select_mode == DS_SELECT_ONLY) || (init_selections != 0)
 637          || (init_selected_var != 0))
 638                 *num_selected = 0;
 639
 640         /*
 641          * Look through any devices the user specified on the command line
 642          * and see if they match known devices.  If so, select them.
 643          */
 644         for (i = 0; (i < *num_selections) && (num_dev_selections > 0); i++) {
 645                 char tmpstr[80];
 646
 647                 snprintf(tmpstr, sizeof(tmpstr), "%s%d",
 648                          (*dev_select)[i].device_name,
 649                          (*dev_select)[i].unit_number);
 650                 for (j = 0; j < num_dev_selections; j++) {
 651                         if (strcmp(tmpstr, dev_selections[j]) == 0) {
 652                                 /*
 653                                  * Here we do different things based on the
 654                                  * mode we're in.  If we're in add or
 655                                  * addonly mode, we only select this device
 656                                  * if it hasn't already been selected.
 657                                  * Otherwise, we would be unnecessarily
 658                                  * changing the selection order and
 659                                  * incrementing the selection count.  If
 660                                  * we're in only mode, we unconditionally
 661                                  * select this device, since in only mode
 662                                  * any previous selections are erased and
 663                                  * manually specified devices are the first
 664                                  * ones to be selected.  If we're in remove
 665                                  * mode, we de-select the specified device and
 666                                  * decrement the selection count.
 667                                  */
 668                                 switch(select_mode) {
 669                                 case DS_SELECT_ADD:
 670                                 case DS_SELECT_ADDONLY:
 671                                         if ((*dev_select)[i].selected)
 672                                                 break;
 673                                         /* FALLTHROUGH */
 674                                 case DS_SELECT_ONLY:
 675                                         (*dev_select)[i].selected =
 676                                                 ++selection_number;
 677                                         (*num_selected)++;
 678                                         break;
 679                                 case DS_SELECT_REMOVE:
 680                                         (*dev_select)[i].selected = 0;
 681                                         (*num_selected)--;
 682                                         /*
 683                                          * This isn't passed back out, we
 684                                          * just use it to keep track of
 685                                          * how many devices we've removed.
 686                                          */
 687                                         num_dev_selections--;
 688                                         break;
 689                                 }
 690                                 break;
 691                         }
 692                 }
 693         }
 694
 695         /*
 696          * Go through the user's device type expressions and select devices
 697          * accordingly.  We only do this if the number of devices already
 698          * selected is less than the maximum number we can show.
 699          */
 700         for (i = 0; (i < num_matches) && (*num_selected < maxshowdevs); i++) {
 701                 /* We should probably indicate some error here */
 702                 if ((matches[i].match_fields == DEVSTAT_MATCH_NONE)
 703                  || (matches[i].num_match_categories <= 0))
 704                         continue;
 705
 706                 for (j = 0; j < numdevs; j++) {
 707                         int num_match_categories;
 708
 709                         num_match_categories = matches[i].num_match_categories;
 710
 711                         /*
 712                          * Determine whether or not the current device
 713                          * matches the given matching expression.  This if
 714                          * statement consists of three components:
 715                          *   - the device type check
 716                          *   - the device interface check
 717                          *   - the passthrough check
 718                          * If a the matching test is successful, it
 719                          * decrements the number of matching categories,
 720                          * and if we've reached the last element that
 721                          * needed to be matched, the if statement succeeds.
 722                          *
 723                          */
 724                         if ((((matches[i].match_fields & DEVSTAT_MATCH_TYPE)!=0)
 725                           && ((devices[j].device_type & DEVSTAT_TYPE_MASK) ==
 726                                 (matches[i].device_type & DEVSTAT_TYPE_MASK))
 727                           &&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 728                            || (((matches[i].match_fields &
 729                                 DEVSTAT_MATCH_PASS) == 0)
 730                             && ((devices[j].device_type &
 731                                 DEVSTAT_TYPE_PASS) == 0)))
 732                           && (--num_match_categories == 0))
 733                          || (((matches[i].match_fields & DEVSTAT_MATCH_IF) != 0)
 734                           && ((devices[j].device_type & DEVSTAT_TYPE_IF_MASK) ==
 735                                 (matches[i].device_type & DEVSTAT_TYPE_IF_MASK))
 736                           &&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 737                            || (((matches[i].match_fields &
 738                                 DEVSTAT_MATCH_PASS) == 0)
 739                             && ((devices[j].device_type &
 740                                 DEVSTAT_TYPE_PASS) == 0)))
 741                           && (--num_match_categories == 0))
 742                          || (((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 743                           && ((devices[j].device_type & DEVSTAT_TYPE_PASS) != 0)
 744                           && (--num_match_categories == 0))) {
 745
 746                                 /*
 747                                  * This is probably a non-optimal solution
 748                                  * to the problem that the devices in the
 749                                  * device list will not be in the same
 750                                  * order as the devices in the selection
 751                                  * array.
 752                                  */
 753                                 for (k = 0; k < numdevs; k++) {
 754                                         if ((*dev_select)[k].position == j) {
 755                                                 found = 1;
 756                                                 break;
 757                                         }
 758                                 }
 759
 760                                 /*
 761                                  * There shouldn't be a case where a device
 762                                  * in the device list is not in the
 763                                  * selection list...but it could happen.
 764                                  */
 765                                 if (found != 1) {
 766                                         fprintf(stderr, "selectdevs: couldn't"
 767                                                 " find %s%d in selection "
 768                                                 "list\n",
 769                                                 devices[j].device_name,
 770                                                 devices[j].unit_number);
 771                                         break;
 772                                 }
 773
 774                                 /*
 775                                  * We do different things based upon the
 776                                  * mode we're in.  If we're in add or only
 777                                  * mode, we go ahead and select this device
 778                                  * if it hasn't already been selected.  If
 779                                  * it has already been selected, we leave
 780                                  * it alone so we don't mess up the
 781                                  * selection ordering.  Manually specified
 782                                  * devices have already been selected, and
 783                                  * they have higher priority than pattern
 784                                  * matched devices.  If we're in remove
 785                                  * mode, we de-select the given device and
 786                                  * decrement the selected count.
 787                                  */
 788                                 switch(select_mode) {
 789                                 case DS_SELECT_ADD:
 790                                 case DS_SELECT_ADDONLY:
 791                                 case DS_SELECT_ONLY:
 792                                         if ((*dev_select)[k].selected != 0)
 793                                                 break;
 794                                         (*dev_select)[k].selected =
 795                                                 ++selection_number;
 796                                         (*num_selected)++;
 797                                         break;
 798                                 case DS_SELECT_REMOVE:
 799                                         (*dev_select)[k].selected = 0;
 800                                         (*num_selected)--;
 801                                         break;
 802                                 }
 803                         }
 804                 }
 805         }
 806
 807         /*
 808          * Here we implement "top" mode.  Devices are sorted in the
 809          * selection array based on two criteria:  whether or not they are
 810          * selected (not selection number, just the fact that they are
 811          * selected!) and the number of bytes in the "bytes" field of the
 812          * selection structure.  The bytes field generally must be kept up
 813          * by the user.  In the future, it may be maintained by library
 814          * functions, but for now the user has to do the work.
 815          *
 816          * At first glance, it may seem wrong that we don't go through and
 817          * select every device in the case where the user hasn't specified
 818          * any devices or patterns.  In fact, though, it won't make any
 819          * difference in the device sorting.  In that particular case (i.e.
 820          * when we're in "add" or "only" mode, and the user hasn't
 821          * specified anything) the first time through no devices will be
 822          * selected, so the only criterion used to sort them will be their
 823          * performance.  The second time through, and every time thereafter,
 824          * all devices will be selected, so again selection won't matter.
 825          */
 826         if (perf_select != 0) {
 827
 828                 /* Sort the device array by throughput  */
 829                 qsort(*dev_select, *num_selections,
 830                       sizeof(struct device_selection),
 831                       compare_select);
 832
 833                 if (*num_selected == 0) {
 834                         /*
 835                          * Here we select every device in the array, if it
 836                          * isn't already selected.  Because the 'selected'
 837                          * variable in the selection array entries contains
 838                          * the selection order, the devstats routine can show
 839                          * the devices that were selected first.
 840                          */
 841                         for (i = 0; i < *num_selections; i++) {
 842                                 if ((*dev_select)[i].selected == 0) {
 843                                         (*dev_select)[i].selected =
 844                                                 ++selection_number;
 845                                         (*num_selected)++;
 846                                 }
 847                         }
 848                 } else {
 849                         selection_number = 0;
 850                         for (i = 0; i < *num_selections; i++) {
 851                                 if ((*dev_select)[i].selected != 0) {
 852                                         (*dev_select)[i].selected =
 853                                                 ++selection_number;
 854                                 }
 855                         }
 856                 }
 857         }
 858
 859         /*
 860          * If we're in the "add" selection mode and if we haven't already
 861          * selected maxshowdevs number of devices, go through the array and
 862          * select any unselected devices.  If we're in "only" mode, we
 863          * obviously don't want to select anything other than what the user
 864          * specifies.  If we're in "remove" mode, it probably isn't a good
 865          * idea to go through and select any more devices, since we might
 866          * end up selecting something that the user wants removed.  Through
 867          * more complicated logic, we could actually figure this out, but
 868          * that would probably require combining this loop with the various
 869          * selections loops above.
 870          */
 871         if ((select_mode == DS_SELECT_ADD) && (*num_selected < maxshowdevs)) {
 872                 for (i = 0; i < *num_selections; i++)
 873                         if ((*dev_select)[i].selected == 0) {
 874                                 (*dev_select)[i].selected = ++selection_number;
 875                                 (*num_selected)++;
 876                         }
 877         }
 878
 879         /*
 880          * Look at the number of devices that have been selected.  If it
 881          * has changed, set the changed variable.  Otherwise, if we've
 882          * made a backup of the selection list, compare it to the current
 883          * selection list to see if the selected devices have changed.
 884          */
 885         if ((changed == 0) && (old_num_selected != *num_selected))
 886                 changed = 1;
 887         else if ((changed == 0) && (old_dev_select != NULL)) {
 888                 /*
 889                  * Now we go through the selection list and we look at
 890                  * it three different ways.
 891                  */
 892                 for (i = 0; (i < *num_selections) && (changed == 0) &&
 893                      (i < old_num_selections); i++) {
 894                         /*
 895                          * If the device at index i in both the new and old
 896                          * selection arrays has the same device number and
 897                          * selection status, it hasn't changed.  We
 898                          * continue on to the next index.
 899                          */
 900                         if (((*dev_select)[i].device_number ==
 901                              old_dev_select[i].device_number)
 902                          && ((*dev_select)[i].selected ==
 903                              old_dev_select[i].selected))
 904                                 continue;
 905
 906                         /*
 907                          * Now, if we're still going through the if
 908                          * statement, the above test wasn't true.  So we
 909                          * check here to see if the device at index i in
 910                          * the current array is the same as the device at
 911                          * index i in the old array.  If it is, that means
 912                          * that its selection number has changed.  Set
 913                          * changed to 1 and exit the loop.
 914                          */
 915                         else if ((*dev_select)[i].device_number ==
 916                                   old_dev_select[i].device_number) {
 917                                 changed = 1;
 918                                 break;
 919                         }
 920                         /*
 921                          * If we get here, then the device at index i in
 922                          * the current array isn't the same device as the
 923                          * device at index i in the old array.
 924                          */
 925                         else {
 926                                 found = 0;
 927
 928                                 /*
 929                                  * Search through the old selection array
 930                                  * looking for a device with the same
 931                                  * device number as the device at index i
 932                                  * in the current array.  If the selection
 933                                  * status is the same, then we mark it as
 934                                  * found.  If the selection status isn't
 935                                  * the same, we break out of the loop.
 936                                  * Since found isn't set, changed will be
 937                                  * set to 1 below.
 938                                  */
 939                                 for (j = 0; j < old_num_selections; j++) {
 940                                         if (((*dev_select)[i].device_number ==
 941                                               old_dev_select[j].device_number)
 942                                          && ((*dev_select)[i].selected ==
 943                                               old_dev_select[j].selected)){
 944                                                 found = 1;
 945                                                 break;
 946                                         }
 947                                         else if ((*dev_select)[i].device_number
 948                                             == old_dev_select[j].device_number)
 949                                                 break;
 950                                 }
 951                                 if (found == 0)
 952                                         changed = 1;
 953                         }
 954                 }
 955         }
 956         if (old_dev_select != NULL)
 957                 free(old_dev_select);
 958
 959         return(changed);
 960 }
 961
 962 /*
 963  * Comparison routine for qsort() above.  Note that the comparison here is
 964  * backwards -- generally, it should return a value to indicate whether
 965  * arg1 is <, =, or > arg2.  Instead, it returns the opposite.  The reason
 966  * it returns the opposite is so that the selection array will be sorted in
 967  * order of decreasing performance.  We sort on two parameters.  The first
 968  * sort key is whether or not one or the other of the devices in question
 969  * has been selected.  If one of them has, and the other one has not, the
 970  * selected device is automatically more important than the unselected
 971  * device.  If neither device is selected, we judge the devices based upon
 972  * performance.
 973  */
 974 static int
 975 compare_select(const void *arg1, const void *arg2)
 976 {
 977         if ((((const struct device_selection *)arg1)->selected)
 978          && (((const struct device_selection *)arg2)->selected == 0))
 979                 return(-1);
 980         else if ((((const struct device_selection *)arg1)->selected == 0)
 981               && (((const struct device_selection *)arg2)->selected))
 982                 return(1);
 983         else if (((const struct device_selection *)arg2)->bytes <
 984                  ((const struct device_selection *)arg1)->bytes)
 985                 return(-1);
 986         else if (((const struct device_selection *)arg2)->bytes >
 987                  ((const struct device_selection *)arg1)->bytes)
 988                 return(1);
 989         else
 990                 return(0);
 991 }
 992
 993 /*
 994  * Take a string with the general format "arg1,arg2,arg3", and build a
 995  * device matching expression from it.
 996  */
 997 int
 998 devstat_buildmatch(char *match_str, struct devstat_match **matches,
 999                    int *num_matches)
1000 {
1001         char *tstr[5];
1002         char **tempstr;
1003         int num_args;
1004         int i, j;
1005
1006         /* We can't do much without a string to parse */
1007         if (match_str == NULL) {
1008                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1009                          "%s: no match expression", __func__);
1010                 return(-1);
1011         }
1012
1013         /*
1014          * Break the (comma delimited) input string out into separate strings.
1015          */
1016         for (tempstr = tstr, num_args  = 0;
1017              (*tempstr = strsep(&match_str, ",")) != NULL && (num_args < 5);)
1018                 if (**tempstr != '\0') {
1019                         num_args++;
1020                         if (++tempstr >= &tstr[5])
1021                                 break;
1022                 }
1023
1024         /* The user gave us too many type arguments */
1025         if (num_args > 3) {
1026                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1027                          "%s: too many type arguments", __func__);
1028                 return(-1);
1029         }
1030
1031         /*
1032          * Since you can't realloc a pointer that hasn't been malloced
1033          * first, we malloc first and then realloc.
1034          */
1035         if (*num_matches == 0)
1036                 *matches = (struct devstat_match *)malloc(
1037                            sizeof(struct devstat_match));
1038         else
1039                 *matches = (struct devstat_match *)realloc(*matches,
1040                           sizeof(struct devstat_match) * (*num_matches + 1));
1041
1042         /* Make sure the current entry is clear */
1043         bzero(&matches[0][*num_matches], sizeof(struct devstat_match));
1044
1045         /*
1046          * Step through the arguments the user gave us and build a device
1047          * matching expression from them.
1048          */
1049         for (i = 0; i < num_args; i++) {
1050                 char *tempstr2, *tempstr3;
1051
1052                 /*
1053                  * Get rid of leading white space.
1054                  */
1055                 tempstr2 = tstr[i];
1056                 while (isspace(*tempstr2) && (*tempstr2 != '\0'))
1057                         tempstr2++;
1058
1059                 /*
1060                  * Get rid of trailing white space.
1061                  */
1062                 tempstr3 = &tempstr2[strlen(tempstr2) - 1];
1063
1064                 while ((*tempstr3 != '\0') && (tempstr3 > tempstr2)
1065                     && (isspace(*tempstr3))) {
1066                         *tempstr3 = '\0';
1067                         tempstr3--;
1068                 }
1069
1070                 /*
1071                  * Go through the match table comparing the user's
1072                  * arguments to known device types, interfaces, etc.
1073                  */
1074                 for (j = 0; match_table[j].match_str != NULL; j++) {
1075                         /*
1076                          * We do case-insensitive matching, in case someone
1077                          * wants to enter "SCSI" instead of "scsi" or
1078                          * something like that.  Only compare as many
1079                          * characters as are in the string in the match
1080                          * table.  This should help if someone tries to use
1081                          * a super-long match expression.
1082                          */
1083                         if (strncasecmp(tempstr2, match_table[j].match_str,
1084                             strlen(match_table[j].match_str)) == 0) {
1085                                 /*
1086                                  * Make sure the user hasn't specified two
1087                                  * items of the same type, like "da" and
1088                                  * "cd".  One device cannot be both.
1089                                  */
1090                                 if (((*matches)[*num_matches].match_fields &
1091                                     match_table[j].match_field) != 0) {
1092                                         snprintf(devstat_errbuf,
1093                                                  sizeof(devstat_errbuf),
1094                                                  "%s: cannot have more than "
1095                                                  "one match item in a single "
1096                                                  "category", __func__);
1097                                         return(-1);
1098                                 }
1099                                 /*
1100                                  * If we've gotten this far, we have a
1101                                  * winner.  Set the appropriate fields in
1102                                  * the match entry.
1103                                  */
1104                                 (*matches)[*num_matches].match_fields |=
1105                                         match_table[j].match_field;
1106                                 (*matches)[*num_matches].device_type |=
1107                                         match_table[j].type;
1108                                 (*matches)[*num_matches].num_match_categories++;
1109                                 break;
1110                         }
1111                 }
1112                 /*
1113                  * We should have found a match in the above for loop.  If
1114                  * not, that means the user entered an invalid device type
1115                  * or interface.
1116                  */
1117                 if ((*matches)[*num_matches].num_match_categories != (i + 1)) {
1118                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1119                                  "%s: unknown match item \"%s\"", __func__,
1120                                  tstr[i]);
1121                         return(-1);
1122                 }
1123         }
1124
1125         (*num_matches)++;
1126
1127         return(0);
1128 }
1129
1130 /*
1131  * Compute a number of device statistics.  Only one field is mandatory, and
1132  * that is "current".  Everything else is optional.  The caller passes in
1133  * pointers to variables to hold the various statistics he desires.  If he
1134  * doesn't want a particular staistic, he should pass in a NULL pointer.
1135  * Return values:
1136  * 0   -- success
1137  * -1  -- failure
1138  */
1139 int
1140 compute_stats(struct devstat *current, struct devstat *previous,
1141               long double etime, u_int64_t *total_bytes,
1142               u_int64_t *total_transfers, u_int64_t *total_blocks,
1143               long double *kb_per_transfer, long double *transfers_per_second,
1144               long double *mb_per_second, long double *blocks_per_second,
1145               long double *ms_per_transaction)
1146 {
1147         return(devstat_compute_statistics(current, previous, etime,
1148                total_bytes ? DSM_TOTAL_BYTES : DSM_SKIP,
1149                total_bytes,
1150                total_transfers ? DSM_TOTAL_TRANSFERS : DSM_SKIP,
1151                total_transfers,
1152                total_blocks ? DSM_TOTAL_BLOCKS : DSM_SKIP,
1153                total_blocks,
1154                kb_per_transfer ? DSM_KB_PER_TRANSFER : DSM_SKIP,
1155                kb_per_transfer,
1156                transfers_per_second ? DSM_TRANSFERS_PER_SECOND : DSM_SKIP,
1157                transfers_per_second,
1158                mb_per_second ? DSM_MB_PER_SECOND : DSM_SKIP,
1159                mb_per_second,
1160                blocks_per_second ? DSM_BLOCKS_PER_SECOND : DSM_SKIP,
1161                blocks_per_second,
1162                ms_per_transaction ? DSM_MS_PER_TRANSACTION : DSM_SKIP,
1163                ms_per_transaction,
1164                DSM_NONE));
1165 }
1166
1167
1168 /* This is 1/2^64 */
1169 #define BINTIME_SCALE 5.42101086242752217003726400434970855712890625e-20
1170
1171 long double
1172 devstat_compute_etime(struct bintime *cur_time, struct bintime *prev_time)
1173 {
1174         long double etime;
1175
1176         etime = cur_time->sec;
1177         etime += cur_time->frac * BINTIME_SCALE;
1178         if (prev_time != NULL) {
1179                 etime -= prev_time->sec;
1180                 etime -= prev_time->frac * BINTIME_SCALE;
1181         }
1182         return(etime);
1183 }
1184
1185 #define DELTA(field, index)                             \
1186         (current->field[(index)] - (previous ? previous->field[(index)] : 0))
1187
1188 #define DELTA_T(field)                                  \
1189         devstat_compute_etime(&current->field,          \
1190         (previous ? &previous->field : NULL))
1191
1192 int
1193 devstat_compute_statistics(struct devstat *current, struct devstat *previous,
1194                            long double etime, ...)
1195 {
1196         u_int64_t totalbytes, totalbytesread, totalbyteswrite, totalbytesfree;
1197         u_int64_t totaltransfers, totaltransfersread, totaltransferswrite;
1198         u_int64_t totaltransfersother, totalblocks, totalblocksread;
1199         u_int64_t totalblockswrite, totaltransfersfree, totalblocksfree;
1200         va_list ap;
1201         devstat_metric metric;
1202         u_int64_t *destu64;
1203         long double *destld;
1204         int retval, i;
1205
1206         retval = 0;
1207
1208         /*
1209          * current is the only mandatory field.
1210          */
1211         if (current == NULL) {
1212                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1213                          "%s: current stats structure was NULL", __func__);
1214                 return(-1);
1215         }
1216
1217         totalbytesread = DELTA(bytes, DEVSTAT_READ);
1218         totalbyteswrite = DELTA(bytes, DEVSTAT_WRITE);
1219         totalbytesfree = DELTA(bytes, DEVSTAT_FREE);
1220         totalbytes = totalbytesread + totalbyteswrite + totalbytesfree;
1221
1222         totaltransfersread = DELTA(operations, DEVSTAT_READ);
1223         totaltransferswrite = DELTA(operations, DEVSTAT_WRITE);
1224         totaltransfersother = DELTA(operations, DEVSTAT_NO_DATA);
1225         totaltransfersfree = DELTA(operations, DEVSTAT_FREE);
1226         totaltransfers = totaltransfersread + totaltransferswrite +
1227                          totaltransfersother + totaltransfersfree;
1228
1229         totalblocks = totalbytes;
1230         totalblocksread = totalbytesread;
1231         totalblockswrite = totalbyteswrite;
1232         totalblocksfree = totalbytesfree;
1233
1234         if (current->block_size > 0) {
1235                 totalblocks /= current->block_size;
1236                 totalblocksread /= current->block_size;
1237                 totalblockswrite /= current->block_size;
1238                 totalblocksfree /= current->block_size;
1239         } else {
1240                 totalblocks /= 512;
1241                 totalblocksread /= 512;
1242                 totalblockswrite /= 512;
1243                 totalblocksfree /= 512;
1244         }
1245
1246         va_start(ap, etime);
1247
1248         while ((metric = (devstat_metric)va_arg(ap, devstat_metric)) != 0) {
1249
1250                 if (metric == DSM_NONE)
1251                         break;
1252
1253                 if (metric >= DSM_MAX) {
1254                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1255                                  "%s: metric %d is out of range", __func__,
1256                                  metric);
1257                         retval = -1;
1258                         goto bailout;
1259                 }
1260
1261                 switch (devstat_arg_list[metric].argtype) {
1262                 case DEVSTAT_ARG_UINT64:
1263                         destu64 = (u_int64_t *)va_arg(ap, u_int64_t *);
1264                         break;
1265                 case DEVSTAT_ARG_LD:
1266                         destld = (long double *)va_arg(ap, long double *);
1267                         break;
1268                 case DEVSTAT_ARG_SKIP:
1269                         destld = (long double *)va_arg(ap, long double *);
1270                         break;
1271                 default:
1272                         retval = -1;
1273                         goto bailout;
1274                         break; /* NOTREACHED */
1275                 }
1276
1277                 if (devstat_arg_list[metric].argtype == DEVSTAT_ARG_SKIP)
1278                         continue;
1279
1280                 switch (metric) {
1281                 case DSM_TOTAL_BYTES:
1282                         *destu64 = totalbytes;
1283                         break;
1284                 case DSM_TOTAL_BYTES_READ:
1285                         *destu64 = totalbytesread;
1286                         break;
1287                 case DSM_TOTAL_BYTES_WRITE:
1288                         *destu64 = totalbyteswrite;
1289                         break;
1290                 case DSM_TOTAL_BYTES_FREE:
1291                         *destu64 = totalbytesfree;
1292                         break;
1293                 case DSM_TOTAL_TRANSFERS:
1294                         *destu64 = totaltransfers;
1295                         break;
1296                 case DSM_TOTAL_TRANSFERS_READ:
1297                         *destu64 = totaltransfersread;
1298                         break;
1299                 case DSM_TOTAL_TRANSFERS_WRITE:
1300                         *destu64 = totaltransferswrite;
1301                         break;
1302                 case DSM_TOTAL_TRANSFERS_FREE:
1303                         *destu64 = totaltransfersfree;
1304                         break;
1305                 case DSM_TOTAL_TRANSFERS_OTHER:
1306                         *destu64 = totaltransfersother;
1307                         break;
1308                 case DSM_TOTAL_BLOCKS:
1309                         *destu64 = totalblocks;
1310                         break;
1311                 case DSM_TOTAL_BLOCKS_READ:
1312                         *destu64 = totalblocksread;
1313                         break;
1314                 case DSM_TOTAL_BLOCKS_WRITE:
1315                         *destu64 = totalblockswrite;
1316                         break;
1317                 case DSM_TOTAL_BLOCKS_FREE:
1318                         *destu64 = totalblocksfree;
1319                         break;
1320                 case DSM_KB_PER_TRANSFER:
1321                         *destld = totalbytes;
1322                         *destld /= 1024;
1323                         if (totaltransfers > 0)
1324                                 *destld /= totaltransfers;
1325                         else
1326                                 *destld = 0.0;
1327                         break;
1328                 case DSM_KB_PER_TRANSFER_READ:
1329                         *destld = totalbytesread;
1330                         *destld /= 1024;
1331                         if (totaltransfersread > 0)
1332                                 *destld /= totaltransfersread;
1333                         else
1334                                 *destld = 0.0;
1335                         break;
1336                 case DSM_KB_PER_TRANSFER_WRITE:
1337                         *destld = totalbyteswrite;
1338                         *destld /= 1024;
1339                         if (totaltransferswrite > 0)
1340                                 *destld /= totaltransferswrite;
1341                         else
1342                                 *destld = 0.0;
1343                         break;
1344                 case DSM_KB_PER_TRANSFER_FREE:
1345                         *destld = totalbytesfree;
1346                         *destld /= 1024;
1347                         if (totaltransfersfree > 0)
1348                                 *destld /= totaltransfersfree;
1349                         else
1350                                 *destld = 0.0;
1351                         break;
1352                 case DSM_TRANSFERS_PER_SECOND:
1353                         if (etime > 0.0) {
1354                                 *destld = totaltransfers;
1355                                 *destld /= etime;
1356                         } else
1357                                 *destld = 0.0;
1358                         break;
1359                 case DSM_TRANSFERS_PER_SECOND_READ:
1360                         if (etime > 0.0) {
1361                                 *destld = totaltransfersread;
1362                                 *destld /= etime;
1363                         } else
1364                                 *destld = 0.0;
1365                         break;
1366                 case DSM_TRANSFERS_PER_SECOND_WRITE:
1367                         if (etime > 0.0) {
1368                                 *destld = totaltransferswrite;
1369                                 *destld /= etime;
1370                         } else
1371                                 *destld = 0.0;
1372                         break;
1373                 case DSM_TRANSFERS_PER_SECOND_FREE:
1374                         if (etime > 0.0) {
1375                                 *destld = totaltransfersfree;
1376                                 *destld /= etime;
1377                         } else
1378                                 *destld = 0.0;
1379                         break;
1380                 case DSM_TRANSFERS_PER_SECOND_OTHER:
1381                         if (etime > 0.0) {
1382                                 *destld = totaltransfersother;
1383                                 *destld /= etime;
1384                         } else
1385                                 *destld = 0.0;
1386                         break;
1387                 case DSM_MB_PER_SECOND:
1388                         *destld = totalbytes;
1389                         *destld /= 1024 * 1024;
1390                         if (etime > 0.0)
1391                                 *destld /= etime;
1392                         else
1393                                 *destld = 0.0;
1394                         break;
1395                 case DSM_MB_PER_SECOND_READ:
1396                         *destld = totalbytesread;
1397                         *destld /= 1024 * 1024;
1398                         if (etime > 0.0)
1399                                 *destld /= etime;
1400                         else
1401                                 *destld = 0.0;
1402                         break;
1403                 case DSM_MB_PER_SECOND_WRITE:
1404                         *destld = totalbyteswrite;
1405                         *destld /= 1024 * 1024;
1406                         if (etime > 0.0)
1407                                 *destld /= etime;
1408                         else
1409                                 *destld = 0.0;
1410                         break;
1411                 case DSM_MB_PER_SECOND_FREE:
1412                         *destld = totalbytesfree;
1413                         *destld /= 1024 * 1024;
1414                         if (etime > 0.0)
1415                                 *destld /= etime;
1416                         else
1417                                 *destld = 0.0;
1418                         break;
1419                 case DSM_BLOCKS_PER_SECOND:
1420                         *destld = totalblocks;
1421                         if (etime > 0.0)
1422                                 *destld /= etime;
1423                         else
1424                                 *destld = 0.0;
1425                         break;
1426                 case DSM_BLOCKS_PER_SECOND_READ:
1427                         *destld = totalblocksread;
1428                         if (etime > 0.0)
1429                                 *destld /= etime;
1430                         else
1431                                 *destld = 0.0;
1432                         break;
1433                 case DSM_BLOCKS_PER_SECOND_WRITE:
1434                         *destld = totalblockswrite;
1435                         if (etime > 0.0)
1436                                 *destld /= etime;
1437                         else
1438                                 *destld = 0.0;
1439                         break;
1440                 case DSM_BLOCKS_PER_SECOND_FREE:
1441                         *destld = totalblocksfree;
1442                         if (etime > 0.0)
1443                                 *destld /= etime;
1444                         else
1445                                 *destld = 0.0;
1446                         break;
1447                 /*
1448                  * This calculation is somewhat bogus.  It simply divides
1449                  * the elapsed time by the total number of transactions
1450                  * completed.  While that does give the caller a good
1451                  * picture of the average rate of transaction completion,
1452                  * it doesn't necessarily give the caller a good view of
1453                  * how long transactions took to complete on average.
1454                  * Those two numbers will be different for a device that
1455                  * can handle more than one transaction at a time.  e.g.
1456                  * SCSI disks doing tagged queueing.
1457                  *
1458                  * The only way to accurately determine the real average
1459                  * time per transaction would be to compute and store the
1460                  * time on a per-transaction basis.  That currently isn't
1461                  * done in the kernel, and would only be desireable if it
1462                  * could be implemented in a somewhat non-intrusive and high
1463                  * performance way.
1464                  */
1465                 case DSM_MS_PER_TRANSACTION:
1466                         if (totaltransfers > 0) {
1467                                 *destld = 0;
1468                                 for (i = 0; i < DEVSTAT_N_TRANS_FLAGS; i++)
1469                                         *destld += DELTA_T(duration[i]);
1470                                 *destld /= totaltransfers;
1471                                 *destld *= 1000;
1472                         } else
1473                                 *destld = 0.0;
1474                         break;
1475                 /*
1476                  * As above, these next two really only give the average
1477                  * rate of completion for read and write transactions, not
1478                  * the average time the transaction took to complete.
1479                  */
1480                 case DSM_MS_PER_TRANSACTION_READ:
1481                         if (totaltransfersread > 0) {
1482                                 *destld = DELTA_T(duration[DEVSTAT_READ]);
1483                                 *destld /= totaltransfersread;
1484                                 *destld *= 1000;
1485                         } else
1486                                 *destld = 0.0;
1487                         break;
1488                 case DSM_MS_PER_TRANSACTION_WRITE:
1489                         if (totaltransferswrite > 0) {
1490                                 *destld = DELTA_T(duration[DEVSTAT_WRITE]);
1491                                 *destld /= totaltransferswrite;
1492                                 *destld *= 1000;
1493                         } else
1494                                 *destld = 0.0;
1495                         break;
1496                 case DSM_MS_PER_TRANSACTION_FREE:
1497                         if (totaltransfersfree > 0) {
1498                                 *destld = DELTA_T(duration[DEVSTAT_FREE]);
1499                                 *destld /= totaltransfersfree;
1500                                 *destld *= 1000;
1501                         } else
1502                                 *destld = 0.0;
1503                         break;
1504                 case DSM_MS_PER_TRANSACTION_OTHER:
1505                         if (totaltransfersother > 0) {
1506                                 *destld = DELTA_T(duration[DEVSTAT_NO_DATA]);
1507                                 *destld /= totaltransfersother;
1508                                 *destld *= 1000;
1509                         } else
1510                                 *destld = 0.0;
1511                         break;
1512                 case DSM_BUSY_PCT:
1513                         *destld = DELTA_T(busy_time);
1514                         if (*destld < 0)
1515                                 *destld = 0;
1516                         *destld /= etime;
1517                         *destld *= 100;
1518                         if (*destld < 0)
1519                                 *destld = 0;
1520                         break;
1521                 case DSM_QUEUE_LENGTH:
1522                         *destu64 = current->start_count - current->end_count;
1523                         break;
1524 /*
1525  * XXX: comment out the default block to see if any case's are missing.
1526  */
1527 #if 1
1528                 default:
1529                         /*
1530                          * This shouldn't happen, since we should have
1531                          * caught any out of range metrics at the top of
1532                          * the loop.
1533                          */
1534                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1535                                  "%s: unknown metric %d", __func__, metric);
1536                         retval = -1;
1537                         goto bailout;
1538                         break; /* NOTREACHED */
1539 #endif
1540                 }
1541         }
1542
1543 bailout:
1544
1545         va_end(ap);
1546         return(retval);
1547 }
1548
1549 static int
1550 readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes)
1551 {
1552
1553         if (kvm_read(kd, addr, buf, nbytes) == -1) {
1554                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1555                          "%s: error reading value (kvm_read): %s", __func__,
1556                          kvm_geterr(kd));
1557                 return(-1);
1558         }
1559         return(0);
1560 }
1561
1562 static int
1563 readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes)
1564 {
1565         struct nlist nl[2];
1566
1567         nl[0].n_name = (char *)name;
1568         nl[1].n_name = NULL;
1569
1570         if (kvm_nlist(kd, nl) == -1) {
1571                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1572                          "%s: error getting name list (kvm_nlist): %s",
1573                          __func__, kvm_geterr(kd));
1574                 return(-1);
1575         }
1576         return(readkmem(kd, nl[0].n_value, buf, nbytes));
1577 }
1578
1579 /*
1580  * This duplicates the functionality of the kernel sysctl handler for poking
1581  * through crash dumps.
1582  */
1583 static char *
1584 get_devstat_kvm(kvm_t *kd)
1585 {
1586         int i, wp;
1587         long gen;
1588         struct devstat *nds;
1589         struct devstat ds;
1590         struct devstatlist dhead;
1591         int num_devs;
1592         char *rv = NULL;
1593
1594         if ((num_devs = devstat_getnumdevs(kd)) <= 0)
1595                 return(NULL);
1596         if (KREADNL(kd, X_DEVICE_STATQ, dhead) == -1)
1597                 return(NULL);
1598
1599         nds = STAILQ_FIRST(&dhead);
1600
1601         if ((rv = malloc(sizeof(gen))) == NULL) {
1602                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1603                          "%s: out of memory (initial malloc failed)",
1604                          __func__);
1605                 return(NULL);
1606         }
1607         gen = devstat_getgeneration(kd);
1608         memcpy(rv, &gen, sizeof(gen));
1609         wp = sizeof(gen);
1610         /*
1611          * Now push out all the devices.
1612          */
1613         for (i = 0; (nds != NULL) && (i < num_devs);
1614              nds = STAILQ_NEXT(nds, dev_links), i++) {
1615                 if (readkmem(kd, (long)nds, &ds, sizeof(ds)) == -1) {
1616                         free(rv);
1617                         return(NULL);
1618                 }
1619                 nds = &ds;
1620                 rv = (char *)reallocf(rv, sizeof(gen) +
1621                                       sizeof(ds) * (i + 1));
1622                 if (rv == NULL) {
1623                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1624                                  "%s: out of memory (malloc failed)",
1625                                  __func__);
1626                         return(NULL);
1627                 }
1628                 memcpy(rv + wp, &ds, sizeof(ds));
1629                 wp += sizeof(ds);
1630         }
1631         return(rv);
1632 }