lib/libdevstat/devstat.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 1997, 1998 Kenneth D. Merry.
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. The name of the author may not be used to endorse or promote products
  16  *    derived from this software without specific prior written permission.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  */
  30
  31 #include <sys/types.h>
  32 #include <sys/sysctl.h>
  33 #include <sys/errno.h>
  34 #include <sys/resource.h>
  35 #include <sys/queue.h>
  36
  37 #include <ctype.h>
  38 #include <err.h>
  39 #include <fcntl.h>
  40 #include <limits.h>
  41 #include <stdio.h>
  42 #include <stdlib.h>
  43 #include <string.h>
  44 #include <stdarg.h>
  45 #include <kvm.h>
  46 #include <nlist.h>
  47
  48 #include "devstat.h"
  49
  50 int
  51 compute_stats(struct devstat *current, struct devstat *previous,
  52               long double etime, u_int64_t *total_bytes,
  53               u_int64_t *total_transfers, u_int64_t *total_blocks,
  54               long double *kb_per_transfer, long double *transfers_per_second,
  55               long double *mb_per_second, long double *blocks_per_second,
  56               long double *ms_per_transaction);
  57
  58 typedef enum {
  59         DEVSTAT_ARG_NOTYPE,
  60         DEVSTAT_ARG_UINT64,
  61         DEVSTAT_ARG_LD,
  62         DEVSTAT_ARG_SKIP
  63 } devstat_arg_type;
  64
  65 char devstat_errbuf[DEVSTAT_ERRBUF_SIZE];
  66
  67 /*
  68  * Table to match descriptive strings with device types.  These are in
  69  * order from most common to least common to speed search time.
  70  */
  71 struct devstat_match_table match_table[] = {
  72         {"da",          DEVSTAT_TYPE_DIRECT,    DEVSTAT_MATCH_TYPE},
  73         {"cd",          DEVSTAT_TYPE_CDROM,     DEVSTAT_MATCH_TYPE},
  74         {"scsi",        DEVSTAT_TYPE_IF_SCSI,   DEVSTAT_MATCH_IF},
  75         {"ide",         DEVSTAT_TYPE_IF_IDE,    DEVSTAT_MATCH_IF},
  76         {"other",       DEVSTAT_TYPE_IF_OTHER,  DEVSTAT_MATCH_IF},
  77         {"nvme",        DEVSTAT_TYPE_IF_NVME,   DEVSTAT_MATCH_IF},
  78         {"worm",        DEVSTAT_TYPE_WORM,      DEVSTAT_MATCH_TYPE},
  79         {"sa",          DEVSTAT_TYPE_SEQUENTIAL,DEVSTAT_MATCH_TYPE},
  80         {"pass",        DEVSTAT_TYPE_PASS,      DEVSTAT_MATCH_PASS},
  81         {"optical",     DEVSTAT_TYPE_OPTICAL,   DEVSTAT_MATCH_TYPE},
  82         {"array",       DEVSTAT_TYPE_STORARRAY, DEVSTAT_MATCH_TYPE},
  83         {"changer",     DEVSTAT_TYPE_CHANGER,   DEVSTAT_MATCH_TYPE},
  84         {"scanner",     DEVSTAT_TYPE_SCANNER,   DEVSTAT_MATCH_TYPE},
  85         {"printer",     DEVSTAT_TYPE_PRINTER,   DEVSTAT_MATCH_TYPE},
  86         {"floppy",      DEVSTAT_TYPE_FLOPPY,    DEVSTAT_MATCH_TYPE},
  87         {"proc",        DEVSTAT_TYPE_PROCESSOR, DEVSTAT_MATCH_TYPE},
  88         {"comm",        DEVSTAT_TYPE_COMM,      DEVSTAT_MATCH_TYPE},
  89         {"enclosure",   DEVSTAT_TYPE_ENCLOSURE, DEVSTAT_MATCH_TYPE},
  90         {NULL,          0,                      0}
  91 };
  92
  93 struct devstat_args {
  94         devstat_metric          metric;
  95         devstat_arg_type        argtype;
  96 } devstat_arg_list[] = {
  97         { DSM_NONE, DEVSTAT_ARG_NOTYPE },
  98         { DSM_TOTAL_BYTES, DEVSTAT_ARG_UINT64 },
  99         { DSM_TOTAL_BYTES_READ, DEVSTAT_ARG_UINT64 },
 100         { DSM_TOTAL_BYTES_WRITE, DEVSTAT_ARG_UINT64 },
 101         { DSM_TOTAL_TRANSFERS, DEVSTAT_ARG_UINT64 },
 102         { DSM_TOTAL_TRANSFERS_READ, DEVSTAT_ARG_UINT64 },
 103         { DSM_TOTAL_TRANSFERS_WRITE, DEVSTAT_ARG_UINT64 },
 104         { DSM_TOTAL_TRANSFERS_OTHER, DEVSTAT_ARG_UINT64 },
 105         { DSM_TOTAL_BLOCKS, DEVSTAT_ARG_UINT64 },
 106         { DSM_TOTAL_BLOCKS_READ, DEVSTAT_ARG_UINT64 },
 107         { DSM_TOTAL_BLOCKS_WRITE, DEVSTAT_ARG_UINT64 },
 108         { DSM_KB_PER_TRANSFER, DEVSTAT_ARG_LD },
 109         { DSM_KB_PER_TRANSFER_READ, DEVSTAT_ARG_LD },
 110         { DSM_KB_PER_TRANSFER_WRITE, DEVSTAT_ARG_LD },
 111         { DSM_TRANSFERS_PER_SECOND, DEVSTAT_ARG_LD },
 112         { DSM_TRANSFERS_PER_SECOND_READ, DEVSTAT_ARG_LD },
 113         { DSM_TRANSFERS_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
 114         { DSM_TRANSFERS_PER_SECOND_OTHER, DEVSTAT_ARG_LD },
 115         { DSM_MB_PER_SECOND, DEVSTAT_ARG_LD },
 116         { DSM_MB_PER_SECOND_READ, DEVSTAT_ARG_LD },
 117         { DSM_MB_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
 118         { DSM_BLOCKS_PER_SECOND, DEVSTAT_ARG_LD },
 119         { DSM_BLOCKS_PER_SECOND_READ, DEVSTAT_ARG_LD },
 120         { DSM_BLOCKS_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
 121         { DSM_MS_PER_TRANSACTION, DEVSTAT_ARG_LD },
 122         { DSM_MS_PER_TRANSACTION_READ, DEVSTAT_ARG_LD },
 123         { DSM_MS_PER_TRANSACTION_WRITE, DEVSTAT_ARG_LD },
 124         { DSM_SKIP, DEVSTAT_ARG_SKIP },
 125         { DSM_TOTAL_BYTES_FREE, DEVSTAT_ARG_UINT64 },
 126         { DSM_TOTAL_TRANSFERS_FREE, DEVSTAT_ARG_UINT64 },
 127         { DSM_TOTAL_BLOCKS_FREE, DEVSTAT_ARG_UINT64 },
 128         { DSM_KB_PER_TRANSFER_FREE, DEVSTAT_ARG_LD },
 129         { DSM_MB_PER_SECOND_FREE, DEVSTAT_ARG_LD },
 130         { DSM_TRANSFERS_PER_SECOND_FREE, DEVSTAT_ARG_LD },
 131         { DSM_BLOCKS_PER_SECOND_FREE, DEVSTAT_ARG_LD },
 132         { DSM_MS_PER_TRANSACTION_OTHER, DEVSTAT_ARG_LD },
 133         { DSM_MS_PER_TRANSACTION_FREE, DEVSTAT_ARG_LD },
 134         { DSM_BUSY_PCT, DEVSTAT_ARG_LD },
 135         { DSM_QUEUE_LENGTH, DEVSTAT_ARG_UINT64 },
 136         { DSM_TOTAL_DURATION, DEVSTAT_ARG_LD },
 137         { DSM_TOTAL_DURATION_READ, DEVSTAT_ARG_LD },
 138         { DSM_TOTAL_DURATION_WRITE, DEVSTAT_ARG_LD },
 139         { DSM_TOTAL_DURATION_FREE, DEVSTAT_ARG_LD },
 140         { DSM_TOTAL_DURATION_OTHER, DEVSTAT_ARG_LD },
 141         { DSM_TOTAL_BUSY_TIME, DEVSTAT_ARG_LD },
 142 };
 143
 144 static const char *namelist[] = {
 145 #define X_NUMDEVS       0
 146         "_devstat_num_devs",
 147 #define X_GENERATION    1
 148         "_devstat_generation",
 149 #define X_VERSION       2
 150         "_devstat_version",
 151 #define X_DEVICE_STATQ  3
 152         "_device_statq",
 153 #define X_TIME_UPTIME   4
 154         "_time_uptime",
 155 #define X_END           5
 156 };
 157
 158 /*
 159  * Local function declarations.
 160  */
 161 static int compare_select(const void *arg1, const void *arg2);
 162 static int readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes);
 163 static int readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes);
 164 static char *get_devstat_kvm(kvm_t *kd);
 165
 166 #define KREADNL(kd, var, val) \
 167         readkmem_nl(kd, namelist[var], &val, sizeof(val))
 168
 169 int
 170 devstat_getnumdevs(kvm_t *kd)
 171 {
 172         size_t numdevsize;
 173         int numdevs;
 174
 175         numdevsize = sizeof(int);
 176
 177         /*
 178          * Find out how many devices we have in the system.
 179          */
 180         if (kd == NULL) {
 181                 if (sysctlbyname("kern.devstat.numdevs", &numdevs,
 182                                  &numdevsize, NULL, 0) == -1) {
 183                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 184                                  "%s: error getting number of devices\n"
 185                                  "%s: %s", __func__, __func__,
 186                                  strerror(errno));
 187                         return(-1);
 188                 } else
 189                         return(numdevs);
 190         } else {
 191
 192                 if (KREADNL(kd, X_NUMDEVS, numdevs) == -1)
 193                         return(-1);
 194                 else
 195                         return(numdevs);
 196         }
 197 }
 198
 199 /*
 200  * This is an easy way to get the generation number, but the generation is
 201  * supplied in a more atmoic manner by the kern.devstat.all sysctl.
 202  * Because this generation sysctl is separate from the statistics sysctl,
 203  * the device list and the generation could change between the time that
 204  * this function is called and the device list is retrieved.
 205  */
 206 long
 207 devstat_getgeneration(kvm_t *kd)
 208 {
 209         size_t gensize;
 210         long generation;
 211
 212         gensize = sizeof(long);
 213
 214         /*
 215          * Get the current generation number.
 216          */
 217         if (kd == NULL) {
 218                 if (sysctlbyname("kern.devstat.generation", &generation,
 219                                  &gensize, NULL, 0) == -1) {
 220                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 221                                  "%s: error getting devstat generation\n%s: %s",
 222                                  __func__, __func__, strerror(errno));
 223                         return(-1);
 224                 } else
 225                         return(generation);
 226         } else {
 227                 if (KREADNL(kd, X_GENERATION, generation) == -1)
 228                         return(-1);
 229                 else
 230                         return(generation);
 231         }
 232 }
 233
 234 /*
 235  * Get the current devstat version.  The return value of this function
 236  * should be compared with DEVSTAT_VERSION, which is defined in
 237  * sys/devicestat.h.  This will enable userland programs to determine
 238  * whether they are out of sync with the kernel.
 239  */
 240 int
 241 devstat_getversion(kvm_t *kd)
 242 {
 243         size_t versize;
 244         int version;
 245
 246         versize = sizeof(int);
 247
 248         /*
 249          * Get the current devstat version.
 250          */
 251         if (kd == NULL) {
 252                 if (sysctlbyname("kern.devstat.version", &version, &versize,
 253                                  NULL, 0) == -1) {
 254                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 255                                  "%s: error getting devstat version\n%s: %s",
 256                                  __func__, __func__, strerror(errno));
 257                         return(-1);
 258                 } else
 259                         return(version);
 260         } else {
 261                 if (KREADNL(kd, X_VERSION, version) == -1)
 262                         return(-1);
 263                 else
 264                         return(version);
 265         }
 266 }
 267
 268 /*
 269  * Check the devstat version we know about against the devstat version the
 270  * kernel knows about.  If they don't match, print an error into the
 271  * devstat error buffer, and return -1.  If they match, return 0.
 272  */
 273 int
 274 devstat_checkversion(kvm_t *kd)
 275 {
 276         int buflen, res, retval = 0, version;
 277
 278         version = devstat_getversion(kd);
 279
 280         if (version != DEVSTAT_VERSION) {
 281                 /*
 282                  * If getversion() returns an error (i.e. -1), then it
 283                  * has printed an error message in the buffer.  Therefore,
 284                  * we need to add a \n to the end of that message before we
 285                  * print our own message in the buffer.
 286                  */
 287                 if (version == -1)
 288                         buflen = strlen(devstat_errbuf);
 289                 else
 290                         buflen = 0;
 291
 292                 res = snprintf(devstat_errbuf + buflen,
 293                                DEVSTAT_ERRBUF_SIZE - buflen,
 294                                "%s%s: userland devstat version %d is not "
 295                                "the same as the kernel\n%s: devstat "
 296                                "version %d\n", version == -1 ? "\n" : "",
 297                                __func__, DEVSTAT_VERSION, __func__, version);
 298
 299                 if (res < 0)
 300                         devstat_errbuf[buflen] = '\0';
 301
 302                 buflen = strlen(devstat_errbuf);
 303                 if (version < DEVSTAT_VERSION)
 304                         res = snprintf(devstat_errbuf + buflen,
 305                                        DEVSTAT_ERRBUF_SIZE - buflen,
 306                                        "%s: libdevstat newer than kernel\n",
 307                                        __func__);
 308                 else
 309                         res = snprintf(devstat_errbuf + buflen,
 310                                        DEVSTAT_ERRBUF_SIZE - buflen,
 311                                        "%s: kernel newer than libdevstat\n",
 312                                        __func__);
 313
 314                 if (res < 0)
 315                         devstat_errbuf[buflen] = '\0';
 316
 317                 retval = -1;
 318         }
 319
 320         return(retval);
 321 }
 322
 323 /*
 324  * Get the current list of devices and statistics, and the current
 325  * generation number.
 326  *
 327  * Return values:
 328  * -1  -- error
 329  *  0  -- device list is unchanged
 330  *  1  -- device list has changed
 331  */
 332 int
 333 devstat_getdevs(kvm_t *kd, struct statinfo *stats)
 334 {
 335         int error;
 336         size_t dssize;
 337         long oldgeneration;
 338         int retval = 0;
 339         struct devinfo *dinfo;
 340         struct timespec ts;
 341
 342         dinfo = stats->dinfo;
 343
 344         if (dinfo == NULL) {
 345                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 346                          "%s: stats->dinfo was NULL", __func__);
 347                 return(-1);
 348         }
 349
 350         oldgeneration = dinfo->generation;
 351
 352         if (kd == NULL) {
 353                 clock_gettime(CLOCK_MONOTONIC, &ts);
 354                 stats->snap_time = ts.tv_sec + ts.tv_nsec * 1e-9;
 355
 356                 /* If this is our first time through, mem_ptr will be null. */
 357                 if (dinfo->mem_ptr == NULL) {
 358                         /*
 359                          * Get the number of devices.  If it's negative, it's an
 360                          * error.  Don't bother setting the error string, since
 361                          * getnumdevs() has already done that for us.
 362                          */
 363                         if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 364                                 return(-1);
 365
 366                         /*
 367                          * The kern.devstat.all sysctl returns the current
 368                          * generation number, as well as all the devices.
 369                          * So we need four bytes more.
 370                          */
 371                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 372                                  sizeof(long);
 373                         dinfo->mem_ptr = (u_int8_t *)malloc(dssize);
 374                         if (dinfo->mem_ptr == NULL) {
 375                                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 376                                          "%s: Cannot allocate memory for mem_ptr element",
 377                                          __func__);
 378                                 return(-1);
 379                         }
 380                 } else
 381                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 382                                  sizeof(long);
 383
 384                 /*
 385                  * Request all of the devices.  We only really allow for one
 386                  * ENOMEM failure.  It would, of course, be possible to just go
 387                  * in a loop and keep reallocing the device structure until we
 388                  * don't get ENOMEM back.  I'm not sure it's worth it, though.
 389                  * If devices are being added to the system that quickly, maybe
 390                  * the user can just wait until all devices are added.
 391                  */
 392                 for (;;) {
 393                         error = sysctlbyname("kern.devstat.all",
 394                                              dinfo->mem_ptr,
 395                                              &dssize, NULL, 0);
 396                         if (error != -1 || errno != EBUSY)
 397                                 break;
 398                 }
 399                 if (error == -1) {
 400                         /*
 401                          * If we get ENOMEM back, that means that there are
 402                          * more devices now, so we need to allocate more
 403                          * space for the device array.
 404                          */
 405                         if (errno == ENOMEM) {
 406                                 /*
 407                                  * No need to set the error string here,
 408                                  * devstat_getnumdevs() will do that if it fails.
 409                                  */
 410                                 if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 411                                         return(-1);
 412
 413                                 dssize = (dinfo->numdevs *
 414                                         sizeof(struct devstat)) + sizeof(long);
 415                                 dinfo->mem_ptr = (u_int8_t *)
 416                                         realloc(dinfo->mem_ptr, dssize);
 417                                 if ((error = sysctlbyname("kern.devstat.all",
 418                                     dinfo->mem_ptr, &dssize, NULL, 0)) == -1) {
 419                                         snprintf(devstat_errbuf,
 420                                                  sizeof(devstat_errbuf),
 421                                                  "%s: error getting device "
 422                                                  "stats\n%s: %s", __func__,
 423                                                  __func__, strerror(errno));
 424                                         return(-1);
 425                                 }
 426                         } else {
 427                                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 428                                          "%s: error getting device stats\n"
 429                                          "%s: %s", __func__, __func__,
 430                                          strerror(errno));
 431                                 return(-1);
 432                         }
 433                 }
 434
 435         } else {
 436                 if (KREADNL(kd, X_TIME_UPTIME, ts.tv_sec) == -1)
 437                         return(-1);
 438                 else
 439                         stats->snap_time = ts.tv_sec;
 440
 441                 /*
 442                  * This is of course non-atomic, but since we are working
 443                  * on a core dump, the generation is unlikely to change
 444                  */
 445                 if ((dinfo->numdevs = devstat_getnumdevs(kd)) == -1)
 446                         return(-1);
 447                 if ((dinfo->mem_ptr = (u_int8_t *)get_devstat_kvm(kd)) == NULL)
 448                         return(-1);
 449         }
 450         /*
 451          * The sysctl spits out the generation as the first four bytes,
 452          * then all of the device statistics structures.
 453          */
 454         dinfo->generation = *(long *)dinfo->mem_ptr;
 455
 456         /*
 457          * If the generation has changed, and if the current number of
 458          * devices is not the same as the number of devices recorded in the
 459          * devinfo structure, it is likely that the device list has shrunk.
 460          * The reason that it is likely that the device list has shrunk in
 461          * this case is that if the device list has grown, the sysctl above
 462          * will return an ENOMEM error, and we will reset the number of
 463          * devices and reallocate the device array.  If the second sysctl
 464          * fails, we will return an error and therefore never get to this
 465          * point.  If the device list has shrunk, the sysctl will not
 466          * return an error since we have more space allocated than is
 467          * necessary.  So, in the shrinkage case, we catch it here and
 468          * reallocate the array so that we don't use any more space than is
 469          * necessary.
 470          */
 471         if (oldgeneration != dinfo->generation) {
 472                 if (devstat_getnumdevs(kd) != dinfo->numdevs) {
 473                         if ((dinfo->numdevs = devstat_getnumdevs(kd)) < 0)
 474                                 return(-1);
 475                         dssize = (dinfo->numdevs * sizeof(struct devstat)) +
 476                                 sizeof(long);
 477                         dinfo->mem_ptr = (u_int8_t *)realloc(dinfo->mem_ptr,
 478                                                              dssize);
 479                 }
 480                 retval = 1;
 481         }
 482
 483         dinfo->devices = (struct devstat *)(dinfo->mem_ptr + sizeof(long));
 484
 485         return(retval);
 486 }
 487
 488 /*
 489  * selectdevs():
 490  *
 491  * Devices are selected/deselected based upon the following criteria:
 492  * - devices specified by the user on the command line
 493  * - devices matching any device type expressions given on the command line
 494  * - devices with the highest I/O, if 'top' mode is enabled
 495  * - the first n unselected devices in the device list, if maxshowdevs
 496  *   devices haven't already been selected and if the user has not
 497  *   specified any devices on the command line and if we're in "add" mode.
 498  *
 499  * Input parameters:
 500  * - device selection list (dev_select)
 501  * - current number of devices selected (num_selected)
 502  * - total number of devices in the selection list (num_selections)
 503  * - devstat generation as of the last time selectdevs() was called
 504  *   (select_generation)
 505  * - current devstat generation (current_generation)
 506  * - current list of devices and statistics (devices)
 507  * - number of devices in the current device list (numdevs)
 508  * - compiled version of the command line device type arguments (matches)
 509  *   - This is optional.  If the number of devices is 0, this will be ignored.
 510  *   - The matching code pays attention to the current selection mode.  So
 511  *     if you pass in a matching expression, it will be evaluated based
 512  *     upon the selection mode that is passed in.  See below for details.
 513  * - number of device type matching expressions (num_matches)
 514  *   - Set to 0 to disable the matching code.
 515  * - list of devices specified on the command line by the user (dev_selections)
 516  * - number of devices selected on the command line by the user
 517  *   (num_dev_selections)
 518  * - Our selection mode.  There are four different selection modes:
 519  *      - add mode.  (DS_SELECT_ADD) Any devices matching devices explicitly
 520  *        selected by the user or devices matching a pattern given by the
 521  *        user will be selected in addition to devices that are already
 522  *        selected.  Additional devices will be selected, up to maxshowdevs
 523  *        number of devices.
 524  *      - only mode. (DS_SELECT_ONLY)  Only devices matching devices
 525  *        explicitly given by the user or devices matching a pattern
 526  *        given by the user will be selected.  No other devices will be
 527  *        selected.
 528  *      - addonly mode.  (DS_SELECT_ADDONLY)  This is similar to add and
 529  *        only.  Basically, this will not de-select any devices that are
 530  *        current selected, as only mode would, but it will also not
 531  *        gratuitously select up to maxshowdevs devices as add mode would.
 532  *      - remove mode.  (DS_SELECT_REMOVE)  Any devices matching devices
 533  *        explicitly selected by the user or devices matching a pattern
 534  *        given by the user will be de-selected.
 535  * - maximum number of devices we can select (maxshowdevs)
 536  * - flag indicating whether or not we're in 'top' mode (perf_select)
 537  *
 538  * Output data:
 539  * - the device selection list may be modified and passed back out
 540  * - the number of devices selected and the total number of items in the
 541  *   device selection list may be changed
 542  * - the selection generation may be changed to match the current generation
 543  *
 544  * Return values:
 545  * -1  -- error
 546  *  0  -- selected devices are unchanged
 547  *  1  -- selected devices changed
 548  */
 549 int
 550 devstat_selectdevs(struct device_selection **dev_select, int *num_selected,
 551                    int *num_selections, long *select_generation,
 552                    long current_generation, struct devstat *devices,
 553                    int numdevs, struct devstat_match *matches, int num_matches,
 554                    char **dev_selections, int num_dev_selections,
 555                    devstat_select_mode select_mode, int maxshowdevs,
 556                    int perf_select)
 557 {
 558         int i, j, k;
 559         int init_selections = 0, init_selected_var = 0;
 560         struct device_selection *old_dev_select = NULL;
 561         int old_num_selections = 0, old_num_selected;
 562         int selection_number = 0;
 563         int changed = 0, found = 0;
 564
 565         if ((dev_select == NULL) || (devices == NULL) || (numdevs < 0))
 566                 return(-1);
 567
 568         /*
 569          * We always want to make sure that we have as many dev_select
 570          * entries as there are devices.
 571          */
 572         /*
 573          * In this case, we haven't selected devices before.
 574          */
 575         if (*dev_select == NULL) {
 576                 *dev_select = (struct device_selection *)malloc(numdevs *
 577                         sizeof(struct device_selection));
 578                 *select_generation = current_generation;
 579                 init_selections = 1;
 580                 changed = 1;
 581         /*
 582          * In this case, we have selected devices before, but the device
 583          * list has changed since we last selected devices, so we need to
 584          * either enlarge or reduce the size of the device selection list.
 585          * But delay the resizing until after copying the data to old_dev_select
 586          * as to not lose any data in the case of reducing the size.
 587          */
 588         } else if (*num_selections != numdevs) {
 589                 *select_generation = current_generation;
 590                 init_selections = 1;
 591         /*
 592          * In this case, we've selected devices before, and the selection
 593          * list is the same size as it was the last time, but the device
 594          * list has changed.
 595          */
 596         } else if (*select_generation < current_generation) {
 597                 *select_generation = current_generation;
 598                 init_selections = 1;
 599         }
 600
 601         if (*dev_select == NULL) {
 602                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 603                          "%s: Cannot (re)allocate memory for dev_select argument",
 604                          __func__);
 605                 return(-1);
 606         }
 607
 608         /*
 609          * If we're in "only" mode, we want to clear out the selected
 610          * variable since we're going to select exactly what the user wants
 611          * this time through.
 612          */
 613         if (select_mode == DS_SELECT_ONLY)
 614                 init_selected_var = 1;
 615
 616         /*
 617          * In all cases, we want to back up the number of selected devices.
 618          * It is a quick and accurate way to determine whether the selected
 619          * devices have changed.
 620          */
 621         old_num_selected = *num_selected;
 622
 623         /*
 624          * We want to make a backup of the current selection list if
 625          * the list of devices has changed, or if we're in performance
 626          * selection mode.  In both cases, we don't want to make a backup
 627          * if we already know for sure that the list will be different.
 628          * This is certainly the case if this is our first time through the
 629          * selection code.
 630          */
 631         if (((init_selected_var != 0) || (init_selections != 0)
 632          || (perf_select != 0)) && (changed == 0)){
 633                 old_dev_select = (struct device_selection *)malloc(
 634                     *num_selections * sizeof(struct device_selection));
 635                 if (old_dev_select == NULL) {
 636                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
 637                                  "%s: Cannot allocate memory for selection list backup",
 638                                  __func__);
 639                         return(-1);
 640                 }
 641                 old_num_selections = *num_selections;
 642                 bcopy(*dev_select, old_dev_select,
 643                     sizeof(struct device_selection) * *num_selections);
 644         }
 645
 646         if (!changed && *num_selections != numdevs) {
 647                 *dev_select = (struct device_selection *)reallocf(*dev_select,
 648                         numdevs * sizeof(struct device_selection));
 649         }
 650
 651         if (init_selections != 0) {
 652                 bzero(*dev_select, sizeof(struct device_selection) * numdevs);
 653
 654                 for (i = 0; i < numdevs; i++) {
 655                         (*dev_select)[i].device_number =
 656                                 devices[i].device_number;
 657                         strncpy((*dev_select)[i].device_name,
 658                                 devices[i].device_name,
 659                                 DEVSTAT_NAME_LEN);
 660                         (*dev_select)[i].device_name[DEVSTAT_NAME_LEN - 1]='\0';
 661                         (*dev_select)[i].unit_number = devices[i].unit_number;
 662                         (*dev_select)[i].position = i;
 663                 }
 664                 *num_selections = numdevs;
 665         } else if (init_selected_var != 0) {
 666                 for (i = 0; i < numdevs; i++)
 667                         (*dev_select)[i].selected = 0;
 668         }
 669
 670         /* we haven't gotten around to selecting anything yet.. */
 671         if ((select_mode == DS_SELECT_ONLY) || (init_selections != 0)
 672          || (init_selected_var != 0))
 673                 *num_selected = 0;
 674
 675         /*
 676          * Look through any devices the user specified on the command line
 677          * and see if they match known devices.  If so, select them.
 678          */
 679         for (i = 0; (i < *num_selections) && (num_dev_selections > 0); i++) {
 680                 char tmpstr[80];
 681
 682                 snprintf(tmpstr, sizeof(tmpstr), "%s%d",
 683                          (*dev_select)[i].device_name,
 684                          (*dev_select)[i].unit_number);
 685                 for (j = 0; j < num_dev_selections; j++) {
 686                         if (strcmp(tmpstr, dev_selections[j]) == 0) {
 687                                 /*
 688                                  * Here we do different things based on the
 689                                  * mode we're in.  If we're in add or
 690                                  * addonly mode, we only select this device
 691                                  * if it hasn't already been selected.
 692                                  * Otherwise, we would be unnecessarily
 693                                  * changing the selection order and
 694                                  * incrementing the selection count.  If
 695                                  * we're in only mode, we unconditionally
 696                                  * select this device, since in only mode
 697                                  * any previous selections are erased and
 698                                  * manually specified devices are the first
 699                                  * ones to be selected.  If we're in remove
 700                                  * mode, we de-select the specified device and
 701                                  * decrement the selection count.
 702                                  */
 703                                 switch(select_mode) {
 704                                 case DS_SELECT_ADD:
 705                                 case DS_SELECT_ADDONLY:
 706                                         if ((*dev_select)[i].selected)
 707                                                 break;
 708                                         /* FALLTHROUGH */
 709                                 case DS_SELECT_ONLY:
 710                                         (*dev_select)[i].selected =
 711                                                 ++selection_number;
 712                                         (*num_selected)++;
 713                                         break;
 714                                 case DS_SELECT_REMOVE:
 715                                         (*dev_select)[i].selected = 0;
 716                                         (*num_selected)--;
 717                                         /*
 718                                          * This isn't passed back out, we
 719                                          * just use it to keep track of
 720                                          * how many devices we've removed.
 721                                          */
 722                                         num_dev_selections--;
 723                                         break;
 724                                 }
 725                                 break;
 726                         }
 727                 }
 728         }
 729
 730         /*
 731          * Go through the user's device type expressions and select devices
 732          * accordingly.  We only do this if the number of devices already
 733          * selected is less than the maximum number we can show.
 734          */
 735         for (i = 0; (i < num_matches) && (*num_selected < maxshowdevs); i++) {
 736                 /* We should probably indicate some error here */
 737                 if ((matches[i].match_fields == DEVSTAT_MATCH_NONE)
 738                  || (matches[i].num_match_categories <= 0))
 739                         continue;
 740
 741                 for (j = 0; j < numdevs; j++) {
 742                         int num_match_categories;
 743
 744                         num_match_categories = matches[i].num_match_categories;
 745
 746                         /*
 747                          * Determine whether or not the current device
 748                          * matches the given matching expression.  This if
 749                          * statement consists of three components:
 750                          *   - the device type check
 751                          *   - the device interface check
 752                          *   - the passthrough check
 753                          * If a the matching test is successful, it
 754                          * decrements the number of matching categories,
 755                          * and if we've reached the last element that
 756                          * needed to be matched, the if statement succeeds.
 757                          *
 758                          */
 759                         if ((((matches[i].match_fields & DEVSTAT_MATCH_TYPE)!=0)
 760                           && ((devices[j].device_type & DEVSTAT_TYPE_MASK) ==
 761                                 (matches[i].device_type & DEVSTAT_TYPE_MASK))
 762                           &&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 763                            || (((matches[i].match_fields &
 764                                 DEVSTAT_MATCH_PASS) == 0)
 765                             && ((devices[j].device_type &
 766                                 DEVSTAT_TYPE_PASS) == 0)))
 767                           && (--num_match_categories == 0))
 768                          || (((matches[i].match_fields & DEVSTAT_MATCH_IF) != 0)
 769                           && ((devices[j].device_type & DEVSTAT_TYPE_IF_MASK) ==
 770                                 (matches[i].device_type & DEVSTAT_TYPE_IF_MASK))
 771                           &&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 772                            || (((matches[i].match_fields &
 773                                 DEVSTAT_MATCH_PASS) == 0)
 774                             && ((devices[j].device_type &
 775                                 DEVSTAT_TYPE_PASS) == 0)))
 776                           && (--num_match_categories == 0))
 777                          || (((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
 778                           && ((devices[j].device_type & DEVSTAT_TYPE_PASS) != 0)
 779                           && (--num_match_categories == 0))) {
 780
 781                                 /*
 782                                  * This is probably a non-optimal solution
 783                                  * to the problem that the devices in the
 784                                  * device list will not be in the same
 785                                  * order as the devices in the selection
 786                                  * array.
 787                                  */
 788                                 for (k = 0; k < numdevs; k++) {
 789                                         if ((*dev_select)[k].position == j) {
 790                                                 found = 1;
 791                                                 break;
 792                                         }
 793                                 }
 794
 795                                 /*
 796                                  * There shouldn't be a case where a device
 797                                  * in the device list is not in the
 798                                  * selection list...but it could happen.
 799                                  */
 800                                 if (found != 1) {
 801                                         fprintf(stderr, "selectdevs: couldn't"
 802                                                 " find %s%d in selection "
 803                                                 "list\n",
 804                                                 devices[j].device_name,
 805                                                 devices[j].unit_number);
 806                                         break;
 807                                 }
 808
 809                                 /*
 810                                  * We do different things based upon the
 811                                  * mode we're in.  If we're in add or only
 812                                  * mode, we go ahead and select this device
 813                                  * if it hasn't already been selected.  If
 814                                  * it has already been selected, we leave
 815                                  * it alone so we don't mess up the
 816                                  * selection ordering.  Manually specified
 817                                  * devices have already been selected, and
 818                                  * they have higher priority than pattern
 819                                  * matched devices.  If we're in remove
 820                                  * mode, we de-select the given device and
 821                                  * decrement the selected count.
 822                                  */
 823                                 switch(select_mode) {
 824                                 case DS_SELECT_ADD:
 825                                 case DS_SELECT_ADDONLY:
 826                                 case DS_SELECT_ONLY:
 827                                         if ((*dev_select)[k].selected != 0)
 828                                                 break;
 829                                         (*dev_select)[k].selected =
 830                                                 ++selection_number;
 831                                         (*num_selected)++;
 832                                         break;
 833                                 case DS_SELECT_REMOVE:
 834                                         (*dev_select)[k].selected = 0;
 835                                         (*num_selected)--;
 836                                         break;
 837                                 }
 838                         }
 839                 }
 840         }
 841
 842         /*
 843          * Here we implement "top" mode.  Devices are sorted in the
 844          * selection array based on two criteria:  whether or not they are
 845          * selected (not selection number, just the fact that they are
 846          * selected!) and the number of bytes in the "bytes" field of the
 847          * selection structure.  The bytes field generally must be kept up
 848          * by the user.  In the future, it may be maintained by library
 849          * functions, but for now the user has to do the work.
 850          *
 851          * At first glance, it may seem wrong that we don't go through and
 852          * select every device in the case where the user hasn't specified
 853          * any devices or patterns.  In fact, though, it won't make any
 854          * difference in the device sorting.  In that particular case (i.e.
 855          * when we're in "add" or "only" mode, and the user hasn't
 856          * specified anything) the first time through no devices will be
 857          * selected, so the only criterion used to sort them will be their
 858          * performance.  The second time through, and every time thereafter,
 859          * all devices will be selected, so again selection won't matter.
 860          */
 861         if (perf_select != 0) {
 862
 863                 /* Sort the device array by throughput  */
 864                 qsort(*dev_select, *num_selections,
 865                       sizeof(struct device_selection),
 866                       compare_select);
 867
 868                 if (*num_selected == 0) {
 869                         /*
 870                          * Here we select every device in the array, if it
 871                          * isn't already selected.  Because the 'selected'
 872                          * variable in the selection array entries contains
 873                          * the selection order, the devstats routine can show
 874                          * the devices that were selected first.
 875                          */
 876                         for (i = 0; i < *num_selections; i++) {
 877                                 if ((*dev_select)[i].selected == 0) {
 878                                         (*dev_select)[i].selected =
 879                                                 ++selection_number;
 880                                         (*num_selected)++;
 881                                 }
 882                         }
 883                 } else {
 884                         selection_number = 0;
 885                         for (i = 0; i < *num_selections; i++) {
 886                                 if ((*dev_select)[i].selected != 0) {
 887                                         (*dev_select)[i].selected =
 888                                                 ++selection_number;
 889                                 }
 890                         }
 891                 }
 892         }
 893
 894         /*
 895          * If we're in the "add" selection mode and if we haven't already
 896          * selected maxshowdevs number of devices, go through the array and
 897          * select any unselected devices.  If we're in "only" mode, we
 898          * obviously don't want to select anything other than what the user
 899          * specifies.  If we're in "remove" mode, it probably isn't a good
 900          * idea to go through and select any more devices, since we might
 901          * end up selecting something that the user wants removed.  Through
 902          * more complicated logic, we could actually figure this out, but
 903          * that would probably require combining this loop with the various
 904          * selections loops above.
 905          */
 906         if ((select_mode == DS_SELECT_ADD) && (*num_selected < maxshowdevs)) {
 907                 for (i = 0; i < *num_selections; i++)
 908                         if ((*dev_select)[i].selected == 0) {
 909                                 (*dev_select)[i].selected = ++selection_number;
 910                                 (*num_selected)++;
 911                         }
 912         }
 913
 914         /*
 915          * Look at the number of devices that have been selected.  If it
 916          * has changed, set the changed variable.  Otherwise, if we've
 917          * made a backup of the selection list, compare it to the current
 918          * selection list to see if the selected devices have changed.
 919          */
 920         if ((changed == 0) && (old_num_selected != *num_selected))
 921                 changed = 1;
 922         else if ((changed == 0) && (old_dev_select != NULL)) {
 923                 /*
 924                  * Now we go through the selection list and we look at
 925                  * it three different ways.
 926                  */
 927                 for (i = 0; (i < *num_selections) && (changed == 0) &&
 928                      (i < old_num_selections); i++) {
 929                         /*
 930                          * If the device at index i in both the new and old
 931                          * selection arrays has the same device number and
 932                          * selection status, it hasn't changed.  We
 933                          * continue on to the next index.
 934                          */
 935                         if (((*dev_select)[i].device_number ==
 936                              old_dev_select[i].device_number)
 937                          && ((*dev_select)[i].selected ==
 938                              old_dev_select[i].selected))
 939                                 continue;
 940
 941                         /*
 942                          * Now, if we're still going through the if
 943                          * statement, the above test wasn't true.  So we
 944                          * check here to see if the device at index i in
 945                          * the current array is the same as the device at
 946                          * index i in the old array.  If it is, that means
 947                          * that its selection number has changed.  Set
 948                          * changed to 1 and exit the loop.
 949                          */
 950                         else if ((*dev_select)[i].device_number ==
 951                                   old_dev_select[i].device_number) {
 952                                 changed = 1;
 953                                 break;
 954                         }
 955                         /*
 956                          * If we get here, then the device at index i in
 957                          * the current array isn't the same device as the
 958                          * device at index i in the old array.
 959                          */
 960                         else {
 961                                 found = 0;
 962
 963                                 /*
 964                                  * Search through the old selection array
 965                                  * looking for a device with the same
 966                                  * device number as the device at index i
 967                                  * in the current array.  If the selection
 968                                  * status is the same, then we mark it as
 969                                  * found.  If the selection status isn't
 970                                  * the same, we break out of the loop.
 971                                  * Since found isn't set, changed will be
 972                                  * set to 1 below.
 973                                  */
 974                                 for (j = 0; j < old_num_selections; j++) {
 975                                         if (((*dev_select)[i].device_number ==
 976                                               old_dev_select[j].device_number)
 977                                          && ((*dev_select)[i].selected ==
 978                                               old_dev_select[j].selected)){
 979                                                 found = 1;
 980                                                 break;
 981                                         }
 982                                         else if ((*dev_select)[i].device_number
 983                                             == old_dev_select[j].device_number)
 984                                                 break;
 985                                 }
 986                                 if (found == 0)
 987                                         changed = 1;
 988                         }
 989                 }
 990         }
 991         if (old_dev_select != NULL)
 992                 free(old_dev_select);
 993
 994         return(changed);
 995 }
 996
 997 /*
 998  * Comparison routine for qsort() above.  Note that the comparison here is
 999  * backwards -- generally, it should return a value to indicate whether
1000  * arg1 is <, =, or > arg2.  Instead, it returns the opposite.  The reason
1001  * it returns the opposite is so that the selection array will be sorted in
1002  * order of decreasing performance.  We sort on two parameters.  The first
1003  * sort key is whether or not one or the other of the devices in question
1004  * has been selected.  If one of them has, and the other one has not, the
1005  * selected device is automatically more important than the unselected
1006  * device.  If neither device is selected, we judge the devices based upon
1007  * performance.
1008  */
1009 static int
1010 compare_select(const void *arg1, const void *arg2)
1011 {
1012         if ((((const struct device_selection *)arg1)->selected)
1013          && (((const struct device_selection *)arg2)->selected == 0))
1014                 return(-1);
1015         else if ((((const struct device_selection *)arg1)->selected == 0)
1016               && (((const struct device_selection *)arg2)->selected))
1017                 return(1);
1018         else if (((const struct device_selection *)arg2)->bytes <
1019                  ((const struct device_selection *)arg1)->bytes)
1020                 return(-1);
1021         else if (((const struct device_selection *)arg2)->bytes >
1022                  ((const struct device_selection *)arg1)->bytes)
1023                 return(1);
1024         else
1025                 return(0);
1026 }
1027
1028 /*
1029  * Take a string with the general format "arg1,arg2,arg3", and build a
1030  * device matching expression from it.
1031  */
1032 int
1033 devstat_buildmatch(char *match_str, struct devstat_match **matches,
1034                    int *num_matches)
1035 {
1036         char *tstr[5];
1037         char **tempstr;
1038         int num_args;
1039         int i, j;
1040
1041         /* We can't do much without a string to parse */
1042         if (match_str == NULL) {
1043                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1044                          "%s: no match expression", __func__);
1045                 return(-1);
1046         }
1047
1048         /*
1049          * Break the (comma delimited) input string out into separate strings.
1050          */
1051         for (tempstr = tstr, num_args  = 0;
1052              (*tempstr = strsep(&match_str, ",")) != NULL && (num_args < 5);)
1053                 if (**tempstr != '\0') {
1054                         num_args++;
1055                         if (++tempstr >= &tstr[5])
1056                                 break;
1057                 }
1058
1059         /* The user gave us too many type arguments */
1060         if (num_args > 3) {
1061                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1062                          "%s: too many type arguments", __func__);
1063                 return(-1);
1064         }
1065
1066         if (*num_matches == 0)
1067                 *matches = NULL;
1068
1069         *matches = (struct devstat_match *)reallocf(*matches,
1070                   sizeof(struct devstat_match) * (*num_matches + 1));
1071
1072         if (*matches == NULL) {
1073                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1074                          "%s: Cannot allocate memory for matches list", __func__);
1075                 return(-1);
1076         }
1077
1078         /* Make sure the current entry is clear */
1079         bzero(&matches[0][*num_matches], sizeof(struct devstat_match));
1080
1081         /*
1082          * Step through the arguments the user gave us and build a device
1083          * matching expression from them.
1084          */
1085         for (i = 0; i < num_args; i++) {
1086                 char *tempstr2, *tempstr3;
1087
1088                 /*
1089                  * Get rid of leading white space.
1090                  */
1091                 tempstr2 = tstr[i];
1092                 while (isspace(*tempstr2) && (*tempstr2 != '\0'))
1093                         tempstr2++;
1094
1095                 /*
1096                  * Get rid of trailing white space.
1097                  */
1098                 tempstr3 = &tempstr2[strlen(tempstr2) - 1];
1099
1100                 while ((*tempstr3 != '\0') && (tempstr3 > tempstr2)
1101                     && (isspace(*tempstr3))) {
1102                         *tempstr3 = '\0';
1103                         tempstr3--;
1104                 }
1105
1106                 /*
1107                  * Go through the match table comparing the user's
1108                  * arguments to known device types, interfaces, etc.
1109                  */
1110                 for (j = 0; match_table[j].match_str != NULL; j++) {
1111                         /*
1112                          * We do case-insensitive matching, in case someone
1113                          * wants to enter "SCSI" instead of "scsi" or
1114                          * something like that.  Only compare as many
1115                          * characters as are in the string in the match
1116                          * table.  This should help if someone tries to use
1117                          * a super-long match expression.
1118                          */
1119                         if (strncasecmp(tempstr2, match_table[j].match_str,
1120                             strlen(match_table[j].match_str)) == 0) {
1121                                 /*
1122                                  * Make sure the user hasn't specified two
1123                                  * items of the same type, like "da" and
1124                                  * "cd".  One device cannot be both.
1125                                  */
1126                                 if (((*matches)[*num_matches].match_fields &
1127                                     match_table[j].match_field) != 0) {
1128                                         snprintf(devstat_errbuf,
1129                                                  sizeof(devstat_errbuf),
1130                                                  "%s: cannot have more than "
1131                                                  "one match item in a single "
1132                                                  "category", __func__);
1133                                         return(-1);
1134                                 }
1135                                 /*
1136                                  * If we've gotten this far, we have a
1137                                  * winner.  Set the appropriate fields in
1138                                  * the match entry.
1139                                  */
1140                                 (*matches)[*num_matches].match_fields |=
1141                                         match_table[j].match_field;
1142                                 (*matches)[*num_matches].device_type |=
1143                                         match_table[j].type;
1144                                 (*matches)[*num_matches].num_match_categories++;
1145                                 break;
1146                         }
1147                 }
1148                 /*
1149                  * We should have found a match in the above for loop.  If
1150                  * not, that means the user entered an invalid device type
1151                  * or interface.
1152                  */
1153                 if ((*matches)[*num_matches].num_match_categories != (i + 1)) {
1154                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1155                                  "%s: unknown match item \"%s\"", __func__,
1156                                  tstr[i]);
1157                         return(-1);
1158                 }
1159         }
1160
1161         (*num_matches)++;
1162
1163         return(0);
1164 }
1165
1166 /*
1167  * Compute a number of device statistics.  Only one field is mandatory, and
1168  * that is "current".  Everything else is optional.  The caller passes in
1169  * pointers to variables to hold the various statistics he desires.  If he
1170  * doesn't want a particular staistic, he should pass in a NULL pointer.
1171  * Return values:
1172  * 0   -- success
1173  * -1  -- failure
1174  */
1175 int
1176 compute_stats(struct devstat *current, struct devstat *previous,
1177               long double etime, u_int64_t *total_bytes,
1178               u_int64_t *total_transfers, u_int64_t *total_blocks,
1179               long double *kb_per_transfer, long double *transfers_per_second,
1180               long double *mb_per_second, long double *blocks_per_second,
1181               long double *ms_per_transaction)
1182 {
1183         return(devstat_compute_statistics(current, previous, etime,
1184                total_bytes ? DSM_TOTAL_BYTES : DSM_SKIP,
1185                total_bytes,
1186                total_transfers ? DSM_TOTAL_TRANSFERS : DSM_SKIP,
1187                total_transfers,
1188                total_blocks ? DSM_TOTAL_BLOCKS : DSM_SKIP,
1189                total_blocks,
1190                kb_per_transfer ? DSM_KB_PER_TRANSFER : DSM_SKIP,
1191                kb_per_transfer,
1192                transfers_per_second ? DSM_TRANSFERS_PER_SECOND : DSM_SKIP,
1193                transfers_per_second,
1194                mb_per_second ? DSM_MB_PER_SECOND : DSM_SKIP,
1195                mb_per_second,
1196                blocks_per_second ? DSM_BLOCKS_PER_SECOND : DSM_SKIP,
1197                blocks_per_second,
1198                ms_per_transaction ? DSM_MS_PER_TRANSACTION : DSM_SKIP,
1199                ms_per_transaction,
1200                DSM_NONE));
1201 }
1202
1203
1204 /* This is 1/2^64 */
1205 #define BINTIME_SCALE 5.42101086242752217003726400434970855712890625e-20
1206
1207 long double
1208 devstat_compute_etime(struct bintime *cur_time, struct bintime *prev_time)
1209 {
1210         long double etime;
1211
1212         etime = cur_time->sec;
1213         etime += cur_time->frac * BINTIME_SCALE;
1214         if (prev_time != NULL) {
1215                 etime -= prev_time->sec;
1216                 etime -= prev_time->frac * BINTIME_SCALE;
1217         }
1218         return(etime);
1219 }
1220
1221 #define DELTA(field, index)                             \
1222         (current->field[(index)] - (previous ? previous->field[(index)] : 0))
1223
1224 #define DELTA_T(field)                                  \
1225         devstat_compute_etime(&current->field,          \
1226         (previous ? &previous->field : NULL))
1227
1228 int
1229 devstat_compute_statistics(struct devstat *current, struct devstat *previous,
1230                            long double etime, ...)
1231 {
1232         u_int64_t totalbytes, totalbytesread, totalbyteswrite, totalbytesfree;
1233         u_int64_t totaltransfers, totaltransfersread, totaltransferswrite;
1234         u_int64_t totaltransfersother, totalblocks, totalblocksread;
1235         u_int64_t totalblockswrite, totaltransfersfree, totalblocksfree;
1236         long double totalduration, totaldurationread, totaldurationwrite;
1237         long double totaldurationfree, totaldurationother;
1238         va_list ap;
1239         devstat_metric metric;
1240         u_int64_t *destu64;
1241         long double *destld;
1242         int retval;
1243
1244         retval = 0;
1245
1246         /*
1247          * current is the only mandatory field.
1248          */
1249         if (current == NULL) {
1250                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1251                          "%s: current stats structure was NULL", __func__);
1252                 return(-1);
1253         }
1254
1255         totalbytesread = DELTA(bytes, DEVSTAT_READ);
1256         totalbyteswrite = DELTA(bytes, DEVSTAT_WRITE);
1257         totalbytesfree = DELTA(bytes, DEVSTAT_FREE);
1258         totalbytes = totalbytesread + totalbyteswrite + totalbytesfree;
1259
1260         totaltransfersread = DELTA(operations, DEVSTAT_READ);
1261         totaltransferswrite = DELTA(operations, DEVSTAT_WRITE);
1262         totaltransfersother = DELTA(operations, DEVSTAT_NO_DATA);
1263         totaltransfersfree = DELTA(operations, DEVSTAT_FREE);
1264         totaltransfers = totaltransfersread + totaltransferswrite +
1265                          totaltransfersother + totaltransfersfree;
1266
1267         totalblocks = totalbytes;
1268         totalblocksread = totalbytesread;
1269         totalblockswrite = totalbyteswrite;
1270         totalblocksfree = totalbytesfree;
1271
1272         if (current->block_size > 0) {
1273                 totalblocks /= current->block_size;
1274                 totalblocksread /= current->block_size;
1275                 totalblockswrite /= current->block_size;
1276                 totalblocksfree /= current->block_size;
1277         } else {
1278                 totalblocks /= 512;
1279                 totalblocksread /= 512;
1280                 totalblockswrite /= 512;
1281                 totalblocksfree /= 512;
1282         }
1283
1284         totaldurationread = DELTA_T(duration[DEVSTAT_READ]);
1285         totaldurationwrite = DELTA_T(duration[DEVSTAT_WRITE]);
1286         totaldurationfree = DELTA_T(duration[DEVSTAT_FREE]);
1287         totaldurationother = DELTA_T(duration[DEVSTAT_NO_DATA]);
1288         totalduration = totaldurationread + totaldurationwrite +
1289             totaldurationfree + totaldurationother;
1290
1291         va_start(ap, etime);
1292
1293         while ((metric = (devstat_metric)va_arg(ap, devstat_metric)) != 0) {
1294
1295                 if (metric == DSM_NONE)
1296                         break;
1297
1298                 if (metric >= DSM_MAX) {
1299                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1300                                  "%s: metric %d is out of range", __func__,
1301                                  metric);
1302                         retval = -1;
1303                         goto bailout;
1304                 }
1305
1306                 switch (devstat_arg_list[metric].argtype) {
1307                 case DEVSTAT_ARG_UINT64:
1308                         destu64 = (u_int64_t *)va_arg(ap, u_int64_t *);
1309                         break;
1310                 case DEVSTAT_ARG_LD:
1311                         destld = (long double *)va_arg(ap, long double *);
1312                         break;
1313                 case DEVSTAT_ARG_SKIP:
1314                         destld = (long double *)va_arg(ap, long double *);
1315                         break;
1316                 default:
1317                         retval = -1;
1318                         goto bailout;
1319                         break; /* NOTREACHED */
1320                 }
1321
1322                 if (devstat_arg_list[metric].argtype == DEVSTAT_ARG_SKIP)
1323                         continue;
1324
1325                 switch (metric) {
1326                 case DSM_TOTAL_BYTES:
1327                         *destu64 = totalbytes;
1328                         break;
1329                 case DSM_TOTAL_BYTES_READ:
1330                         *destu64 = totalbytesread;
1331                         break;
1332                 case DSM_TOTAL_BYTES_WRITE:
1333                         *destu64 = totalbyteswrite;
1334                         break;
1335                 case DSM_TOTAL_BYTES_FREE:
1336                         *destu64 = totalbytesfree;
1337                         break;
1338                 case DSM_TOTAL_TRANSFERS:
1339                         *destu64 = totaltransfers;
1340                         break;
1341                 case DSM_TOTAL_TRANSFERS_READ:
1342                         *destu64 = totaltransfersread;
1343                         break;
1344                 case DSM_TOTAL_TRANSFERS_WRITE:
1345                         *destu64 = totaltransferswrite;
1346                         break;
1347                 case DSM_TOTAL_TRANSFERS_FREE:
1348                         *destu64 = totaltransfersfree;
1349                         break;
1350                 case DSM_TOTAL_TRANSFERS_OTHER:
1351                         *destu64 = totaltransfersother;
1352                         break;
1353                 case DSM_TOTAL_BLOCKS:
1354                         *destu64 = totalblocks;
1355                         break;
1356                 case DSM_TOTAL_BLOCKS_READ:
1357                         *destu64 = totalblocksread;
1358                         break;
1359                 case DSM_TOTAL_BLOCKS_WRITE:
1360                         *destu64 = totalblockswrite;
1361                         break;
1362                 case DSM_TOTAL_BLOCKS_FREE:
1363                         *destu64 = totalblocksfree;
1364                         break;
1365                 case DSM_KB_PER_TRANSFER:
1366                         *destld = totalbytes;
1367                         *destld /= 1024;
1368                         if (totaltransfers > 0)
1369                                 *destld /= totaltransfers;
1370                         else
1371                                 *destld = 0.0;
1372                         break;
1373                 case DSM_KB_PER_TRANSFER_READ:
1374                         *destld = totalbytesread;
1375                         *destld /= 1024;
1376                         if (totaltransfersread > 0)
1377                                 *destld /= totaltransfersread;
1378                         else
1379                                 *destld = 0.0;
1380                         break;
1381                 case DSM_KB_PER_TRANSFER_WRITE:
1382                         *destld = totalbyteswrite;
1383                         *destld /= 1024;
1384                         if (totaltransferswrite > 0)
1385                                 *destld /= totaltransferswrite;
1386                         else
1387                                 *destld = 0.0;
1388                         break;
1389                 case DSM_KB_PER_TRANSFER_FREE:
1390                         *destld = totalbytesfree;
1391                         *destld /= 1024;
1392                         if (totaltransfersfree > 0)
1393                                 *destld /= totaltransfersfree;
1394                         else
1395                                 *destld = 0.0;
1396                         break;
1397                 case DSM_TRANSFERS_PER_SECOND:
1398                         if (etime > 0.0) {
1399                                 *destld = totaltransfers;
1400                                 *destld /= etime;
1401                         } else
1402                                 *destld = 0.0;
1403                         break;
1404                 case DSM_TRANSFERS_PER_SECOND_READ:
1405                         if (etime > 0.0) {
1406                                 *destld = totaltransfersread;
1407                                 *destld /= etime;
1408                         } else
1409                                 *destld = 0.0;
1410                         break;
1411                 case DSM_TRANSFERS_PER_SECOND_WRITE:
1412                         if (etime > 0.0) {
1413                                 *destld = totaltransferswrite;
1414                                 *destld /= etime;
1415                         } else
1416                                 *destld = 0.0;
1417                         break;
1418                 case DSM_TRANSFERS_PER_SECOND_FREE:
1419                         if (etime > 0.0) {
1420                                 *destld = totaltransfersfree;
1421                                 *destld /= etime;
1422                         } else
1423                                 *destld = 0.0;
1424                         break;
1425                 case DSM_TRANSFERS_PER_SECOND_OTHER:
1426                         if (etime > 0.0) {
1427                                 *destld = totaltransfersother;
1428                                 *destld /= etime;
1429                         } else
1430                                 *destld = 0.0;
1431                         break;
1432                 case DSM_MB_PER_SECOND:
1433                         *destld = totalbytes;
1434                         *destld /= 1024 * 1024;
1435                         if (etime > 0.0)
1436                                 *destld /= etime;
1437                         else
1438                                 *destld = 0.0;
1439                         break;
1440                 case DSM_MB_PER_SECOND_READ:
1441                         *destld = totalbytesread;
1442                         *destld /= 1024 * 1024;
1443                         if (etime > 0.0)
1444                                 *destld /= etime;
1445                         else
1446                                 *destld = 0.0;
1447                         break;
1448                 case DSM_MB_PER_SECOND_WRITE:
1449                         *destld = totalbyteswrite;
1450                         *destld /= 1024 * 1024;
1451                         if (etime > 0.0)
1452                                 *destld /= etime;
1453                         else
1454                                 *destld = 0.0;
1455                         break;
1456                 case DSM_MB_PER_SECOND_FREE:
1457                         *destld = totalbytesfree;
1458                         *destld /= 1024 * 1024;
1459                         if (etime > 0.0)
1460                                 *destld /= etime;
1461                         else
1462                                 *destld = 0.0;
1463                         break;
1464                 case DSM_BLOCKS_PER_SECOND:
1465                         *destld = totalblocks;
1466                         if (etime > 0.0)
1467                                 *destld /= etime;
1468                         else
1469                                 *destld = 0.0;
1470                         break;
1471                 case DSM_BLOCKS_PER_SECOND_READ:
1472                         *destld = totalblocksread;
1473                         if (etime > 0.0)
1474                                 *destld /= etime;
1475                         else
1476                                 *destld = 0.0;
1477                         break;
1478                 case DSM_BLOCKS_PER_SECOND_WRITE:
1479                         *destld = totalblockswrite;
1480                         if (etime > 0.0)
1481                                 *destld /= etime;
1482                         else
1483                                 *destld = 0.0;
1484                         break;
1485                 case DSM_BLOCKS_PER_SECOND_FREE:
1486                         *destld = totalblocksfree;
1487                         if (etime > 0.0)
1488                                 *destld /= etime;
1489                         else
1490                                 *destld = 0.0;
1491                         break;
1492                 /*
1493                  * Some devstat callers update the duration and some don't.
1494                  * So this will only be accurate if they provide the
1495                  * duration.
1496                  */
1497                 case DSM_MS_PER_TRANSACTION:
1498                         if (totaltransfers > 0) {
1499                                 *destld = totalduration;
1500                                 *destld /= totaltransfers;
1501                                 *destld *= 1000;
1502                         } else
1503                                 *destld = 0.0;
1504                         break;
1505                 case DSM_MS_PER_TRANSACTION_READ:
1506                         if (totaltransfersread > 0) {
1507                                 *destld = totaldurationread;
1508                                 *destld /= totaltransfersread;
1509                                 *destld *= 1000;
1510                         } else
1511                                 *destld = 0.0;
1512                         break;
1513                 case DSM_MS_PER_TRANSACTION_WRITE:
1514                         if (totaltransferswrite > 0) {
1515                                 *destld = totaldurationwrite;
1516                                 *destld /= totaltransferswrite;
1517                                 *destld *= 1000;
1518                         } else
1519                                 *destld = 0.0;
1520                         break;
1521                 case DSM_MS_PER_TRANSACTION_FREE:
1522                         if (totaltransfersfree > 0) {
1523                                 *destld = totaldurationfree;
1524                                 *destld /= totaltransfersfree;
1525                                 *destld *= 1000;
1526                         } else
1527                                 *destld = 0.0;
1528                         break;
1529                 case DSM_MS_PER_TRANSACTION_OTHER:
1530                         if (totaltransfersother > 0) {
1531                                 *destld = totaldurationother;
1532                                 *destld /= totaltransfersother;
1533                                 *destld *= 1000;
1534                         } else
1535                                 *destld = 0.0;
1536                         break;
1537                 case DSM_BUSY_PCT:
1538                         *destld = DELTA_T(busy_time);
1539                         if (*destld < 0)
1540                                 *destld = 0;
1541                         *destld /= etime;
1542                         *destld *= 100;
1543                         if (*destld < 0)
1544                                 *destld = 0;
1545                         break;
1546                 case DSM_QUEUE_LENGTH:
1547                         *destu64 = current->start_count - current->end_count;
1548                         break;
1549                 case DSM_TOTAL_DURATION:
1550                         *destld = totalduration;
1551                         break;
1552                 case DSM_TOTAL_DURATION_READ:
1553                         *destld = totaldurationread;
1554                         break;
1555                 case DSM_TOTAL_DURATION_WRITE:
1556                         *destld = totaldurationwrite;
1557                         break;
1558                 case DSM_TOTAL_DURATION_FREE:
1559                         *destld = totaldurationfree;
1560                         break;
1561                 case DSM_TOTAL_DURATION_OTHER:
1562                         *destld = totaldurationother;
1563                         break;
1564                 case DSM_TOTAL_BUSY_TIME:
1565                         *destld = DELTA_T(busy_time);
1566                         break;
1567 /*
1568  * XXX: comment out the default block to see if any case's are missing.
1569  */
1570 #if 1
1571                 default:
1572                         /*
1573                          * This shouldn't happen, since we should have
1574                          * caught any out of range metrics at the top of
1575                          * the loop.
1576                          */
1577                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1578                                  "%s: unknown metric %d", __func__, metric);
1579                         retval = -1;
1580                         goto bailout;
1581                         break; /* NOTREACHED */
1582 #endif
1583                 }
1584         }
1585
1586 bailout:
1587
1588         va_end(ap);
1589         return(retval);
1590 }
1591
1592 static int
1593 readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes)
1594 {
1595
1596         if (kvm_read(kd, addr, buf, nbytes) == -1) {
1597                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1598                          "%s: error reading value (kvm_read): %s", __func__,
1599                          kvm_geterr(kd));
1600                 return(-1);
1601         }
1602         return(0);
1603 }
1604
1605 static int
1606 readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes)
1607 {
1608         struct nlist nl[2];
1609
1610         nl[0].n_name = (char *)name;
1611         nl[1].n_name = NULL;
1612
1613         if (kvm_nlist(kd, nl) == -1) {
1614                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1615                          "%s: error getting name list (kvm_nlist): %s",
1616                          __func__, kvm_geterr(kd));
1617                 return(-1);
1618         }
1619         return(readkmem(kd, nl[0].n_value, buf, nbytes));
1620 }
1621
1622 /*
1623  * This duplicates the functionality of the kernel sysctl handler for poking
1624  * through crash dumps.
1625  */
1626 static char *
1627 get_devstat_kvm(kvm_t *kd)
1628 {
1629         int i, wp;
1630         long gen;
1631         struct devstat *nds;
1632         struct devstat ds;
1633         struct devstatlist dhead;
1634         int num_devs;
1635         char *rv = NULL;
1636
1637         if ((num_devs = devstat_getnumdevs(kd)) <= 0)
1638                 return(NULL);
1639         if (KREADNL(kd, X_DEVICE_STATQ, dhead) == -1)
1640                 return(NULL);
1641
1642         nds = STAILQ_FIRST(&dhead);
1643
1644         if ((rv = malloc(sizeof(gen))) == NULL) {
1645                 snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1646                          "%s: out of memory (initial malloc failed)",
1647                          __func__);
1648                 return(NULL);
1649         }
1650         gen = devstat_getgeneration(kd);
1651         memcpy(rv, &gen, sizeof(gen));
1652         wp = sizeof(gen);
1653         /*
1654          * Now push out all the devices.
1655          */
1656         for (i = 0; (nds != NULL) && (i < num_devs);
1657              nds = STAILQ_NEXT(nds, dev_links), i++) {
1658                 if (readkmem(kd, (long)nds, &ds, sizeof(ds)) == -1) {
1659                         free(rv);
1660                         return(NULL);
1661                 }
1662                 nds = &ds;
1663                 rv = (char *)reallocf(rv, sizeof(gen) +
1664                                       sizeof(ds) * (i + 1));
1665                 if (rv == NULL) {
1666                         snprintf(devstat_errbuf, sizeof(devstat_errbuf),
1667                                  "%s: out of memory (malloc failed)",
1668                                  __func__);
1669                         return(NULL);
1670                 }
1671                 memcpy(rv + wp, &ds, sizeof(ds));
1672                 wp += sizeof(ds);
1673         }
1674         return(rv);
1675 }