freebsd-skq/lib/libdevstat/devstat.c

1540 lines
46 KiB
C
Raw Normal View History

/*
* Copyright (c) 1997, 1998 Kenneth D. Merry.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. The name of the author may not be used to endorse or promote products
* derived from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
2001-09-16 21:35:07 +00:00
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include <sys/types.h>
#include <sys/sysctl.h>
#include <sys/errno.h>
#include <sys/resource.h>
#include <sys/queue.h>
#include <ctype.h>
#include <err.h>
#include <fcntl.h>
#include <limits.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdarg.h>
#include <kvm.h>
#include "devstat.h"
typedef enum {
DEVSTAT_ARG_NOTYPE,
DEVSTAT_ARG_UINT64,
DEVSTAT_ARG_LD,
DEVSTAT_ARG_SKIP
} devstat_arg_type;
char devstat_errbuf[DEVSTAT_ERRBUF_SIZE];
/*
* Table to match descriptive strings with device types. These are in
* order from most common to least common to speed search time.
*/
struct devstat_match_table match_table[] = {
{"da", DEVSTAT_TYPE_DIRECT, DEVSTAT_MATCH_TYPE},
{"cd", DEVSTAT_TYPE_CDROM, DEVSTAT_MATCH_TYPE},
{"scsi", DEVSTAT_TYPE_IF_SCSI, DEVSTAT_MATCH_IF},
{"ide", DEVSTAT_TYPE_IF_IDE, DEVSTAT_MATCH_IF},
{"other", DEVSTAT_TYPE_IF_OTHER, DEVSTAT_MATCH_IF},
{"worm", DEVSTAT_TYPE_WORM, DEVSTAT_MATCH_TYPE},
{"sa", DEVSTAT_TYPE_SEQUENTIAL,DEVSTAT_MATCH_TYPE},
{"pass", DEVSTAT_TYPE_PASS, DEVSTAT_MATCH_PASS},
{"optical", DEVSTAT_TYPE_OPTICAL, DEVSTAT_MATCH_TYPE},
{"array", DEVSTAT_TYPE_STORARRAY, DEVSTAT_MATCH_TYPE},
{"changer", DEVSTAT_TYPE_CHANGER, DEVSTAT_MATCH_TYPE},
{"scanner", DEVSTAT_TYPE_SCANNER, DEVSTAT_MATCH_TYPE},
{"printer", DEVSTAT_TYPE_PRINTER, DEVSTAT_MATCH_TYPE},
{"floppy", DEVSTAT_TYPE_FLOPPY, DEVSTAT_MATCH_TYPE},
{"proc", DEVSTAT_TYPE_PROCESSOR, DEVSTAT_MATCH_TYPE},
{"comm", DEVSTAT_TYPE_COMM, DEVSTAT_MATCH_TYPE},
{"enclosure", DEVSTAT_TYPE_ENCLOSURE, DEVSTAT_MATCH_TYPE},
{NULL, 0, 0}
};
struct devstat_args {
devstat_metric metric;
devstat_arg_type argtype;
} devstat_arg_list[] = {
{ DSM_NONE, DEVSTAT_ARG_NOTYPE },
{ DSM_TOTAL_BYTES, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_BYTES_READ, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_BYTES_WRITE, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_TRANSFERS, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_TRANSFERS_READ, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_TRANSFERS_WRITE, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_TRANSFERS_OTHER, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_BLOCKS, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_BLOCKS_READ, DEVSTAT_ARG_UINT64 },
{ DSM_TOTAL_BLOCKS_WRITE, DEVSTAT_ARG_UINT64 },
{ DSM_KB_PER_TRANSFER, DEVSTAT_ARG_LD },
{ DSM_KB_PER_TRANSFER_READ, DEVSTAT_ARG_LD },
{ DSM_KB_PER_TRANSFER_WRITE, DEVSTAT_ARG_LD },
{ DSM_TRANSFERS_PER_SECOND, DEVSTAT_ARG_LD },
{ DSM_TRANSFERS_PER_SECOND_READ, DEVSTAT_ARG_LD },
{ DSM_TRANSFERS_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
{ DSM_TRANSFERS_PER_SECOND_OTHER, DEVSTAT_ARG_LD },
{ DSM_MB_PER_SECOND, DEVSTAT_ARG_LD },
{ DSM_MB_PER_SECOND_READ, DEVSTAT_ARG_LD },
{ DSM_MB_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
{ DSM_BLOCKS_PER_SECOND, DEVSTAT_ARG_LD },
{ DSM_BLOCKS_PER_SECOND_READ, DEVSTAT_ARG_LD },
{ DSM_BLOCKS_PER_SECOND_WRITE, DEVSTAT_ARG_LD },
{ DSM_MS_PER_TRANSACTION, DEVSTAT_ARG_LD },
{ DSM_MS_PER_TRANSACTION_READ, DEVSTAT_ARG_LD },
{ DSM_MS_PER_TRANSACTION_WRITE, DEVSTAT_ARG_LD },
{ DSM_SKIP, DEVSTAT_ARG_SKIP }
};
static const char *namelist[] = {
#define X_NUMDEVS 0
"_devstat_num_devs",
#define X_GENERATION 1
"_devstat_generation",
#define X_VERSION 2
"_devstat_version",
#define X_DEVICE_STATQ 3
"_device_statq",
#define X_END 4
};
/*
* Local function declarations.
*/
static int compare_select(const void *arg1, const void *arg2);
static int readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes);
static int readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes);
static char *get_devstat_kvm(kvm_t *kd);
#define KREADNL(kd, var, val) \
readkmem_nl(kd, namelist[var], &val, sizeof(val))
int
devstat_getnumdevs(kvm_t *kd)
{
size_t numdevsize;
int numdevs;
const char *func_name = "devstat_getnumdevs";
numdevsize = sizeof(int);
/*
* Find out how many devices we have in the system.
*/
if (kd == NULL) {
if (sysctlbyname("kern.devstat.numdevs", &numdevs,
&numdevsize, NULL, 0) == -1) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: error getting number of devices\n"
"%s: %s", func_name, func_name,
strerror(errno));
return(-1);
} else
return(numdevs);
} else {
if (KREADNL(kd, X_NUMDEVS, numdevs) == -1)
return(-1);
else
return(numdevs);
}
}
/*
* This is an easy way to get the generation number, but the generation is
* supplied in a more atmoic manner by the kern.devstat.all sysctl.
* Because this generation sysctl is separate from the statistics sysctl,
* the device list and the generation could change between the time that
* this function is called and the device list is retreived.
*/
long
devstat_getgeneration(kvm_t *kd)
{
size_t gensize;
long generation;
const char *func_name = "devstat_getgeneration";
gensize = sizeof(long);
/*
* Get the current generation number.
*/
if (kd == NULL) {
if (sysctlbyname("kern.devstat.generation", &generation,
&gensize, NULL, 0) == -1) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: error getting devstat generation\n%s: %s",
func_name, func_name, strerror(errno));
return(-1);
} else
return(generation);
} else {
if (KREADNL(kd, X_GENERATION, generation) == -1)
return(-1);
else
return(generation);
}
}
/*
* Get the current devstat version. The return value of this function
* should be compared with DEVSTAT_VERSION, which is defined in
* sys/devicestat.h. This will enable userland programs to determine
* whether they are out of sync with the kernel.
*/
int
devstat_getversion(kvm_t *kd)
{
size_t versize;
int version;
const char *func_name = "devstat_getversion";
versize = sizeof(int);
/*
* Get the current devstat version.
*/
if (kd == NULL) {
if (sysctlbyname("kern.devstat.version", &version, &versize,
NULL, 0) == -1) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: error getting devstat version\n%s: %s",
func_name, func_name, strerror(errno));
return(-1);
} else
return(version);
} else {
if (KREADNL(kd, X_VERSION, version) == -1)
return(-1);
else
return(version);
}
}
/*
* Check the devstat version we know about against the devstat version the
* kernel knows about. If they don't match, print an error into the
* devstat error buffer, and return -1. If they match, return 0.
*/
int
devstat_checkversion(kvm_t *kd)
{
const char *func_name = "devstat_checkversion";
int buflen, res, retval = 0, version;
version = devstat_getversion(kd);
if (version != DEVSTAT_VERSION) {
/*
* If getversion() returns an error (i.e. -1), then it
* has printed an error message in the buffer. Therefore,
* we need to add a \n to the end of that message before we
* print our own message in the buffer.
*/
if (version == -1)
buflen = strlen(devstat_errbuf);
else
buflen = 0;
res = snprintf(devstat_errbuf + buflen,
DEVSTAT_ERRBUF_SIZE - buflen,
"%s%s: userland devstat version %d is not "
"the same as the kernel\n%s: devstat "
"version %d\n", version == -1 ? "\n" : "",
func_name, DEVSTAT_VERSION, func_name, version);
if (res < 0)
devstat_errbuf[buflen] = '\0';
buflen = strlen(devstat_errbuf);
if (version < DEVSTAT_VERSION)
res = snprintf(devstat_errbuf + buflen,
DEVSTAT_ERRBUF_SIZE - buflen,
"%s: libdevstat newer than kernel\n",
func_name);
else
res = snprintf(devstat_errbuf + buflen,
DEVSTAT_ERRBUF_SIZE - buflen,
"%s: kernel newer than libdevstat\n",
func_name);
if (res < 0)
devstat_errbuf[buflen] = '\0';
retval = -1;
}
return(retval);
}
/*
* Get the current list of devices and statistics, and the current
* generation number.
*
* Return values:
* -1 -- error
* 0 -- device list is unchanged
* 1 -- device list has changed
*/
int
devstat_getdevs(kvm_t *kd, struct statinfo *stats)
{
int error;
size_t dssize;
int oldnumdevs;
long oldgeneration;
int retval = 0;
struct devinfo *dinfo;
const char *func_name = "devstat_getdevs";
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
struct timespec ts;
dinfo = stats->dinfo;
if (dinfo == NULL) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: stats->dinfo was NULL", func_name);
return(-1);
}
oldnumdevs = dinfo->numdevs;
oldgeneration = dinfo->generation;
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
clock_gettime(CLOCK_MONOTONIC, &ts);
stats->snap_time = ts.tv_sec + ts.tv_nsec * 1e-9;
if (kd == NULL) {
/* If this is our first time through, mem_ptr will be null. */
if (dinfo->mem_ptr == NULL) {
/*
* Get the number of devices. If it's negative, it's an
* error. Don't bother setting the error string, since
* getnumdevs() has already done that for us.
*/
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
if ((dinfo->numdevs = devstat_getnumdevs(NULL)) < 0)
return(-1);
/*
* The kern.devstat.all sysctl returns the current
* generation number, as well as all the devices.
* So we need four bytes more.
*/
dssize = (dinfo->numdevs * sizeof(struct devstat)) +
sizeof(long);
dinfo->mem_ptr = (u_int8_t *)malloc(dssize);
} else
dssize = (dinfo->numdevs * sizeof(struct devstat)) +
sizeof(long);
/*
* Request all of the devices. We only really allow for one
* ENOMEM failure. It would, of course, be possible to just go
* in a loop and keep reallocing the device structure until we
* don't get ENOMEM back. I'm not sure it's worth it, though.
* If devices are being added to the system that quickly, maybe
* the user can just wait until all devices are added.
*/
if ((error = sysctlbyname("kern.devstat.all", dinfo->mem_ptr,
&dssize, NULL, 0)) == -1) {
/*
* If we get ENOMEM back, that means that there are
* more devices now, so we need to allocate more
* space for the device array.
*/
if (errno == ENOMEM) {
/*
* No need to set the error string here,
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
* devstat_getnumdevs() will do that if it fails.
*/
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
if ((dinfo->numdevs = devstat_getnumdevs(NULL)) < 0)
return(-1);
dssize = (dinfo->numdevs *
sizeof(struct devstat)) + sizeof(long);
dinfo->mem_ptr = (u_int8_t *)
realloc(dinfo->mem_ptr, dssize);
if ((error = sysctlbyname("kern.devstat.all",
dinfo->mem_ptr, &dssize, NULL, 0)) == -1) {
snprintf(devstat_errbuf,
sizeof(devstat_errbuf),
"%s: error getting device "
"stats\n%s: %s", func_name,
func_name, strerror(errno));
return(-1);
}
} else {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: error getting device stats\n"
"%s: %s", func_name, func_name,
strerror(errno));
return(-1);
}
}
} else {
/*
* This is of course non-atomic, but since we are working
* on a core dump, the generation is unlikely to change
*/
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
if ((dinfo->numdevs = devstat_getnumdevs(NULL)) == -1)
return(-1);
if ((dinfo->mem_ptr = get_devstat_kvm(kd)) == NULL)
return(-1);
}
/*
* The sysctl spits out the generation as the first four bytes,
* then all of the device statistics structures.
*/
dinfo->generation = *(long *)dinfo->mem_ptr;
/*
* If the generation has changed, and if the current number of
* devices is not the same as the number of devices recorded in the
* devinfo structure, it is likely that the device list has shrunk.
* The reason that it is likely that the device list has shrunk in
* this case is that if the device list has grown, the sysctl above
* will return an ENOMEM error, and we will reset the number of
* devices and reallocate the device array. If the second sysctl
* fails, we will return an error and therefore never get to this
* point. If the device list has shrunk, the sysctl will not
* return an error since we have more space allocated than is
* necessary. So, in the shrinkage case, we catch it here and
* reallocate the array so that we don't use any more space than is
* necessary.
*/
if (oldgeneration != dinfo->generation) {
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
if (devstat_getnumdevs(NULL) != dinfo->numdevs) {
if ((dinfo->numdevs = devstat_getnumdevs(NULL)) < 0)
return(-1);
dssize = (dinfo->numdevs * sizeof(struct devstat)) +
sizeof(long);
dinfo->mem_ptr = (u_int8_t *)realloc(dinfo->mem_ptr,
dssize);
}
retval = 1;
}
dinfo->devices = (struct devstat *)(dinfo->mem_ptr + sizeof(long));
return(retval);
}
/*
* selectdevs():
*
* Devices are selected/deselected based upon the following criteria:
* - devices specified by the user on the command line
* - devices matching any device type expressions given on the command line
* - devices with the highest I/O, if 'top' mode is enabled
* - the first n unselected devices in the device list, if maxshowdevs
* devices haven't already been selected and if the user has not
* specified any devices on the command line and if we're in "add" mode.
*
* Input parameters:
* - device selection list (dev_select)
* - current number of devices selected (num_selected)
* - total number of devices in the selection list (num_selections)
* - devstat generation as of the last time selectdevs() was called
* (select_generation)
* - current devstat generation (current_generation)
* - current list of devices and statistics (devices)
* - number of devices in the current device list (numdevs)
* - compiled version of the command line device type arguments (matches)
* - This is optional. If the number of devices is 0, this will be ignored.
* - The matching code pays attention to the current selection mode. So
* if you pass in a matching expression, it will be evaluated based
* upon the selection mode that is passed in. See below for details.
* - number of device type matching expressions (num_matches)
* - Set to 0 to disable the matching code.
* - list of devices specified on the command line by the user (dev_selections)
* - number of devices selected on the command line by the user
* (num_dev_selections)
* - Our selection mode. There are four different selection modes:
* - add mode. (DS_SELECT_ADD) Any devices matching devices explicitly
* selected by the user or devices matching a pattern given by the
* user will be selected in addition to devices that are already
* selected. Additional devices will be selected, up to maxshowdevs
* number of devices.
* - only mode. (DS_SELECT_ONLY) Only devices matching devices
* explicitly given by the user or devices matching a pattern
* given by the user will be selected. No other devices will be
* selected.
* - addonly mode. (DS_SELECT_ADDONLY) This is similar to add and
* only. Basically, this will not de-select any devices that are
* current selected, as only mode would, but it will also not
* gratuitously select up to maxshowdevs devices as add mode would.
* - remove mode. (DS_SELECT_REMOVE) Any devices matching devices
* explicitly selected by the user or devices matching a pattern
* given by the user will be de-selected.
* - maximum number of devices we can select (maxshowdevs)
* - flag indicating whether or not we're in 'top' mode (perf_select)
*
* Output data:
* - the device selection list may be modified and passed back out
* - the number of devices selected and the total number of items in the
* device selection list may be changed
* - the selection generation may be changed to match the current generation
*
* Return values:
* -1 -- error
* 0 -- selected devices are unchanged
* 1 -- selected devices changed
*/
int
devstat_selectdevs(struct device_selection **dev_select, int *num_selected,
int *num_selections, long *select_generation,
long current_generation, struct devstat *devices,
int numdevs, struct devstat_match *matches, int num_matches,
char **dev_selections, int num_dev_selections,
devstat_select_mode select_mode, int maxshowdevs,
int perf_select)
{
2002-03-21 23:39:28 +00:00
int i, j, k;
int init_selections = 0, init_selected_var = 0;
struct device_selection *old_dev_select = NULL;
int old_num_selections = 0, old_num_selected;
int selection_number = 0;
int changed = 0, found = 0;
if ((dev_select == NULL) || (devices == NULL) || (numdevs <= 0))
return(-1);
/*
* We always want to make sure that we have as many dev_select
* entries as there are devices.
*/
/*
* In this case, we haven't selected devices before.
*/
if (*dev_select == NULL) {
*dev_select = (struct device_selection *)malloc(numdevs *
sizeof(struct device_selection));
*select_generation = current_generation;
init_selections = 1;
changed = 1;
/*
* In this case, we have selected devices before, but the device
* list has changed since we last selected devices, so we need to
* either enlarge or reduce the size of the device selection list.
*/
} else if (*num_selections != numdevs) {
*dev_select = (struct device_selection *)realloc(*dev_select,
numdevs * sizeof(struct device_selection));
*select_generation = current_generation;
init_selections = 1;
/*
* In this case, we've selected devices before, and the selection
* list is the same size as it was the last time, but the device
* list has changed.
*/
} else if (*select_generation < current_generation) {
*select_generation = current_generation;
init_selections = 1;
}
/*
* If we're in "only" mode, we want to clear out the selected
* variable since we're going to select exactly what the user wants
* this time through.
*/
if (select_mode == DS_SELECT_ONLY)
init_selected_var = 1;
/*
* In all cases, we want to back up the number of selected devices.
* It is a quick and accurate way to determine whether the selected
* devices have changed.
*/
old_num_selected = *num_selected;
/*
* We want to make a backup of the current selection list if
* the list of devices has changed, or if we're in performance
* selection mode. In both cases, we don't want to make a backup
* if we already know for sure that the list will be different.
* This is certainly the case if this is our first time through the
* selection code.
*/
if (((init_selected_var != 0) || (init_selections != 0)
|| (perf_select != 0)) && (changed == 0)){
old_dev_select = (struct device_selection *)malloc(
*num_selections * sizeof(struct device_selection));
old_num_selections = *num_selections;
bcopy(*dev_select, old_dev_select,
sizeof(struct device_selection) * *num_selections);
}
if (init_selections != 0) {
bzero(*dev_select, sizeof(struct device_selection) * numdevs);
for (i = 0; i < numdevs; i++) {
(*dev_select)[i].device_number =
devices[i].device_number;
strncpy((*dev_select)[i].device_name,
devices[i].device_name,
DEVSTAT_NAME_LEN);
(*dev_select)[i].device_name[DEVSTAT_NAME_LEN - 1]='\0';
(*dev_select)[i].unit_number = devices[i].unit_number;
(*dev_select)[i].position = i;
}
*num_selections = numdevs;
} else if (init_selected_var != 0) {
for (i = 0; i < numdevs; i++)
(*dev_select)[i].selected = 0;
}
/* we haven't gotten around to selecting anything yet.. */
if ((select_mode == DS_SELECT_ONLY) || (init_selections != 0)
|| (init_selected_var != 0))
*num_selected = 0;
/*
* Look through any devices the user specified on the command line
* and see if they match known devices. If so, select them.
*/
for (i = 0; (i < *num_selections) && (num_dev_selections > 0); i++) {
char tmpstr[80];
snprintf(tmpstr, sizeof(tmpstr), "%s%d",
(*dev_select)[i].device_name,
(*dev_select)[i].unit_number);
for (j = 0; j < num_dev_selections; j++) {
if (strcmp(tmpstr, dev_selections[j]) == 0) {
/*
* Here we do different things based on the
* mode we're in. If we're in add or
* addonly mode, we only select this device
* if it hasn't already been selected.
* Otherwise, we would be unnecessarily
* changing the selection order and
* incrementing the selection count. If
* we're in only mode, we unconditionally
* select this device, since in only mode
* any previous selections are erased and
* manually specified devices are the first
* ones to be selected. If we're in remove
* mode, we de-select the specified device and
* decrement the selection count.
*/
switch(select_mode) {
case DS_SELECT_ADD:
case DS_SELECT_ADDONLY:
if ((*dev_select)[i].selected)
break;
/* FALLTHROUGH */
case DS_SELECT_ONLY:
(*dev_select)[i].selected =
++selection_number;
(*num_selected)++;
break;
case DS_SELECT_REMOVE:
(*dev_select)[i].selected = 0;
(*num_selected)--;
/*
* This isn't passed back out, we
* just use it to keep track of
* how many devices we've removed.
*/
num_dev_selections--;
break;
}
break;
}
}
}
/*
* Go through the user's device type expressions and select devices
* accordingly. We only do this if the number of devices already
* selected is less than the maximum number we can show.
*/
for (i = 0; (i < num_matches) && (*num_selected < maxshowdevs); i++) {
/* We should probably indicate some error here */
if ((matches[i].match_fields == DEVSTAT_MATCH_NONE)
|| (matches[i].num_match_categories <= 0))
continue;
for (j = 0; j < numdevs; j++) {
int num_match_categories;
num_match_categories = matches[i].num_match_categories;
/*
* Determine whether or not the current device
* matches the given matching expression. This if
* statement consists of three components:
* - the device type check
* - the device interface check
* - the passthrough check
* If a the matching test is successful, it
* decrements the number of matching categories,
* and if we've reached the last element that
* needed to be matched, the if statement succeeds.
*
*/
if ((((matches[i].match_fields & DEVSTAT_MATCH_TYPE)!=0)
&& ((devices[j].device_type & DEVSTAT_TYPE_MASK) ==
(matches[i].device_type & DEVSTAT_TYPE_MASK))
&&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
|| (((matches[i].match_fields &
DEVSTAT_MATCH_PASS) == 0)
&& ((devices[j].device_type &
DEVSTAT_TYPE_PASS) == 0)))
&& (--num_match_categories == 0))
|| (((matches[i].match_fields & DEVSTAT_MATCH_IF) != 0)
&& ((devices[j].device_type & DEVSTAT_TYPE_IF_MASK) ==
(matches[i].device_type & DEVSTAT_TYPE_IF_MASK))
&&(((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
|| (((matches[i].match_fields &
DEVSTAT_MATCH_PASS) == 0)
&& ((devices[j].device_type &
DEVSTAT_TYPE_PASS) == 0)))
&& (--num_match_categories == 0))
|| (((matches[i].match_fields & DEVSTAT_MATCH_PASS)!=0)
&& ((devices[j].device_type & DEVSTAT_TYPE_PASS) != 0)
&& (--num_match_categories == 0))) {
/*
* This is probably a non-optimal solution
* to the problem that the devices in the
* device list will not be in the same
* order as the devices in the selection
* array.
*/
for (k = 0; k < numdevs; k++) {
if ((*dev_select)[k].position == j) {
found = 1;
break;
}
}
/*
* There shouldn't be a case where a device
* in the device list is not in the
* selection list...but it could happen.
*/
if (found != 1) {
fprintf(stderr, "selectdevs: couldn't"
" find %s%d in selection "
"list\n",
devices[j].device_name,
devices[j].unit_number);
break;
}
/*
* We do different things based upon the
* mode we're in. If we're in add or only
* mode, we go ahead and select this device
* if it hasn't already been selected. If
* it has already been selected, we leave
* it alone so we don't mess up the
* selection ordering. Manually specified
* devices have already been selected, and
* they have higher priority than pattern
* matched devices. If we're in remove
* mode, we de-select the given device and
* decrement the selected count.
*/
switch(select_mode) {
case DS_SELECT_ADD:
case DS_SELECT_ADDONLY:
case DS_SELECT_ONLY:
if ((*dev_select)[k].selected != 0)
break;
(*dev_select)[k].selected =
++selection_number;
(*num_selected)++;
break;
case DS_SELECT_REMOVE:
(*dev_select)[k].selected = 0;
(*num_selected)--;
break;
}
}
}
}
/*
* Here we implement "top" mode. Devices are sorted in the
* selection array based on two criteria: whether or not they are
* selected (not selection number, just the fact that they are
* selected!) and the number of bytes in the "bytes" field of the
* selection structure. The bytes field generally must be kept up
* by the user. In the future, it may be maintained by library
* functions, but for now the user has to do the work.
*
* At first glance, it may seem wrong that we don't go through and
* select every device in the case where the user hasn't specified
* any devices or patterns. In fact, though, it won't make any
* difference in the device sorting. In that particular case (i.e.
* when we're in "add" or "only" mode, and the user hasn't
* specified anything) the first time through no devices will be
* selected, so the only criterion used to sort them will be their
* performance. The second time through, and every time thereafter,
* all devices will be selected, so again selection won't matter.
*/
if (perf_select != 0) {
/* Sort the device array by throughput */
qsort(*dev_select, *num_selections,
sizeof(struct device_selection),
compare_select);
if (*num_selected == 0) {
/*
* Here we select every device in the array, if it
* isn't already selected. Because the 'selected'
* variable in the selection array entries contains
* the selection order, the devstats routine can show
* the devices that were selected first.
*/
for (i = 0; i < *num_selections; i++) {
if ((*dev_select)[i].selected == 0) {
(*dev_select)[i].selected =
++selection_number;
(*num_selected)++;
}
}
} else {
selection_number = 0;
for (i = 0; i < *num_selections; i++) {
if ((*dev_select)[i].selected != 0) {
(*dev_select)[i].selected =
++selection_number;
}
}
}
}
/*
* If we're in the "add" selection mode and if we haven't already
* selected maxshowdevs number of devices, go through the array and
* select any unselected devices. If we're in "only" mode, we
* obviously don't want to select anything other than what the user
* specifies. If we're in "remove" mode, it probably isn't a good
* idea to go through and select any more devices, since we might
* end up selecting something that the user wants removed. Through
* more complicated logic, we could actually figure this out, but
* that would probably require combining this loop with the various
* selections loops above.
*/
if ((select_mode == DS_SELECT_ADD) && (*num_selected < maxshowdevs)) {
for (i = 0; i < *num_selections; i++)
if ((*dev_select)[i].selected == 0) {
(*dev_select)[i].selected = ++selection_number;
(*num_selected)++;
}
}
/*
* Look at the number of devices that have been selected. If it
* has changed, set the changed variable. Otherwise, if we've
* made a backup of the selection list, compare it to the current
* selection list to see if the selected devices have changed.
*/
if ((changed == 0) && (old_num_selected != *num_selected))
changed = 1;
else if ((changed == 0) && (old_dev_select != NULL)) {
/*
* Now we go through the selection list and we look at
* it three different ways.
*/
for (i = 0; (i < *num_selections) && (changed == 0) &&
(i < old_num_selections); i++) {
/*
* If the device at index i in both the new and old
* selection arrays has the same device number and
* selection status, it hasn't changed. We
* continue on to the next index.
*/
if (((*dev_select)[i].device_number ==
old_dev_select[i].device_number)
&& ((*dev_select)[i].selected ==
old_dev_select[i].selected))
continue;
/*
* Now, if we're still going through the if
* statement, the above test wasn't true. So we
* check here to see if the device at index i in
* the current array is the same as the device at
* index i in the old array. If it is, that means
* that its selection number has changed. Set
* changed to 1 and exit the loop.
*/
else if ((*dev_select)[i].device_number ==
old_dev_select[i].device_number) {
changed = 1;
break;
}
/*
* If we get here, then the device at index i in
* the current array isn't the same device as the
* device at index i in the old array.
*/
else {
found = 0;
/*
* Search through the old selection array
* looking for a device with the same
* device number as the device at index i
* in the current array. If the selection
* status is the same, then we mark it as
* found. If the selection status isn't
* the same, we break out of the loop.
* Since found isn't set, changed will be
* set to 1 below.
*/
for (j = 0; j < old_num_selections; j++) {
if (((*dev_select)[i].device_number ==
old_dev_select[j].device_number)
&& ((*dev_select)[i].selected ==
old_dev_select[j].selected)){
found = 1;
break;
}
else if ((*dev_select)[i].device_number
== old_dev_select[j].device_number)
break;
}
if (found == 0)
changed = 1;
}
}
}
if (old_dev_select != NULL)
free(old_dev_select);
return(changed);
}
/*
* Comparison routine for qsort() above. Note that the comparison here is
* backwards -- generally, it should return a value to indicate whether
* arg1 is <, =, or > arg2. Instead, it returns the opposite. The reason
* it returns the opposite is so that the selection array will be sorted in
* order of decreasing performance. We sort on two parameters. The first
* sort key is whether or not one or the other of the devices in question
* has been selected. If one of them has, and the other one has not, the
* selected device is automatically more important than the unselected
* device. If neither device is selected, we judge the devices based upon
* performance.
*/
static int
compare_select(const void *arg1, const void *arg2)
{
if ((((const struct device_selection *)arg1)->selected)
&& (((const struct device_selection *)arg2)->selected == 0))
return(-1);
else if ((((const struct device_selection *)arg1)->selected == 0)
&& (((const struct device_selection *)arg2)->selected))
return(1);
else if (((const struct device_selection *)arg2)->bytes <
((const struct device_selection *)arg1)->bytes)
return(-1);
else if (((const struct device_selection *)arg2)->bytes >
((const struct device_selection *)arg1)->bytes)
return(1);
else
return(0);
}
/*
* Take a string with the general format "arg1,arg2,arg3", and build a
* device matching expression from it.
*/
int
devstat_buildmatch(char *match_str, struct devstat_match **matches,
int *num_matches)
{
char *tstr[5];
char **tempstr;
int num_args;
2002-03-21 23:39:28 +00:00
int i, j;
const char *func_name = "devstat_buildmatch";
/* We can't do much without a string to parse */
if (match_str == NULL) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: no match expression", func_name);
return(-1);
}
/*
* Break the (comma delimited) input string out into separate strings.
*/
for (tempstr = tstr, num_args = 0;
(*tempstr = strsep(&match_str, ",")) != NULL && (num_args < 5);
num_args++)
if (**tempstr != '\0')
if (++tempstr >= &tstr[5])
break;
/* The user gave us too many type arguments */
if (num_args > 3) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: too many type arguments", func_name);
return(-1);
}
/*
* Since you can't realloc a pointer that hasn't been malloced
* first, we malloc first and then realloc.
*/
if (*num_matches == 0)
*matches = (struct devstat_match *)malloc(
sizeof(struct devstat_match));
else
*matches = (struct devstat_match *)realloc(*matches,
sizeof(struct devstat_match) * (*num_matches + 1));
/* Make sure the current entry is clear */
bzero(&matches[0][*num_matches], sizeof(struct devstat_match));
/*
* Step through the arguments the user gave us and build a device
* matching expression from them.
*/
for (i = 0; i < num_args; i++) {
char *tempstr2, *tempstr3;
/*
* Get rid of leading white space.
*/
tempstr2 = tstr[i];
while (isspace(*tempstr2) && (*tempstr2 != '\0'))
tempstr2++;
/*
* Get rid of trailing white space.
*/
tempstr3 = &tempstr2[strlen(tempstr2) - 1];
while ((*tempstr3 != '\0') && (tempstr3 > tempstr2)
&& (isspace(*tempstr3))) {
*tempstr3 = '\0';
tempstr3--;
}
/*
* Go through the match table comparing the user's
* arguments to known device types, interfaces, etc.
*/
for (j = 0; match_table[j].match_str != NULL; j++) {
/*
* We do case-insensitive matching, in case someone
* wants to enter "SCSI" instead of "scsi" or
* something like that. Only compare as many
* characters as are in the string in the match
* table. This should help if someone tries to use
* a super-long match expression.
*/
if (strncasecmp(tempstr2, match_table[j].match_str,
strlen(match_table[j].match_str)) == 0) {
/*
* Make sure the user hasn't specified two
* items of the same type, like "da" and
* "cd". One device cannot be both.
*/
if (((*matches)[*num_matches].match_fields &
match_table[j].match_field) != 0) {
snprintf(devstat_errbuf,
sizeof(devstat_errbuf),
"%s: cannot have more than "
"one match item in a single "
"category", func_name);
return(-1);
}
/*
* If we've gotten this far, we have a
* winner. Set the appropriate fields in
* the match entry.
*/
(*matches)[*num_matches].match_fields |=
match_table[j].match_field;
(*matches)[*num_matches].device_type |=
match_table[j].type;
(*matches)[*num_matches].num_match_categories++;
break;
}
}
/*
* We should have found a match in the above for loop. If
* not, that means the user entered an invalid device type
* or interface.
*/
if ((*matches)[*num_matches].num_match_categories != (i + 1)) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: unknown match item \"%s\"", func_name,
tstr[i]);
return(-1);
}
}
(*num_matches)++;
return(0);
}
/*
* Compute a number of device statistics. Only one field is mandatory, and
* that is "current". Everything else is optional. The caller passes in
* pointers to variables to hold the various statistics he desires. If he
* doesn't want a particular staistic, he should pass in a NULL pointer.
* Return values:
* 0 -- success
* -1 -- failure
*/
int
compute_stats(struct devstat *current, struct devstat *previous,
long double etime, u_int64_t *total_bytes,
u_int64_t *total_transfers, u_int64_t *total_blocks,
long double *kb_per_transfer, long double *transfers_per_second,
long double *mb_per_second, long double *blocks_per_second,
long double *ms_per_transaction)
{
return(devstat_compute_statistics(current, previous, etime,
total_bytes ? DSM_TOTAL_BYTES : DSM_SKIP,
total_bytes,
total_transfers ? DSM_TOTAL_TRANSFERS : DSM_SKIP,
total_transfers,
total_blocks ? DSM_TOTAL_BLOCKS : DSM_SKIP,
total_blocks,
kb_per_transfer ? DSM_KB_PER_TRANSFER : DSM_SKIP,
kb_per_transfer,
transfers_per_second ? DSM_TRANSFERS_PER_SECOND : DSM_SKIP,
transfers_per_second,
mb_per_second ? DSM_MB_PER_SECOND : DSM_SKIP,
mb_per_second,
blocks_per_second ? DSM_BLOCKS_PER_SECOND : DSM_SKIP,
blocks_per_second,
ms_per_transaction ? DSM_MS_PER_TRANSACTION : DSM_SKIP,
ms_per_transaction,
DSM_NONE));
}
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
/* This is 1/2^64 */
#define BINTIME_SCALE 5.42101086242752217003726400434970855712890625e-20
long double
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
devstat_compute_etime(struct bintime *cur_time, struct bintime *prev_time)
{
long double etime;
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
etime = cur_time->sec;
etime += cur_time->frac * BINTIME_SCALE;
if (prev_time != NULL) {
etime -= prev_time->sec;
etime -= prev_time->frac * BINTIME_SCALE;
}
return(etime);
}
int
devstat_compute_statistics(struct devstat *current, struct devstat *previous,
long double etime, ...)
{
const char *func_name = "devstat_compute_statistics";
u_int64_t totalbytes, totalbytesread, totalbyteswrite;
u_int64_t totaltransfers, totaltransfersread, totaltransferswrite;
u_int64_t totaltransfersother, totalblocks, totalblocksread;
u_int64_t totalblockswrite;
va_list ap;
devstat_metric metric;
u_int64_t *destu64;
long double *destld;
int retval;
retval = 0;
/*
* current is the only mandatory field.
*/
if (current == NULL) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: current stats structure was NULL", func_name);
return(-1);
}
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
totalbytesread = current->bytes[DEVSTAT_READ] -
((previous) ? previous->bytes[DEVSTAT_READ] : 0);
totalbyteswrite = current->bytes[DEVSTAT_WRITE] -
((previous) ? previous->bytes[DEVSTAT_WRITE] : 0);
totalbytes = totalbytesread + totalbyteswrite;
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
totaltransfersread = current->operations[DEVSTAT_READ] -
((previous) ? previous->operations[DEVSTAT_READ] : 0);
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
totaltransferswrite = current->operations[DEVSTAT_WRITE] -
((previous) ? previous->operations[DEVSTAT_WRITE] : 0);
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
totaltransfersother = current->operations[DEVSTAT_NO_DATA] -
((previous) ? previous->operations[DEVSTAT_NO_DATA] : 0);
totaltransfers = totaltransfersread + totaltransferswrite +
totaltransfersother;
totalblocks = totalbytes;
totalblocksread = totalbytesread;
totalblockswrite = totalbyteswrite;
if (current->block_size > 0) {
totalblocks /= current->block_size;
totalblocksread /= current->block_size;
totalblockswrite /= current->block_size;
} else {
totalblocks /= 512;
totalblocksread /= 512;
totalblockswrite /= 512;
}
va_start(ap, etime);
while ((metric = (devstat_metric)va_arg(ap, devstat_metric)) != 0) {
if (metric == DSM_NONE)
break;
if (metric >= DSM_MAX) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: metric %d is out of range", func_name,
metric);
retval = -1;
goto bailout;
}
switch (devstat_arg_list[metric].argtype) {
case DEVSTAT_ARG_UINT64:
destu64 = (u_int64_t *)va_arg(ap, u_int64_t *);
break;
case DEVSTAT_ARG_LD:
destld = (long double *)va_arg(ap, long double *);
break;
case DEVSTAT_ARG_SKIP:
destld = (long double *)va_arg(ap, long double *);
break;
default:
retval = -1;
goto bailout;
break; /* NOTREACHED */
}
if (devstat_arg_list[metric].argtype == DEVSTAT_ARG_SKIP)
continue;
switch (metric) {
case DSM_TOTAL_BYTES:
*destu64 = totalbytes;
break;
case DSM_TOTAL_BYTES_READ:
*destu64 = totalbytesread;
break;
case DSM_TOTAL_BYTES_WRITE:
*destu64 = totalbyteswrite;
break;
case DSM_TOTAL_TRANSFERS:
*destu64 = totaltransfers;
break;
case DSM_TOTAL_TRANSFERS_READ:
*destu64 = totaltransfersread;
break;
case DSM_TOTAL_TRANSFERS_WRITE:
*destu64 = totaltransferswrite;
break;
case DSM_TOTAL_TRANSFERS_OTHER:
*destu64 = totaltransfersother;
break;
case DSM_TOTAL_BLOCKS:
*destu64 = totalblocks;
break;
case DSM_TOTAL_BLOCKS_READ:
*destu64 = totalblocksread;
break;
case DSM_TOTAL_BLOCKS_WRITE:
*destu64 = totalblockswrite;
break;
case DSM_KB_PER_TRANSFER:
*destld = totalbytes;
*destld /= 1024;
if (totaltransfers > 0)
*destld /= totaltransfers;
else
*destld = 0.0;
break;
case DSM_KB_PER_TRANSFER_READ:
*destld = totalbytesread;
*destld /= 1024;
if (totaltransfersread > 0)
*destld /= totaltransfersread;
else
*destld = 0.0;
break;
case DSM_KB_PER_TRANSFER_WRITE:
*destld = totalbyteswrite;
*destld /= 1024;
if (totaltransferswrite > 0)
*destld /= totaltransferswrite;
else
*destld = 0.0;
break;
case DSM_TRANSFERS_PER_SECOND:
if (etime > 0.0) {
*destld = totaltransfers;
*destld /= etime;
} else
*destld = 0.0;
break;
case DSM_TRANSFERS_PER_SECOND_READ:
if (etime > 0.0) {
*destld = totaltransfersread;
*destld /= etime;
} else
*destld = 0.0;
break;
case DSM_TRANSFERS_PER_SECOND_WRITE:
if (etime > 0.0) {
*destld = totaltransferswrite;
*destld /= etime;
} else
*destld = 0.0;
break;
case DSM_TRANSFERS_PER_SECOND_OTHER:
if (etime > 0.0) {
*destld = totaltransfersother;
*destld /= etime;
} else
*destld = 0.0;
break;
case DSM_MB_PER_SECOND:
*destld = totalbytes;
*destld /= 1024 * 1024;
if (etime > 0.0)
*destld /= etime;
else
*destld = 0.0;
break;
case DSM_MB_PER_SECOND_READ:
*destld = totalbytesread;
*destld /= 1024 * 1024;
if (etime > 0.0)
*destld /= etime;
else
*destld = 0.0;
break;
case DSM_MB_PER_SECOND_WRITE:
*destld = totalbyteswrite;
*destld /= 1024 * 1024;
if (etime > 0.0)
*destld /= etime;
else
*destld = 0.0;
break;
case DSM_BLOCKS_PER_SECOND:
*destld = totalblocks;
if (etime > 0.0)
*destld /= etime;
else
*destld = 0.0;
break;
case DSM_BLOCKS_PER_SECOND_READ:
*destld = totalblocksread;
if (etime > 0.0)
*destld /= etime;
else
*destld = 0.0;
break;
case DSM_BLOCKS_PER_SECOND_WRITE:
*destld = totalblockswrite;
if (etime > 0.0)
*destld /= etime;
else
*destld = 0.0;
break;
/*
* This calculation is somewhat bogus. It simply divides
* the elapsed time by the total number of transactions
* completed. While that does give the caller a good
* picture of the average rate of transaction completion,
* it doesn't necessarily give the caller a good view of
* how long transactions took to complete on average.
* Those two numbers will be different for a device that
* can handle more than one transaction at a time. e.g.
* SCSI disks doing tagged queueing.
*
* The only way to accurately determine the real average
* time per transaction would be to compute and store the
* time on a per-transaction basis. That currently isn't
* done in the kernel, and would only be desireable if it
* could be implemented in a somewhat non-intrusive and high
* performance way.
*/
case DSM_MS_PER_TRANSACTION:
if (totaltransfers > 0) {
*destld = etime;
*destld /= totaltransfers;
*destld *= 1000;
} else
*destld = 0.0;
break;
/*
* As above, these next two really only give the average
* rate of completion for read and write transactions, not
* the average time the transaction took to complete.
*/
case DSM_MS_PER_TRANSACTION_READ:
if (totaltransfersread > 0) {
*destld = etime;
*destld /= totaltransfersread;
*destld *= 1000;
} else
*destld = 0.0;
break;
case DSM_MS_PER_TRANSACTION_WRITE:
if (totaltransferswrite > 0) {
*destld = etime;
*destld /= totaltransferswrite;
*destld *= 1000;
} else
*destld = 0.0;
break;
default:
/*
* This shouldn't happen, since we should have
* caught any out of range metrics at the top of
* the loop.
*/
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: unknown metric %d", func_name, metric);
retval = -1;
goto bailout;
break; /* NOTREACHED */
}
}
bailout:
va_end(ap);
return(retval);
}
static int
readkmem(kvm_t *kd, unsigned long addr, void *buf, size_t nbytes)
{
const char *func_name = "readkmem";
if (kvm_read(kd, addr, buf, nbytes) == -1) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: error reading value (kvm_read): %s", func_name,
kvm_geterr(kd));
return(-1);
}
return(0);
}
static int
readkmem_nl(kvm_t *kd, const char *name, void *buf, size_t nbytes)
{
const char *func_name = "readkmem_nl";
struct nlist nl[2];
(const char *)nl[0].n_name = name;
nl[1].n_name = NULL;
if (kvm_nlist(kd, nl) == -1) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: error getting name list (kvm_nlist): %s",
func_name, kvm_geterr(kd));
return(-1);
}
return(readkmem(kd, nl[0].n_value, buf, nbytes));
}
/*
* This duplicates the functionality of the kernel sysctl handler for poking
* through crash dumps.
*/
static char *
get_devstat_kvm(kvm_t *kd)
{
int error, i, wp;
long gen;
struct devstat *nds;
struct devstat ds;
struct devstatlist dhead;
int num_devs;
char *rv = NULL;
const char *func_name = "get_devstat_kvm";
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
if ((num_devs = devstat_getnumdevs(kd)) <= 0)
return(NULL);
error = 0;
if (KREADNL(kd, X_DEVICE_STATQ, dhead) == -1)
return(NULL);
nds = STAILQ_FIRST(&dhead);
if ((rv = malloc(sizeof(gen))) == NULL) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: out of memory (initial malloc failed)",
func_name);
return(NULL);
}
Run a revision of the devstat interface: Kernel: Change statistics to use the *uptime() timescale (ie: relative to boottime) rather than the UTC aligned timescale. This makes the device statistics code oblivious to clock steps. Change timestamps to bintime format, they are cheaper. Remove the "busy_count", and replace it with two counter fields: "start_count" and "end_count", which are updated in the down and up paths respectively. This removes the locking constraint on devstat. Add a timestamp argument to devstat_start_transaction(), this will normally be a timestamp set by the *_bio() function in bp->bio_t0. Use this field to calculate duration of I/O operations. Add two timestamp arguments to devstat_end_transaction(), one is the current time, a NULL pointer means "take timestamp yourself", the other is the timestamp of when this transaction started (see above). Change calculation of busy_time to operate on "the salami principle": Only when we are idle, which we can determine by the start+end counts being identical, do we update the "busy_from" field in the down path. In the up path we accumulate the timeslice in busy_time and update busy_from. Change the byte_* and num_* fields into two arrays: bytes[] and operations[]. Userland: Change the misleading "busy_time" name to be called "snap_time" and make the time long double since that is what most users need anyway, fill it using clock_gettime(CLOCK_MONOTONIC) to put it on the same timescale as the kernel fields. Change devstat_compute_etime() to operate on struct bintime. Remove the version 2 legacy interface: the change to bintime makes compatibility far too expensive. Fix a bug in systat's "vm" page where boot relative busy times would be bogus. Bump __FreeBSD_version to 500107 Review & Collaboration by: ken
2003-03-15 21:59:06 +00:00
gen = devstat_getgeneration(kd);
memcpy(rv, &gen, sizeof(gen));
wp = sizeof(gen);
/*
* Now push out all the devices.
*/
for (i = 0; (nds != NULL) && (i < num_devs);
nds = STAILQ_NEXT(nds, dev_links), i++) {
if (readkmem(kd, (long)nds, &ds, sizeof(ds)) == -1) {
free(rv);
return(NULL);
}
nds = &ds;
rv = (char *)reallocf(rv, sizeof(gen) +
sizeof(ds) * (i + 1));
if (rv == NULL) {
snprintf(devstat_errbuf, sizeof(devstat_errbuf),
"%s: out of memory (malloc failed)",
func_name);
return(NULL);
}
memcpy(rv + wp, &ds, sizeof(ds));
wp += sizeof(ds);
}
return(rv);
}