man.bsd.lv manual page server

Manual Page Search Parameters
DEVSTAT(9) Kernel Developer's Manual DEVSTAT(9)

devstat, devstat_add_entry, devstat_end_transaction, devstat_end_transaction_buf, devstat_remove_entry, devstat_start_transactionkernel interface for keeping device statistics

#include <sys/devicestat.h>

void
devstat_add_entry(struct devstat *ds, const char *dev_name, int unit_number, u_int32_t block_size, devstat_support_flags flags, devstat_type_flags device_type, devstat_priority priority);

void
devstat_remove_entry(struct devstat *ds);

void
devstat_start_transaction(struct devstat *ds);

void
devstat_end_transaction(struct devstat *ds, u_int32_t bytes, devstat_tag_type tag_type, devstat_trans_flags flags);

void
devstat_end_transaction_buf(struct devstat *ds, struct buf *bp);

The devstat subsystem is an interface for recording device statistics, as its name implies. The idea is to keep reasonably detailed statistics while utilizing a minimum amount of CPU time to record them. Thus, no statistical calculations are actually performed in the kernel portion of the devstat code. Instead, that is left for user programs to handle.

() registers a device with the devstat subsystem. The caller is expected to have already allocated and zeroed the devstat structure before calling this function. devstat_add_entry() takes several arguments:

ds
The devstat structure, allocated and zeroed by the client.
dev_name
The device name. e.g. da, cd, sa.
unit_number
Device unit number.
block_size
Block size of the device, if supported. If the device does not support a block size, or if the blocksize is unknown at the time the device is added to the devstat list, it should be set to 0.
flags
Flags indicating operations supported or not supported by the device. See below for details.
device_type
The device type. This is broken into three sections: base device type (e.g. direct access, CDROM, sequential access), interface type (IDE, SCSI or other) and a passthrough flag to indicate passthrough devices. See below for a complete list of types.
priority
The device priority. The priority is used to determine how devices are sorted within devstat's list of devices. Devices are sorted first by priority (highest to lowest), and then by attach order. See below for a complete list of available priorities.

() removes a device from the devstat subsystem. It takes the devstat structure for the device in question as an argument. The devstat generation number is incremented and the number of devices is decremented.

() registers the start of a transaction with the devstat subsystem. The busy count is incremented with each transaction start. When a device goes from idle to busy, the system uptime is recorded in the start_time field of the devstat structure.

() registers the end of a transaction with the devstat subsystem. It takes four arguments:

ds
The devstat structure for the device in question.
bytes
The number of bytes transferred in this transaction.
tag_type
Transaction tag type. See below for tag types.
flags
Transaction flags indicating whether the transaction was a read, write, or whether no data was transferred.

() is a wrapper for devstat_end_transaction() which pulls all the information from a struct buf which is ready for ().

The devstat structure is composed of the following fields:

dev_links
Each devstat structure is placed in a linked list when it is registered. The dev_links field contains a pointer to the next entry in the list of devstat structures.
device_number
The device number is a unique identifier for each device. The device number is incremented for each new device that is registered. The device number is currently only a 32-bit integer, but it could be enlarged if someone has a system with more than four billion device arrival events.
device_name
The device name is a text string given by the registering driver to identify itself. (e.g. “da”, “cd”, “sa”, etc.)
unit_number
The unit number identifies the particular instance of the peripheral driver in question.
bytes_written
This is the number of bytes that have been written to the device. This number is currently an unsigned 64 bit integer. This will hopefully eliminate the counter wrap that would come very quickly on some systems if 32 bit integers were used.
bytes_read
This is the number of bytes that have been read from the device.
bytes_freed
This is the number of bytes that have been freed/erased on the device.
num_reads
This is the number of reads from the device.
num_writes
This is the number of writes to the device.
num_frees
This is the number of free/erase operations on the device.
num_other
This is the number of transactions to the device which are neither reads or writes. For instance, SCSI drivers often send a test unit ready command to SCSI devices. The test unit ready command does not read or write any data. It merely causes the device to return its status.
busy_count
This is the current number of outstanding transactions for the device. This should never go below zero, and on an idle device it should be zero. If either one of these conditions is not true, it indicates a problem in the way () and devstat_end_transaction() are being called in client code. There should be one and only one transaction start event and one transaction end event for each transaction.
block_size
This is the block size of the device, if the device has a block size.
tag_types
This is an array of counters to record the number of various tag types that are sent to a device. See below for a list of tag types.
dev_creation_time
This is the time, as reported by () that the device was registered.
busy_time
This is the amount of time that the device busy count has been greater than zero. This is only updated when the busy count returns to zero.
start_time
This is the time, as reported by () that the device busy count went from zero to one.
last_comp_time
This is the time as reported by getmicrouptime() that a transaction last completed. It is used along with start_time to calculate the device busy time.
flags
These flags indicate which statistics measurements are supported by a particular device. These flags are primarily intended to serve as an aid to userland programs that decipher the statistics.
device_type
This is the device type. It consists of three parts: the device type (e.g. direct access, CDROM, sequential access, etc.), the interface (IDE, SCSI or other) and whether or not the device in question is a passthrough driver. See below for a complete list of device types.
priority
This is the priority. This is the first parameter used to determine where to insert a device in the devstat list. The second parameter is attach order. See below for a list of available priorities.

Each device is given a device type. Passthrough devices have the same underlying device type and interface as the device they provide an interface for, but they also have the passthrough flag set. The base device types are identical to the SCSI device type numbers, so with SCSI peripherals, the device type returned from an inquiry is usually ORed with the SCSI interface type and the passthrough flag if appropriate. The device type flags are as follows:

typedef enum {
	DEVSTAT_TYPE_DIRECT	= 0x000,
	DEVSTAT_TYPE_SEQUENTIAL	= 0x001,
	DEVSTAT_TYPE_PRINTER	= 0x002,
	DEVSTAT_TYPE_PROCESSOR	= 0x003,
	DEVSTAT_TYPE_WORM	= 0x004,
	DEVSTAT_TYPE_CDROM	= 0x005,
	DEVSTAT_TYPE_SCANNER	= 0x006,
	DEVSTAT_TYPE_OPTICAL	= 0x007,
	DEVSTAT_TYPE_CHANGER	= 0x008,
	DEVSTAT_TYPE_COMM	= 0x009,
	DEVSTAT_TYPE_ASC0	= 0x00a,
	DEVSTAT_TYPE_ASC1	= 0x00b,
	DEVSTAT_TYPE_STORARRAY	= 0x00c,
	DEVSTAT_TYPE_ENCLOSURE	= 0x00d,
	DEVSTAT_TYPE_FLOPPY	= 0x00e,
	DEVSTAT_TYPE_MASK	= 0x00f,
	DEVSTAT_TYPE_IF_SCSI	= 0x010,
	DEVSTAT_TYPE_IF_IDE	= 0x020,
	DEVSTAT_TYPE_IF_OTHER	= 0x030,
	DEVSTAT_TYPE_IF_MASK	= 0x0f0,
	DEVSTAT_TYPE_PASS	= 0x100
} devstat_type_flags;

Devices have a priority associated with them, which controls roughly where they are placed in the devstat list. The priorities are as follows:

typedef enum {
	DEVSTAT_PRIORITY_MIN	= 0x000,
	DEVSTAT_PRIORITY_OTHER	= 0x020,
	DEVSTAT_PRIORITY_PASS	= 0x030,
	DEVSTAT_PRIORITY_FD	= 0x040,
	DEVSTAT_PRIORITY_WFD	= 0x050,
	DEVSTAT_PRIORITY_TAPE	= 0x060,
	DEVSTAT_PRIORITY_CD	= 0x090,
	DEVSTAT_PRIORITY_DISK	= 0x110,
	DEVSTAT_PRIORITY_ARRAY	= 0x120,
	DEVSTAT_PRIORITY_MAX	= 0xfff
} devstat_priority;

Each device has associated with it flags to indicate what operations are supported or not supported. The devstat_support_flags values are as follows:

DEVSTAT_ALL_SUPPORTED
Every statistic type is supported by the device.
DEVSTAT_NO_BLOCKSIZE
This device does not have a blocksize.
DEVSTAT_NO_ORDERED_TAGS
This device does not support ordered tags.
DEVSTAT_BS_UNAVAILABLE
This device supports a blocksize, but it is currently unavailable. This flag is most often used with removable media drives.

Transactions to a device fall into one of three categories, which are represented in the flags passed into (). The transaction types are as follows:

typedef enum {
	DEVSTAT_NO_DATA	= 0x00,
	DEVSTAT_READ	= 0x01,
	DEVSTAT_WRITE	= 0x02,
	DEVSTAT_FREE	= 0x03
} devstat_trans_flags;

There are four possible values for the tag_type argument to ():

DEVSTAT_TAG_SIMPLE
The transaction had a simple tag.
DEVSTAT_TAG_HEAD
The transaction had a head of queue tag.
DEVSTAT_TAG_ORDERED
The transaction had an ordered tag.
DEVSTAT_TAG_NONE
The device doesn't support tags.

The tag type values correspond to the lower four bits of the SCSI tag definitions. In CAM, for instance, the tag_action from the CCB is ORed with 0xf to determine the tag type to pass in to ().

There is a macro, DEVSTAT_VERSION that is defined in <sys/devicestat.h>. This is the current version of the devstat subsystem, and it should be incremented each time a change is made that would require recompilation of userland programs that access devstat statistics. Userland programs use this version, via the kern.devstat.version sysctl variable to determine whether they are in sync with the kernel devstat structures.

systat(1), devstat(3), iostat(8), rpc.rstatd(8), vmstat(8)

The devstat statistics system appeared in FreeBSD 3.0.

Kenneth Merry <ken@FreeBSD.org>

There may be a need for some of the list manipulation code to be inside a critical section to ensure, for example, that the list of devices is not changed while someone is fetching the kern.devstat.all sysctl variable.

It is impossible with the current devstat architecture to accurately measure time per transaction. The only feasible way to accurately measure time per transaction would be to record a timestamp for every transaction. This measurement is probably not worthwhile for most people as it would adversely affect the performance of the system and cost space to store the timestamps for individual transactions.

May 22, 1998 DragonFly-5.6.1