DISK(9) FreeBSD Kernel Developer's Manual DISK(9)
NAME
disk, disk_init, disk_attach, disk_begindetach, disk_detach,
disk_destroy, disk_wait, disk_busy, disk_unbusy, disk_isbusy, disk_find,
disk_set_info - generic disk framework
SYNOPSIS
#include <sys/types.h>
#include <sys/disklabel.h>
#include <sys/disk.h>
void
disk_init(struct disk *, const char *name,
const struct dkdriver *driver);
void
disk_attach(struct disk *);
void
disk_begindetach(struct disk *, int (*lastclose)(device_t),
device_t self, int flags);
void
disk_detach(struct disk *);
void
disk_destroy(struct disk *);
void
disk_wait(struct disk *);
void
disk_busy(struct disk *);
void
disk_unbusy(struct disk *, long bcount, int read);
bool
disk_isbusy(struct disk *);
struct disk *
disk_find(const char *);
void
disk_set_info(device_t, struct disk *, const char *type);
DESCRIPTION
The NetBSD generic disk framework is designed to provide flexible,
scalable, and consistent handling of disk state and metrics information.
The fundamental component of this framework is the disk structure, which
is defined as follows:
struct disk {
TAILQ_ENTRY(disk) dk_link; /* link in global disklist */
const char *dk_name; /* disk name */
prop_dictionary_t dk_info; /* reference to disk-info dictionary */
int dk_bopenmask; /* block devices open */
int dk_copenmask; /* character devices open */
int dk_openmask; /* composite (bopen|copen) */
int dk_state; /* label state ### */
int dk_blkshift; /* shift to convert DEV_BSIZE to blks */
int dk_byteshift; /* shift to convert bytes to blks */
/*
* Metrics data; note that some metrics may have no meaning
* on certain types of disks.
*/
struct io_stats *dk_stats;
const struct dkdriver *dk_driver; /* pointer to driver */
/*
* Information required to be the parent of a disk wedge.
*/
kmutex_t dk_rawlock; /* lock on these fields */
u_int dk_rawopens; /* # of opens of rawvp */
struct vnode *dk_rawvp; /* vnode for the RAW_PART bdev */
kmutex_t dk_openlock; /* lock on these and openmask */
u_int dk_nwedges; /* # of configured wedges */
/* all wedges on this disk */
LIST_HEAD(, dkwedge_softc) dk_wedges;
/*
* Disk label information. Storage for the in-core disk label
* must be dynamically allocated, otherwise the size of this
* structure becomes machine-dependent.
*/
daddr_t dk_labelsector; /* sector containing label */
struct disklabel *dk_label; /* label */
struct cpu_disklabel *dk_cpulabel;
};
The system maintains a global linked-list of all disks attached to the
system. This list, called disklist, may grow or shrink over time as
disks are dynamically added and removed from the system. Drivers which
currently make use of the detachment capability of the framework are the
ccd, dm, and vnd pseudo-device drivers.
The following is a brief description of each function in the framework:
disk_init() Initialize the disk structure.
disk_attach() Attach a disk; allocate storage for the disklabel, set the
"attached time" timestamp, insert the disk into the
disklist, and increment the system disk count.
disk_begindetach()
Check whether the disk is open, and if not, return 0. If
the disk is open, and DETACH_FORCE is not set in flags,
return EBUSY. Otherwise, call the provided lastclose
routine (if not NULL) and return its exit code.
disk_detach() Detach a disk; free storage for the disklabel, remove the
disk from the disklist, and decrement the system disk
count. If the count drops below zero, panic.
disk_destroy()
Release resources used by the disk structure when it is no
longer required.
disk_wait() Disk timings are measured by counting the number of queued
requests (wait counter) and requests issued to the
hardware (busy counter) and keeping timestamp when the
counters change. The time interval between two changes of
a counter is accumulated into a total and also multiplied
by the counter value and the accumulated into a sum. Both
values can be used to determine how much time is spent in
the driver queue or in-flight to the hardware as well as
the average number of requests in either state.
disk_wait() increment the disk's wait counter and handles
the accumulation.
disk_busy() Decrements the disk's wait counter and increments the
disk's "busy counter", and handles either accumulation.
If the wait counter is still zero, it is assumed that the
driver hasn't been updated to call disk_wait(), then only
the values from the busy counter are available.
disk_unbusy() Decrement the disk's busy counter and handles the
accumulation. The third argument read specifies the
direction of I/O; if non-zero it means reading from the
disk, otherwise it means writing to the disk.
disk_isbusy() Returns true if disk is marked as busy and false if it is
not.
disk_find() Return a pointer to the disk structure corresponding to
the name provided, or NULL if the disk does not exist.
disk_set_info()
Setup disk-info dictionary and other dependent values of
the disk structure, the driver must have initialized the
dk_geom member of struct disk with suitable values. If
type is not NULL, it will be added to the dictionary.
The functions typically called by device drivers are disk_init()
disk_attach(), disk_begindetach(), disk_detach(), disk_destroy(),
disk_wait(), disk_busy(), disk_unbusy(), and disk_set_info(). The
function disk_find() is provided as a utility function.
DISK IOCTLS
The following ioctls should be implemented by disk drivers:
DIOCGDINFO struct disklabel
Get disklabel.
DIOCSDINFO struct disklabel
Set in-memory disklabel.
DIOCWDINFO struct disklabel
Set in-memory disklabel and write on-disk disklabel.
DIOCGPART struct partinfo
Get partition information. This is used internally.
DIOCRFORMAT struct format_op
Read format.
DIOCWFORMAT struct format_op
Write format.
DIOCSSTEP int
Set step rate.
DIOCSRETRIES int
Set number of retries.
DIOCKLABEL int
Specify whether to keep or drop the in-memory disklabel when the
device is closed.
DIOCWLABEL int
Enable or disable writing to the part of the disk that contains
the label.
DIOCSBAD struct dkbad
Set kernel dkbad.
DIOCEJECT int
Eject removable disk.
DIOCLOCK int
Lock or unlock disk pack. For devices with removable media,
locking is intended to prevent the operator from removing the
media.
DIOCGDEFLABEL struct disklabel
Get default label.
DIOCCLRLABEL
Clear disk label.
DIOCGCACHE int
Get status of disk read and write caches. The result is a
bitmask containing the following values:
DKCACHE_READ Read cache enabled.
DKCACHE_WRITE Write(back) cache enabled.
DKCACHE_RCHANGE Read cache enable is changeable.
DKCACHE_WCHANGE Write cache enable is changeable.
DKCACHE_SAVE Cache parameters may be saved, so that they
persist across reboots or device detach/attach
cycles.
DIOCSCACHE int
Set status of disk read and write caches. The input is a bitmask
in the same format as used for DIOCGCACHE.
DIOCCACHESYNC int
Synchronise the disk cache. This causes information in the
disk's write cache (if any) to be flushed to stable storage. The
argument specifies whether or not to force a flush even if the
kernel believes that there is no outstanding data.
DIOCBSLIST struct disk_badsecinfo
Get bad sector list.
DIOCBSFLUSH
Flush bad sector list.
DIOCAWEDGE struct dkwedge_info
Add wedge.
DIOCGWEDGEINFO struct dkwedge_info
Get wedge information.
DIOCDWEDGE struct dkwedge_info
Delete wedge.
DIOCLWEDGES struct dkwedge_list
List wedges.
DIOCGSTRATEGY struct disk_strategy
Get disk buffer queue strategy.
DIOCSSTRATEGY struct disk_strategy
Set disk buffer queue strategy.
DIOCGDISKINFO struct plistref
Get disk-info dictionary.
DIOCGMEDIASIZE off_t
Get disk size in bytes.
DIOCGSECTORSIZE u_int
Get sector size in bytes.
USING THE FRAMEWORK
This section includes a description on basic use of the framework and
example usage of its functions. Actual implementation of a device driver
which uses the framework may vary.
Each device in the system uses a "softc" structure which contains
autoconfiguration and state information for that device. In the case of
disks, the softc should also contain one instance of the disk structure,
e.g.:
struct foo_softc {
device_t sc_dev; /* generic device information */
struct disk sc_dk; /* generic disk information */
[ . . . more . . . ]
};
In order for the system to gather metrics data about a disk, the disk
must be registered with the system. The disk_attach() routine performs
all of the functions currently required to register a disk with the
system including allocation of disklabel storage space, recording of the
time since boot that the disk was attached, and insertion into the
disklist. Note that since this function allocates storage space for the
disklabel, it must be called before the disklabel is read from the media
or used in any other way. Before disk_attach() is called, a portions of
the disk structure must be initialized with data specific to that disk.
For example, in the "foo" disk driver, the following would be performed
in the autoconfiguration "attach" routine:
void
fooattach(device_t parent, device_t self, void *aux)
{
struct foo_softc *sc = device_private(self);
[ . . . ]
/* Initialize and attach the disk structure. */
disk_init(&sc->sc_dk, device_xname(self), &foodkdriver);
disk_attach(&sc->sc_dk);
/* Read geometry and fill in pertinent parts of disklabel. */
/* Initialize geometry values of the disk structure */
[ . . . ]
disk_set_info(&self>, &sc->sc_dk, type);
}
The foodkdriver above is the disk's "driver" switch. This switch
currently includes pointers to several driver entry points, where only
the d_strategy entry point is used by the disk framework. This switch
needs to have global scope and should be initialized as follows:
void (foostrategy)(struct buf *);
void (foominphys)(struct buf *);
int (fooopen)(dev_t, int, int, struct lwp *);
int (fooclose)(dev_t, int, int, struct lwp *);
int (foo_discard)(device_t, off_t, off_t);
int (foo_diskstart)(device_t, struct buf *);
void (foo_iosize)(device_t, int *);
int (foo_dumpblocks)(device_t, void *, daddr_t, int);
int (foo_lastclose)(device_t);
int (foo_firstopen)(device_t, dev_t, int, int);
int (foo_label)(device_t, struct disklabel *);
const struct dkdriver foodkdriver = {
.d_open = fooopen,
.d_close = fooclose,
.d_strategy = foostrategy,
.d_minphys = foominphys,
.d_discard = foo_discard,
.d_diskstart = foo_diskstart, /* optional */
.d_dumpblocks = foo_dumpblocks, /* optional */
.d_iosize = foo_iosize, /* optional */
.d_firstopen = foo_firstopen, /* optional */
.d_lastclose = foo_lastclose, /* optional */
.d_label = foo_label, /* optional */
};
Once the disk is attached, metrics may be gathered on that disk. In
order to gather metrics data, the driver must tell the framework when the
disk queues, starts and stops operations. This functionality is provided
by the disk_wait(), disk_busy() and disk_unbusy() routines. Because
struct disk is part of device driver private data it needs to be guarded.
Mutual exclusion must be done by driver disk_wait(), disk_busy() and
disk_unbusy() are not thread safe. The disk_busy() routine should be
called immediately before a command to the disk is sent, e.g.:
void
foostrategy(struct buf *bp)
{
[ . . . ]
mutex_enter(&sc->sc_dk_mtx);
disk_wait(&sc->sc_dk);
/* Put buffer onto drive's transfer queue */
mutex_exit(&sc->sc_dk_mtx);
foostart(sc);
}
void
foostart(struct foo_softc *sc)
{
[ . . . ]
/* Get buffer from drive's transfer queue. */
[ . . . ]
/* Build command to send to drive. */
[ . . . ]
/* Tell the disk framework we're going busy. */
mutex_enter(&sc->sc_dk_mtx);
disk_busy(&sc->sc_dk);
mutex_exit(&sc->sc_dk_mtx);
/* Send command to the drive. */
[ . . . ]
}
The routine disk_unbusy() performs some consistency checks, such as
ensuring that the calls to disk_busy() and disk_unbusy() are balanced.
It also performs the final steps of the metrics calcuation. A byte count
is added to the disk's running total, and if greater than zero, the
number of transfers the disk has performed is incremented. The third
argument read specifies the direction of I/O; if non-zero it means
reading from the disk, otherwise it means writing to the disk.
void
foodone(xfer)
struct foo_xfer *xfer;
{
struct foo_softc = (struct foo_softc *)xfer->xf_softc;
struct buf *bp = xfer->xf_buf;
long nbytes;
[ . . . ]
/*
* Get number of bytes transferred. If there is no buf
* associated with the xfer, we are being called at the
* end of a non-I/O command.
*/
if (bp == NULL)
nbytes = 0;
else
nbytes = bp->b_bcount - bp->b_resid;
[ . . . ]
mutex_enter(&sc->sc_dk_mtx);
/* Notify the disk framework that we've completed the transfer. */
disk_unbusy(&sc->sc_dk, nbytes,
bp != NULL ? bp->b_flags & B_READ : 0);
mutex_exit(&sc->sc_dk_mtx);
[ . . . ]
}
disk_isbusy() is used to get status of disk device it returns true if
device is currently busy and false if it is not. Like disk_wait(),
disk_busy() and disk_unbusy() it requires explicit locking from user
side.
CODE REFERENCES
The disk framework itself is implemented within the file
sys/kern/subr_disk.c. Data structures and function prototypes for the
framework are located in sys/sys/disk.h.
The NetBSD machine-independent SCSI disk and CD-ROM drivers use the disk
framework. They are located in sys/scsi/sd.c and sys/scsi/cd.c.
The NetBSD ccd, dm, and vnd drivers use the detachment capability of the
framework. They are located in sys/dev/ccd.c, sys/dev/vnd.c, and
sys/dev/dm/device-mapper.c.
SEE ALSO
ccd(4), dm(4), vnd(4), dksubr(9)
HISTORY
The NetBSD generic disk framework appeared in NetBSD 1.2.
AUTHORS
The NetBSD generic disk framework was architected and implemented by
Jason R. Thorpe <
[email protected]>.
FreeBSD 14.1-RELEASE-p8 March 5, 2017 FreeBSD 14.1-RELEASE-p8