Linux 2.3.0

[davej-history.git] / Documentation / scsi-generic.txt

            Notes on Linux's SG driver version 2.1.30
            -----------------------------------------
                                                        990328

Introduction
============
These are notes on the Linux SCSI generic packet device driver (sg)
describing version 2.1.30 . The original driver was written by Lawrence
Foard and has remained in place with minimal changes since circa 1992.
Version 2 of this driver remains backward compatible (binary and
source **) with the original. It adds scatter gather, command queuing,
per file descriptor sequencing, asynchronous notification and better
error reporting.

Sg is one of the four "high level" SCSI device drivers along with
sd, st and sr (disk, tape and CDROM respectively). Sg is more generalized
(but lower level) than its sibling and tends to be used on SCSI devices
that don't fit into the already serviced categories. Thus sg is used for
scanners, cd writers and reading audio cds amongst other things.

The interface and usage of the original sg driver has been documented
by Heiko Eissfeldt in a HOWTO called SCSI-Programming-HOWTO. My copy
of the document is version 1.5 dated 7th May 1996. It can found at
ftp://sunsite.unc.edu/pub/Linux/docs/HOWTO/SCSI-Programming-HOWTO .
Amongst other things it has a lot of tables from the SCSI-2 standard
that are very useful for programming this interface.

** It is possible to write applications that perform differently
depending on whether they are using the original or this version of
the sg device driver. The author is not aware of any useful applications
that have problems with version 2 (yet).


Architecture
============
The SCSI generic packet device driver (sg) is a character based device.
It is one of the four high level device driver in the SCSI sub-system;
the others are sd (for direct-access devices - disks), st (for tapes)
and sr (for data cdroms). The other three devices are block devices.

The unifying layer of the SCSI sub-system in the so-called mid-level.
Below that are all the drivers for the various adapters supported by
Linux.

Since sg is a character device it supports the traditional Unix
system calls of open(), close(), read(), write() and ioctl(). Two other
related system calls: poll() and fcntl() are added to this list and
how they interact with the sg device driver is documented later.
 
An SG device is accessed by write()ing SCSI commands plus any associated 
outgoing data to it; the resulting status codes and any incoming data are
then obtained by a read() call. The device can be opened O_NONBLOCK
(non-blocking) and poll() used to monitor its progress. The device may be
opened O_EXCL which excludes other "sg" users from this device (but not 
"sd", "st" or "sr" users). The buffer given to the write() call is made
up as follows:
        - struct sg_header image (see below)
        - scsi command (6, 10 or 12 bytes long)
        - data to be written to the device (if any)

The buffer received from the corresponding read() call contains:
        - struct sg_header image (check status/errors + sense_buffer)
        - data read back from device (if any)

The given SCSI command has its LUN field overwritten by the LUN value of
the associated sg device that has been open()ed.


sg_header
=========
This is the name of the control structure that conveys information
about the length of data to be read/written by the associated SCSI
command. It also conveys error and status information from the
read() call. An instance of this structure is the first thing that
is placed in the data buffers of both write() and read().

In its original form it looked like this:
struct sg_header {
    int pack_len;
    int reply_len;
    int pack_id;
    int result;
    unsigned int twelve_byte:1;
    unsigned int other_flags:31;
    unsigned char sense_buffer[16];
}; /* this structure is 36 bytes long */

The 'pack_len' is bizzare and ends up having the 'reply_len' put in it
(perhaps it had a use at some stage). 

The 'reply_len' is the length of the data the corresponding read()
will/should request (including the sg_header). 

The 'pack_id' is not acted upon by the sg device driver but is conveyed
back to the corresponding read() so it can be used for sequencing by an
application. 

The 'result' is also bizzare, turning certain types of host codes it 0 (no
error), EBUSY or EIO. With better error reporting now available, the
'result' is best ignored.

The 'twelve_byte' field overrides the internal SCSI command length "guessing"
algorithm for group 6 and 7 commands (ie when 1st byte >= 0xc0) and forces
a command lenth of 12 bytes.
The command length "guessing" algorithm is as follows:
Group:  0    1    2    3    4    5    6    7
Length: 6   10   10   12   12   12   10   10

'other_flags' was originally documented as "not used" but some current
applications assume it has 0 placed in it.

The 'sense_buffer' is the first 16 bytes of SCSI sense buffer that is
returned when the target returns a SCSI status code of CHECK_CONDITION
or COMMAND_TERMINATED [or (driver_status & DRIVER_SENSE) is true]. This
buffer should be at least 18 bytes long and arguably 32 bytes; unfortunately
this is unlikely to happen in the 2.2.x series of kernels.

The new sg_header offered in this driver is:
#define SG_MAX_SENSE 16
struct sg_header
{
    int pack_len;    /* [o] reply_len (ie useless) ignored as input */
    int reply_len;   /* [i] max length of expected reply (inc. sg_header) */
    int pack_id;     /* [io] id number of packet (use ints >= 0) */
    int result;      /* [o] 0==ok, else (+ve) Unix errno code (e.g. EIO) */
    unsigned int twelve_byte:1;
        /* [i] Force 12 byte command length for group 6 & 7 commands  */
    unsigned int target_status:5;   /* [o] scsi status from target */
    unsigned int host_status:8;     /* [o] host status (see "DID" codes) */
    unsigned int driver_status:8;   /* [o] driver status+suggestion */
    unsigned int other_flags:10;    /* unused */
    unsigned char sense_buffer[SG_MAX_SENSE]; /* [o] when target_status is
               CHECK_CONDITION or COMMAND_TERMINATED this is output. */
};      /* This structure is 36 bytes long on i386 */

Firstly the new header is binary compatible with the original. This is
important for keeping existing apps working without recompilation.

Only those elements (or fields) that are new or in some way different
from the original are documented below.

'pack_id' becomes input to a read() when ioctl(sg_fd, SG_SET_FORCE_PACK_ID,
&one) is active. A 'pack_id' of -1 is interpreted as fetch the oldest
waiting packet; any other value will cause the read() to wait (or yield
EAGAIN) until a packet with that 'pack_id' becomes available. In all cases
the value of 'pack_id' available after a read() is the value given to that
variable in the prior, corresponding write().

The 'target_status' field is always output and is the (masked and shifted
1 bit right) SCSI status code from the target device. The allowable
values are (found in <scsi/scsi.h>):
/* N.B. 1 bit offset from usual SCSI status values */
#define GOOD                 0x00
#define CHECK_CONDITION      0x01  
#define CONDITION_GOOD       0x02
#define BUSY                 0x04
#define INTERMEDIATE_GOOD    0x08
#define INTERMEDIATE_C_GOOD  0x0a
#define RESERVATION_CONFLICT 0x0c
#define COMMAND_TERMINATED   0x11
#define QUEUE_FULL           0x14
When the 'target_status' is CHECK_CONDITION or COMMAND_TERMINATED the
'sense_buffer' is output. Note that when (driver_status & DRIVER_SENSE)
is true then the 'sense_buffer' is also output (this seems to occur when
the scsi ide emulation is used). When the 'sense_buffer' is output the 
SCSI Sense Key can be found at (sense_buffer[2] & 0x0f) .

The 'host_status' field is always output and has the following values
whose "defines" are not visible outside the kernel (unfortunately):
#define DID_OK          0x00 /* NO error                                */
#define DID_NO_CONNECT  0x01 /* Couldn't connect before timeout period  */
#define DID_BUS_BUSY    0x02 /* BUS stayed busy through time out period */
#define DID_TIME_OUT    0x03 /* TIMED OUT for other reason              */
#define DID_BAD_TARGET  0x04 /* BAD target.                             */
#define DID_ABORT       0x05 /* Told to abort for some other reason     */
#define DID_PARITY      0x06 /* Parity error                            */
#define DID_ERROR       0x07 /* Internal error                          */
#define DID_RESET       0x08 /* Reset by somebody.                      */
#define DID_BAD_INTR    0x09 /* Got an interrupt we weren't expecting.  */
#define DID_PASSTHROUGH 0x0a /* Force command past mid-layer            */
#define DID_SOFT_ERROR  0x0b /* The low level driver just wish a retry  */

The 'driver_status' field is always output. When ('driver_status' &
DRIVER_SENSE) is true the 'sense_buffer' is also output. The following
values whose "defines" are not visible outside the kernel (unfortunately)
can occur:
#define DRIVER_OK           0x00 /* Typically no suggestion */
#define DRIVER_BUSY         0x01
#define DRIVER_SOFT         0x02
#define DRIVER_MEDIA        0x03
#define DRIVER_ERROR        0x04
#define DRIVER_INVALID      0x05
#define DRIVER_TIMEOUT      0x06
#define DRIVER_HARD         0x07
#define DRIVER_SENSE        0x08
/* above status 'or'ed with one of the following suggestions */
#define SUGGEST_RETRY       0x10
#define SUGGEST_ABORT       0x20
#define SUGGEST_REMAP       0x30
#define SUGGEST_DIE         0x40
#define SUGGEST_SENSE       0x80

'other_flags' still remains as a 10 bit field, so code that places 0 in it
will still be happy. It is not used.


memory
======
Memory is a scarce resource in any computer. Sg needs to reserve memory
suitable for DMA roughly equal in size to the maximum of the write and
read data buffers for each packet. This DMA memory is obtained at the time
of a write() and released when the corresponding read() is called (although
if memory is tight it may be using the buffer reserved by the open() ).

Linux obtaining memory a challenge for several reasons. The memory pool
that sg uses is in common with all other device drivers and all user
processes. In this environment the only way to 99.9% guarantee a driver
will have memory in Linux is to build it into the kernel (ie not as a
module) and then reserve it on initialization before user processes get
a chance. [Of course, another driver initialized before sg could take
all available memory ...] Another problem is the biggest contiguous
chunk of memory that can be obtained from the kernel is 32 * PAGE_SIZE
(which is 128KBytes on i386). As memory gets "splintered" there is a good
chance that buffers won't be available (my machine has 64 MBytes of RAM
and has 3 available at the moment).

The original sg driver used the following technique: grab a SG_BIG_BUFF
sized buffer at driver initialization and use it for all requests greater
than PAGE_SIZE (4096 bytes on i386). By default SG_BIG_BUFF is set to
32 KBytes in the origianl driver but many applications suggest that the
user increases this number. Linux limits the biggest single buffer of
this type to 32 * PAGE_SIZE (128KBytes on i386). Unfortunately if the
sg driver is a module then there is a high chance a contiguous block of
that large size will not be available at module initialization.

The author has found no "silver bullet" solution but uses multiple
techniques hoping that at least one is able provide memory at the critical
time. Listed below are some of these techniques:
        - use scatter gather: then instead of one large buffer needing to
          be found, multiple smaller buffer can be used
        - use memory above the 16MByte level: the original driver limited
          itself to obtaining memory below the 16MByte level (on the i386)
          due to the shortcomings of DMA on ISA adapters. Yet more and more
          people use PCI adapters that don't have this problem. So make
          the decision based on the capabilities of the host adpater
          associated with the current SCSI device
        - reserve some memory at open() for emergencies but otherwise
          fetch and release it on a per packet basis
        - if the kernel is short of memory then dip into the SCSI DMA
          pool (maintained by the mid-level driver) to a limited amount



System Calls
============
What follows are descriptions of the characteristics of the standard
Unix operating system calls when applied to a SCSI generic device
using this version of the device driver.

open
----
The filename should be an 'sg' device such as
/dev/sg[a-z]
/dev/sg[0,1,2,...]
or a symbolic link to one of these. [Devfs has its own sub-directory for
sg devices.] It seems as though SCSI devices are allocated to sg minor
numbers in the same order as they appear in 'cat /proc/scsi/scsi'.
Sg is a "character" based Linux device driver. This means it has an
open/close/read/write/ioctl type interface.

Flags can be either O_RDONLY or O_RDWR or-ed with either
O_EXCL          waits for other opens on sg device to be closed before
                proceeding. If O_NONBLOCK is set then yields EBUSY when
                someone else has the sg device open. The combination of
                O_RDONLY and O_EXCL is disallowed.
O_NONBLOCK      Sets non-blocking mode. Calls that would otherwise block
                yield EAGAIN (eg read() ) or EBUSY (eg open() ).

The original version of sg did not allow the O_RDONLY (yielding a EACCES
error). This version allows it for accessing ioctls (e.g. doing an sg
device scan with the SG_GET_SCSI_ID ioctl) but write()s will not be
allowed.

By default, sequencing is per file descriptor in this version of sg. This
means, for example that 2 processes can independently manipulate the same
sg device at the same time. This may or may not make sense depending on
the application: 2 processes (logically) reading from the same direct access
device (ie a disk) is ok while running 2 instances of cd writing software
on the same device at the same time probably wouldn't be a good idea. The
previous version of sg supported only per device sequencing and this can
still be selected with the SG_SET_MERGE_FD,1 ioctl().

The driver will attempt to reserve SG_SCATTER_SZ bytes (32KBytes in the
current sg.h) on open() for "emergency" situations. If this is unavailable
it will halve its request and try again. It gives up if PAGE_SIZE bytes
(4096 bytes on i386) cannot be obtained so no memory is reserved. In this
case open() will still return successfully. The actual amount of memory
reserved can be found with the SG_GET_RESERVED_SIZE ioctl().

Returns a file descriptor if >= 0 , otherwise -1 implies an error.

Error codes (value in 'errno' after -1 returned):
ENODEV          sg not compiled into kernel or the kernel cannot find the
                sg module (or it can't initialize itself (low memory??))
ENXIO           either scsi sub-system is currently processing some error
                (eg doing a device reset) or the sg driver/module removed
                or corrupted
EBUSY           O_NONBLOCK set and some user of this sg device has O_EXCL
                set while someone is already using this device
EINTR           while waiting for an "exclusive" lock to clear, a signal
                is received, just try again ...        
ENOMEM          An attempt to get memory to store this open's context
                failed (this was _not_ a request to reserve DMA memory)
EACCES          An attempt to use both O_RDONLY and O_EXCL


write
-----
Even though sg is a character-based device driver it sends and receives
packets to/from the associated scsi device. Write() is used to send a
packet containing 2 mandatory parts and 1 optional part. The mandatory
parts are:
  - a control block (an instance of struct sg_header)
  - a SCSI command (6, 10 or 12 bytes long)
The optional part is:
  - outgoing data (eg if a SCSI write command is being sent)
These should appear as one contiguous string in the buffer given to
write() in the above order with no pad characters.

If a write() accepts this packet then at some later time the user should
call a read() to get the result of the SCSI command. The previous sg
driver enforced a strict write()/read()/write()/read() regime so that a
second write() would block until first read() was finished. This sg
driver relaxes that condition and thereby allows command queuing
(limit is SG_MAX_QUEUE (16) outstanding packets per file descriptor).
However, for backward compatibility, command queuing is turned off
by default (#define SG_DEF_COMMAND_Q 0 in sg.h). This can be changed
via the the SG_SET_COMMAND_Q ioctl() [or by recompiling after changing
the above define to 1].

In this sg driver a write() should return more or less immediately.

Returns number of bytes written if > 0 , otherwise -1 implies an error.

Error codes (value in 'errno' after -1 returned):
ENXIO           either scsi sub-system is currently processing some error
                (eg doing a device reset) or the sg driver/module removed
                or corrupted
EACCES          opened with RD_ONLY flag
EIO             incoming buffer too short. It should be at least (6 +
                sizeof(struct sg_header))==42 bytes long
EDOM            a) command queuing off: a packet is already queued
                b) command queuing on: too many packets queued 
                   (SG_MAX_QUEUE exceeded)
EAGAIN          SCSI mid-level out of command blocks (rare), try again.
                This is more likely to happen when queuing commands,
                so wait a bit (eg usleep(10000) ) before trying again
ENOMEM          can't get memory for DMA. Take evasive action ...
                (see section on memory)


read
----
Read() is used to receive a packet containing 1 mandatory part and 1 
optional part. The mandatory part is:
  - a control block (an instance of struct sg_header)
The optional part is:
  - incoming data (eg if a SCSI read command was sent by earlier write() )
The buffer given to a read() and its corresponding count should be
sufficient to accommodate this packet to avoid truncation. Truncation has
occurred if count < sg_header::replylen .

By default, read() will return the oldest packet queued up. If the 
SG_SET_FORCE_PACK_ID,1 ioctl() is active then read() will attempt to
fetch the packet whose pack_id (given earlier to write()) matches the
sg_header::pack_id given to this read(). If not available it will either
wait or yield EAGAIN. As a special case, -1 in sg_header::pack_id given
to read() will match the oldest packet.


Returns number of bytes read if > 0 , otherwise -1 implies an error.
Unfortunately the return value in the non-error case is simply the
same as the count argument. It is not the actual number of bytes
DMA-ed by the SCSI device. This driver is currently unable to provide
such an underrun indication.

Error codes (value in 'errno' after -1 returned):
ENXIO           either scsi sub-system is currently processing some error
                (eg doing a device reset) or the sg driver/module removed
                or corrupted
EAGAIN          either no waiting packet or requested packet is not
                available while O_NONBLOCK flag was set
EINTR           while waiting for a packet, a signal is received, just
                try again ...        
EIO             if the 'count' given to read() is < sizeof(struct sg_header)
                and the 'result' element in sg_header is non-zero. Not a
                recommended error reporting technique


close
-----
Preferably a close() should be done after all issued write()s have had
their corresponding read() calls completed. Unfortunately this is not
always possible. The semantics of close() in Unix are to return more
or less immediately (ie not wait on any event) so the driver needs to
arrange to an orderly cleanup of those packets that are still "in
flight".

A process that has an open file descriptor to an sg device may be aborted
(eg by a kill signal). In this case, the kernel automatically calls close 
(which is called 'sg_release()' in the version 2 driver) to facilitate
the cleanup mentioned above.

A problem persists in version 2.1.8 if the sg driver is a module and is
removed while packets are still "in flight". Hopefully this will be soon
fixed.

Returns 0 if successful, otherwise -1 implies an error.

Error codes (value in 'errno' after -1 returned):
ENXIO           sg driver/module removed or corrupted

ioctl (sg specific)
-------------------
Ken Thompson (or perhaps some other Unix luminary) described ioctl() as 
the "garbage bin of Unix". This driver compounds the situation by adding
around 18 more commands. These commands either yield state information (10
of them), change the driver's characteristics (8 of them) or allow direct
communication with the common SCSI mid-level driver.

Those commands with an appended "+" are new in version 2.

Those commands with an appended "W" are only accessible from file
descriptors opened with O_RDWR. They will yield EACCES otherwise.

SG_GET_TIMEOUT:
Ignores its 3rd argument and _returns_ the timeout value (which will be
>= 0 ). The unit of this timeout is "jiffies" which are currently 10
millisecond intervals on i386 (less on an alpha). Linux supplies
a manifest constant HZ which is the number of "jiffies" in 1 second.

SG_SET_TIMEOUT:
Assumes 3rd argument points to an int containing the new timeout value
for this file descriptor. The unit is a "jiffy". Packets that are
already "in flight" will not be effected. The default value is set
on open() and is SG_DEFAULT_TIMEOUT (defined in sg.h).

SG_EMULATED_HOST:
Assumes 3rd argument points to an int and outputs a flag indicating
whether the host (adapter) is connected to a real SCSI bus or is
emulated one (eg ide-scsi device driver). A value of 1 means emulated
while 0 is not.

SG_SET_FORCE_LOW_DMA +:
Assumes 3rd argument points to an int containing 0 or 1. 0 (default)
means sg decides whether to use memory above 16 Mbyte level (on i386)
based on the host adapter being used by this SCSI device. Typically
PCI SCSI adapters will indicate they can DMA to the whole 32 bit address
space. 
If 1 is given then the host adapter is overridden and only memory below
the 16MB level is used for DMA. A requirement for this should be
extremely rare. If the "reserve" buffer allocated on open() is not in
use then it will be de-allocated and re-allocated under the 16MB level
(and the latter operation could fail yielding ENOMEM).
Only the current file descriptor is effected.

SG_GET_LOW_DMA +:
Assumes 3rd argument points to an int and places 0 or 1 in it. 0
indicates the whole 32 bit address space is being used for DMA transfers
on this file descriptor. 1 indicates the memory below the 16MB level
(on i386) is being used (and this may be the case because the host
adapters setting has been overridden by SG_SET_FORCE_LOW_DMA,1 .

SG_GET_SCSI_ID +:
Assumes 3rd argument is pointing to an object of type Sg_scsi_id and
populates it. That structure contains ints for host_no, channel,
scsi_id, lun and scsi_type. Most of this information is available from
other sources (eg SCSI_IOCTL_GET_IDLUN and SCSI_IOCTL_GET_BUS_NUMBER)
but tends to be awkward to collect.

SG_SET_FORCE_PACK_ID +:
Assumes 3rd argument is pointing to an int. 0 (default) instructs read()
to return the oldest (written) packet if multiple packets are
waiting to be read (when command queuing is being used). 
1 instructs read() to view the sg_header::pack_id as input and return the
oldest packet matching that pack_id or wait until it arrives (or yield
EAGAIN if O_NONBLOCK is in force). As a special case the pack_id of -1
given to read() in the mode will match the oldest packet.
Only the current file descriptor is effected by this command.

SG_GET_LOW_DMA +:
Assumes 3rd argument points to an int and places the pack_id of the
oldest (written) packet in it. If no packet is waiting to be read then
yields -1.

SG_GET_NUM_WAITING +:
Assumes 3rd argument points to an int and places the number of packets
waiting to be read in it.

SG_GET_SG_TABLESIZE +:
Assumes 3rd argument points to an int and places the maximum number of
scatter gather elements supported by the host adapter. 0 indicates that
the adapter does support scatter gather.

SG_SET_RESERVED_SIZE +W:
This is not currently implemented. It is intended for reserving either a
large buffer or scatter gather list that will be available until the
current file descriptor is closed. The requested amount of memory may
not be available so SG_GET_RESERVED_SIZE should be used after this call
to see how much was reserved. (EBUSY error possible)

SG_GET_RESERVED_SIZE +:
Assumes 3rd argument points to an int and places the size in bytes of
the DMA buffer reserved on open() for emergencies. If this is 0 then it
is probably not wise to attempt on operation like burning a CD on this
file descriptor.

SG_SET_MERGE_FD +W:
Assumes 3rd argument is pointing to an int. 0 (the default) causes all
subsequent sequencing to be per file descriptor. 1 causes all subsequent
sequencing to be per device. If this command tries to change the current
state and the is one or more _other_ file descriptors using this sg
device then an EBUSY error occurs. Also if this file descriptor was not
open()ed with the O_RDWR flag then an EACCES error occurs.
Per device sequencing was the original semantics and allowed, for example
different processes to "share" the device, one perhaps write()ing with
the other one read()ing. This command is supplied if anyone needs those
semantics. Per file descriptor sequencing, perhaps with the usage of
the O_EXCL flag, seems more sensible.

SG_GET_MERGE_FD +:
Assumes 3rd argument points to an int and places 0 or 1 in it. 0 implies
sequencing is per file descriptor. 1 implies sequencing is per device
(original sg driver's semantics).

SG_SET_COMMAND_Q +:
Assumes 3rd argument is pointing to an int. 0 (current default, set by
SG_DEF_COMMAND_Q in sg.h) disables command queuing. Attempts to write()
a packet while one is already queued will result in a EDOM error.
1 turns command queuing on.
Changing the queuing state only effects write()s done after the change.
Only the current file descriptor is effected by this command.

SG_GET_COMMAND_Q +:
Assumes 3rd argument points to an int and places 0 or 1 in it. 0 implies
that command queuing is off on this file descriptor. 1 implies command
queuing is on.

SG_SET_DEBUG +:
Assumes 3rd argument is pointing to an int. 0 (default) turns debugging
off. Values > 0 cause the SCSI sense buffer to be decoded and output
to the console/log when a SCSI device error occurs. Values > 8 cause
the current sg device driver's state to be output to the console/log
(this is a "one off" effect).
If you need a _lot_ of the SCSI sub-system debug information (mainly from
the mid-level) then try 'echo "scsi dump 0" > /proc/scsi/scsi' and lots of
debug will appear in your console/log.

ioctl (in common with sd, st + sr)
----------------------------------
The following ioctl()s can be called from any high-level scsi device
driver (ie sd, st, sr + sg). Access permissions may differ a bit from
one device to another, the access information given below is specific to
the sg device driver.

SCSI_IOCTL_GET_IDLUN:
SCSI_IOCTL_GET_BUS_NUMBER:

SCSI_IOCTL_SEND_COMMAND:  W
If open()ed O_RDONLY yields an EACCESS error. Otherwise is forwarded onto
the SCSI mid-level driver for processing.
Don't know much about this one but it looks pretty powerful and
dangerous. Some comments says it is also deprecated.

<any_command_not matching_above>:  W
If open()ed O_RDONLY yields an EACCESS error. Otherwise is forwarded onto
the SCSI mid-level driver for processing.


poll
----
This is a native call in Linux 2.2 but most of its capabilities are available
through the older select() call. Given a choice poll() should probably be
used. Typically poll() is used when a sg scsi device is open()ed O_NONBLOCK
for polling; or alternatively with asynchronous notification using the
fcntl() system call (below) and the SIGPOLL (aka SIGIO) signal.
Only if something drastically is wrong (eg file handle gone stale) will
POLLERR ever be set. POLLPRI, POLLHUP and POLLNVAL are never set.
POLLIN is set when there is one or more packets waiting to be read.
When POLLIN is set it implies that a read() will not block (or yield
EAGAIN in non-blocking mode) but return a packet immediately.
POLLOUT (aka POLLWRNORM) is set when write() is able to accept a packet
(ie will _not_ yield an EDOM error). The setting of POLLOUT is effected
by the SG_SET_COMMAND_Q state: if the state is on then POLLOUT will remain
set until the number of queued packets reaches SG_MAX_QUEUE, if the
state is off then POLLOUT is only set when no packets are queued.
Note that a packet can be queued after write()ing but not available to be
read(); this typically happens when a SCSI read command is issued while
the data is being retreaved.
Poll() is per file descriptor unless SG_SET_MERGE_FD is set in which case
it is per device.


fcntl
-----
There are several uses for this system call in association with a sg
file descriptor. The first pseudo code shows code that is useful for
scanning the sg devices, taking care not to be caught in a wait for
an O_EXCL lock by another process, and when the appropriate device is
found switching to normal blocked io. A working example of this logic
is in the sg_scan.c utility program.

open("/dev/sga", O_RDONLY | O_NONBLOCK)
/* check device, EBUSY means some other process has O_EXCL lock on it */
/* one the device you want is found then ... */
flags = fcntl(sg_fd, F_GETFL)
fcntl(sg_fd, F_SETFL, flags & (~ O_NONBLOCK))
/* for simple apps is is easier to use normal blocked io */


Some work has to be done in Linux to set up for asynchronous notification.
This is a non-blocking mode of operation in which when the driver receives
data back from a device so that a read() can be done, it sends a SIGPOLL
(aka SIGIO) signal to the owning process. A working example of this logic
is in the sg_poll.c test program.

sigemptyset(&sig_set)
sigaddset(&sig_set, SIGPOLL)
sigaction(SIGPOLL, &s_action, 0)
fcntl(sg_fd, F_SETOWN, getpid())
flags = fcntl(sg_fd, F_GETFL);
fcntl(sg_fd, F_SETFL, flags | O_ASYNC)


Utility and Test Programs
=========================