sg timeout problem

sg timeout problem

[Norman Diamond]

struct sg_io_hdr sg_io;
[...]
sg_io.timeout = UINT_MAX;
[...]
ioctl(fd, SG_IO, &sg_io);

As far as I can tell, if sg is running on top of LIBATA, the infinite
timeout is honoured.

But if sg is running on top of an actual SCSI driver, someone aborts the
ioctl after 3 seconds.  The ioctl returns success even though the operation
didn't complete.  I tried this with two Adaptec drivers and one NinjaSCSI 
driver with 100% repro.

What am I missing?  I need the call to run to completion. 

--------------------------------------
Power up the Internet with Yahoo! Toolbar.
http://pr.mail.yahoo.co.jp/toolbar/

Re: sg timeout problem

[Alan Cox]

> But if sg is running on top of an actual SCSI driver, someone aborts the
> ioctl after 3 seconds.  The ioctl returns success even though the operation
> didn't complete.  I tried this with two Adaptec drivers and one NinjaSCSI 
> driver with 100% repro.
> 
> What am I missing?  I need the call to run to completion. 

If you set it to the max of a signed int does it behave, or to a large
value thats well under the limit. That will help tell if there is a
problem with sign handling in the scsi code somewhere or it is getting
multiplied up for some reason.

Re: sg timeout problem

[Norman Diamond]

>> But if sg is running on top of an actual SCSI driver, someone aborts the
>> ioctl after 3 seconds.  The ioctl returns success even though the
>> operation didn't complete.  I tried this with two Adaptec drivers and one
>> NinjaSCSI driver with 100% repro.
>>
>> What am I missing?  I need the call to run to completion.
>
> If you set it to the max of a signed int does it behave, or to a large
> value thats well under the limit. That will help tell if there is a
> problem with sign handling in the scsi code somewhere or it is getting
> multiplied up for some reason.

I have the impression now that the drives are lying and it's not the fault
of the drivers or cards.  The same problem occured with a fourth card and a
non-Linux system.

Anyway I also tried changing the timeout to 86400000 milliseconds (one day)
and it made no difference.  So again I stop blaming the drivers.

--------------------------------------
Power up the Internet with Yahoo! Toolbar.
http://pr.mail.yahoo.co.jp/toolbar/

Re: sg timeout problem

[Douglas Gilbert]

Norman Diamond wrote:
>>> But if sg is running on top of an actual SCSI driver, someone aborts the
>>> ioctl after 3 seconds.  The ioctl returns success even though the
>>> operation didn't complete.  I tried this with two Adaptec drivers and 
>>> one
>>> NinjaSCSI driver with 100% repro.
>>>
>>> What am I missing?  I need the call to run to completion.
>>
>> If you set it to the max of a signed int does it behave, or to a large
>> value thats well under the limit. That will help tell if there is a
>> problem with sign handling in the scsi code somewhere or it is getting
>> multiplied up for some reason.
> 
> I have the impression now that the drives are lying and it's not the fault
> of the drivers or cards.  The same problem occured with a fourth card and a
> non-Linux system.
> 
> Anyway I also tried changing the timeout to 86400000 milliseconds (one day)
> and it made no difference.  So again I stop blaming the drivers.

Norman,
I'm not sure what version or distro of Linux you have
but I just did this quick test on Ubuntu 9.04 in which
a kernel tick seems to be around 4 milliseconds:

$  uname -a
Linux zink 2.6.28-11-generic #42-Ubuntu SMP Fri Apr 17 01:57:59 UTC 2009 i686 
GNU/Linux
$
$ modprobe scsi_debug delay=1000
$ lsscsi -g
[7:0:0:0]    disk    Linux    scsi_debug       0004  /dev/sdb   /dev/sg2

$ time sg_readcap /dev/sg2
Read Capacity results:
    Last logical block address=16383 (0x3fff), Number of blocks=16384
    Logical block length=512 bytes
Hence:
    Device size: 8388608 bytes, 8.0 MiB, 0.01 GB

real	0m3.999s
user	0m0.000s
sys	0m0.000s
$


See 'modinfo scsi_debug' for more about the arguments to the
scsi_debug driver. If the delay is too long then the kernel
may timeout registering the pseudo device. I picked 4 seconds
because you noted a failure above 3 seconds. The timeout on
sg_inq is set at 60 seconds.

Doug Gilbert