Re: Linux 2.4.13-ac7

[Douglas Gilbert <dougg@torque.net>]

Alan,
I getting repeatable lockups on my SMP box (dual Celeron
Abit hack) after sg does direct IO in this new version.

This sequence using sg_rbuf from my sg3_utils package
(www.torque.net/sg) locks at the end of the transfer
just before returning:
   $ echo 1 > /proc/scsi/sg/allow_dio
   $ sg_rbuf -d -b=512 /dev/sg0
# /dev/sg0 is a fast disk.
# this executes multiple READ BUFFER commands each 512 KB

The box is still usable but the "sg_rbuf" process is
unkillable and ps locks just before it would have 
listed the damaged process. Alt-SysRq-T shows that
the sg_rbuf task in "D" state. The stack backtrace is:
  write_chan
  rwsem_down_write_failed
  ????
  ret_from_fork

This is with sg version 3.2.21 utilising alloc_kiovec_sz()
and friend. I didn't see this "feature" when I tested it
against ac4 before submitting it to you. Changing the "nbhs"
value from 0 to 256 doesn't make any difference. I have
just retested with ac4+sg3.1.21 and it works.

When normal double buffering is used (i.e. no "-d" switch
to the above sg_rbuf) there is no lockup.

Doug Gilbert