Requests never returning?

["Alexandr Sandler" <rookie@linuxfrost.org>]

Hi list.

I am working on a logical volume manager (yep, another one) driver for
Linux. I implemented block device driver that used to receive requests and
to redirect them to scsi disks (by changing bh->b_rdev and bh->b_rsector). I
am experiencing a problem with my driver and I was wondering if anyone can
help me with it.

Notice that scsi disks sitting on SAN.

The driver, as it is, seems to be quite stable. I confident it can run for
weeks under the most heavy stress test, despite I only tried 48 hours tests.
But sometimes, very rarely, something goes wrong and process doing I/O (on
the logical volume + file system) getting stacked. Call trace (from kdb)
looks like this. Process trying to do generic_file_read() (or ..._write())
and, eventually, getting into lock_page() - this is where it's getting
stacked. What seems to be obvious (for me - correct me if I am wrong), is
that process supposed to succeed locking page, but buffer_head using this
page is never getting it's b_end_io() and here we go - it's stacked.

This is happening, usually, when one or several computers on SAN getting
restarted and as a result making a lot of noise (LIPs) on fiber. And it's
even more strange because I made special tests to reveal the conditions
needed for this to happen. I caused hundreds of LIPs, but Linux box was
stable and test was running - it was couple of month ago. And now it
happened once again - when I rebooted some lame NetWare host.

I think the problem is not with my driver (yes, I know it's better to be
paranoiac in this things), despite I never tried to do some stress tests
with scsi disks - without my driver.

So, my question is like this. Is there any condition, under which I/O
requests may not return (even with error), when working with SAN?

Here are some specs:
It's Pentium III 800 MHz UP with 256Mb of RAM. Kernel 2.4.16 (as far as I
remember there was no major changes and bugfixes in scsi layer in 2.4.17 and
2.4.18, so I think it doesn't really matter if I use 2.4.16 or 2.4.18 - once
again, correct me if I am wrong) . I am using RedHat 7.1 distro. Two Qlogic
2200 optical HBAs powered by qlogic 4.27b driver (this part may seem to be
crucial - I never tried to work with other versions of their drivers. On the
other hand, I never had a reason to doubt stability of this version). SAN
sitting on Gadzoox Capellix 3000 - arbitrated loop (I got the same behavour
with Brocade switch - don't remember the model).

Thanks in advance for any help.

Alexandr Sandler.