Willem Riede wrote:
> On 2003.05.19 09:01, Douglas Gilbert wrote:
> 
>>Debug: sleeping function called from illegal context at include/linux/rwsem.h:43
>>Call Trace:
>> [<c011f92c>] __might_sleep+0x5c/0x5e
>> [<c011b2d8>] do_page_fault+0x78/0x4a8
>> [<e0823c4b>] idescsi_transfer_pc+0xfb/0x130 [ide_scsi]
> 
> 
> A page fault while in idescsi_transfer_pc?!
> What memory would be accessed that is allowed to be paged out?
> 
> By the way, I have never seen that problem. When ide-scsi fails for me, it
> is in the same way Randy reports. While my change improves mean-time-to-hang
> significantly on my machine, it obviously doesn't for Randy. Back to the
> drawing board :-(

Willem,
When I tried today, my test went for a while then
failed with a timeout and an abort lockup (which
you reported as fixed but I don't have that fix):
   hdb: irq timeout: status=0xd0 { Busy }
   ide-scsi: abort called for 330982
   hdb: ATAPI reset complete
  <<machine lockup, alt-sysrq inactive>>

Attached is a patch to idescsi_queue(). Won't fix the
problems we are seeing now. Changes:
   - returns 0 on error (not 1 which means "busy")
   - yield DID_NO_CONNECT for channel, id or lun
     invalid (this should fix the "responding to
     multiple lun" problem often seen in lk 2.4
   - memset the whole of pc and rq to zero

Doug Gilbert