Re: libata & scsi error handling

Jeff Garzik wrote:

Brad Campbell wrote:

quoted

I think I have this timeout error issue pegged now.

I know this is both wrong, ugly and likely to cause internal kernel 
damage, but for the purpose of pegging what I think may be the culprit 
it works around the error nicely here

brad@srv:/usr/src$ diff -u 
temp/linux-2.6.8.1/drivers/scsi/libata-scsi.c 
linux-2.6.8.1/drivers/scsi/libata-scsi.c

--- temp/linux-2.6.8.1/drivers/scsi/libata-scsi.c       2004-08-14 

14:55:19.000000000 +0400

+++ linux-2.6.8.1/drivers/scsi/libata-scsi.c    2004-08-18 

01:04:11.000000000 +0400

@@ -213,6 +213,7 @@

        ap = (struct ata_port *) &host->hostdata[0];
        ap->ops->eng_timeout(ap);
+       host->host_failed--;

        DPRINTK("EXIT\n");
        return 0;

The issue is that the libata installed eh_strategy_handler does not 
complete the error as
scsi_unjam_host -> scsi_eh_abort_cmds -> scsi_eh_finish_cmd does.

Well, well, well.  If I had a libata Honorary Hacker merit badge, I 
would give it to you.

It is highly likely that your patch is doing the right thing.  Doug 
Ledford, 2.4.x SCSI maintainer, pointed out to me recently that my 2.4.x 
error handling code MUST update a couple variables, otherwise error 
handling would hang as you see.  The reason is that scsi_unjam_host(), 
on both 2.4.x and 2.6.x, is the only ->eh_strategy_handler until libata 
came along.

So, it is likely that there are a few details the scsi_unjam_host() 
performs, that needs to do too.

Thanks much for your excellent detective work, I'll see where to best 
put this change...

Jeff,
It probably doesn't rate any gold stars but while your patching
libata-scsi.c could you slip this fix in as well.

The patch is against lk 2.6.8.1 . The same patch is needed
(give or take fuzz) in lk 2.4.27 .

Changes:
    - send vendor, product and rev strings back for 36 byte
      INQUIRYs
    - set the additional length field to indicate 96 byte
      response is available

Doug Gilbert