2.6.9-rc2-mm4 BIO's still broken

2.6.9-rc2-mm4 BIO's still broken

[jmerkey]

I have more information on the problem with bio requests.  I am seeing the bi_size 
value return through bi_end_io early with an odd size if the interface is passed 
an unaligned 4K write.  What's busted here is that bio_add_page accepts the 
4K unaligned write request, then the callback from the SCSI layer calls back 
with a partial compleation with the bi_size field set to the value of 0x1FE (????)
and no other callback is received.  What's busted here is if you use the 
recommended logic of 

if (bio->bi_size)
    return 1;

then you never get the completed callback and the IO request just sits off in left 
field and the driver never returns any error status through the callback interface.
I am also still seeing the disappearing pages and after tracking through the code, 
I am certain they are related since I am not getting any callbacks from the driver 
layer after I receive the first end_io callback with bi_size set.  

Also, my drdos.com address is blocked from posting to the list again, so I am using 
comcast.net.  What's up with this?

Jeff

Re: 2.6.9-rc2-mm4 BIO's still broken

[Jens Axboe]

(please wrap your out going emails, the lines are way too long)

On Fri, Oct 01 2004, jmerkey@comcast.net wrote:
> 
> I have more information on the problem with bio requests.  I am seeing
> the bi_size value return through bi_end_io early with an odd size if
> the interface is passed an unaligned 4K write.  What's busted here is
> that bio_add_page accepts the 4K unaligned write request, then the
> callback from the SCSI layer calls back with a partial compleation
> with the bi_size field set to the value of 0x1FE (????) and no other
> callback is received.  What's busted here is if you use the
> recommended logic of 
> 
> if (bio->bi_size) return 1;
> 
> then you never get the completed callback and the IO request just sits
> off in left field and the driver never returns any error status
> through the callback interface.  I am also still seeing the
> disappearing pages and after tracking through the code, I am certain
> they are related since I am not getting any callbacks from the driver
> layer after I receive the first end_io callback with bi_size set.  

Please send a test case, thanks.

-- 
Jens Axboe