Re: [RFC] [PATCH] Potential bug fix for blk_tag_queue

[Jens Axboe <jens.axboe@oracle.com>]

(please don't top post!)

On Tue, Aug 26 2008, Abhijeet Joglekar wrote:
> Ok, removing that works too. 
> 
> Shouldn't the BUG_ON be checking upto bqt->real_max_depth instead of
> bqt->depth?  
> 
> BUG_ON(find_first_bit(bqt->tag_map, bqt->real_max_depth) <
>                       bqt->real_max_depth);
> 
> 
> In cases where the tag map gets resized to a value less than the
> originally allocated tag map, blk_queue_resize_tags seems to be setting
> bqt->max_depth to the new_depth. 
> 
> There could still be outstanding tags max_depth <= t <= real_max_depth
> which might not get freed, which the BUG_ON will not capture if it
> checks against max_depth.

max_depth is fine, it's the safer choice. For the real_max_depth to
potentially catch any extra offenders, you would have to do a resize to
a larger depth, get bunch of IO issued, then resize down and shut down
the queue. Using real_max_depth should work as well, but then you have
to audit the ending of tags > max/real_max_depth. It looks ok, though.

-- 
Jens Axboe