Re: [PATCH v4] drivers/block/mtip32xx: Adding new driver mtip32xx

[Jens Axboe <jaxboe@fusionio.com>]

On 2011-09-13 18:46, Eric Seppanen wrote:
> On Tue, Sep 13, 2011 at 5:49 AM, Christoph Hellwig <hch@infradead.org> wrote:
>> Btw, there is another _huge_ issue with the driver, and that is the
>> lack of any internal queueing.  Remember the make_request interface is
>> an extremly thin layer (or rather the lack of it) below the filesystem.
>>
>> So for example if eh_active is non-zero you return -EBUSY to the
>> filesystems.  That's an error code it a) doesn't recognize and b)
>> couldn't handle even if it did.  Similarly mtip_hw_get_scatterlist
>> simply blocks if no tag is currently available instead of queueing
>> it up.
> 
> Just out of curiosity, why is blocking on no-tag-available a bad
> thing?  How is it any different than the blocking that will occur when
> a request queue is full?  When the hardware queue depth is bigger than
> that of a request queue, what extra benefit does queuing give?

The blocking for a free tag is fine, the actual implementation is
definitely not optimal (using a rw semaphore with count initialized to
the tag depth, ugh). You'll need to block for a free tag in any case, or
add a thread to restart things on a free tag. The thread would not help
performance.

But the -EBUSY is definitely a bug, that needs to be a waiting condition
as well.

-- 
Jens Axboe