Re: [PATCH 1/2] mtip32xx: Don't bounce IO requests

[Jens Axboe <axboe@kernel.dk>]

On 03/12/2014 11:18 AM, David Vrabel wrote:
> On 12/03/14 17:08, Felipe Franciosi wrote:
>> On Wed, Mar 12, 2014 at 4:16 PM, Jens Axboe <axboe@kernel.dk> wrote:
>>>
>>> On 03/12/2014 10:05 AM, Felipe Franciosi wrote:
>>>>
>>>> This device support 64-bit DMA and therefore no IO requests needs bouncing.
>>>> On 32-bit systems, blk_queue_bounce() will bounce all IO requests with high
>>>> mem pages. This makes performance really really bad.
> [...]
>>>> --- a/drivers/block/mtip32xx/mtip32xx.c
>>>> +++ b/drivers/block/mtip32xx/mtip32xx.c
>>>> @@ -4082,8 +4082,6 @@ static void mtip_make_request(struct request_queue *queue, struct bio *bio)
>>>>
>>>>          sg = mtip_hw_get_scatterlist(dd, &tag, unaligned);
>>>>          if (likely(sg != NULL)) {
>>>> -               blk_queue_bounce(queue, &bio);
>>>> -
>>>>                  if (unlikely((bio)->bi_vcnt > MTIP_MAX_SG)) {
>>>>                          dev_warn(&dd->pdev->dev,
>>>>                                  "Maximum number of SGL entries exceeded\n");
>>>
>>>
>>> That seems to be because the driver forgets to set the DMA
>>> capabilities. It needs a blk_queue_bounce_limit() call when it sets up
>>> the queue.
>>
>> Hmm... I see what you are saying. David, would you like to comment?
>
> Yes, it should also call blk_queue_bounce_limit(..., BLK_BOUNCE_ANY) in
> mtip_block_initialize() as well.
>
> But, only one other driver calls blk_queue_bounce() so it's difficult to
> see why this driver would need a call here.

That's because most other drivers hook in the appropriate place, and the 
blk_queue_bounce() is then done by the block layer.

-- 
Jens Axboe