Re: [PATCH v3 2/2] ufs: ufs-exynos: set dma_alignment to 4095

[Can Guo <cang@codeaurora.org>]

On 2020-12-22 11:17, Kiwoong Kim wrote:
>> On 2020-12-22 10:21, Kiwoong Kim wrote:
>> > Exynos requires one scatterlist entry for smaller than page size, i.e.
>> > 4KB. For the cases of dispatching commands with more than one
>> > scatterlist entry and under 4KB size, Exynos behaves as follows:
>> >
>> > Given that a command to read something from device is dispatched with
>> > two scatterlist entries that are named AAA and BBB. After dispatching,
>> > host builds two PRDT entries and during transmission, device sends
>> > just one DATA IN because device doesn't care on host dma.
>> 
>> If my understanding is correct, above is same to all hosts, only below
>> part is Exynos's behavior. Please correct me if I am wrong.
> You're correct.
> 
>> 
>> > The host then tranfers
>> > the whole data from start address of the area named AAA.
>> > In consequebnce, the area that follows AAA would be corrupted.
>> 
>> In consequence
>> 
>> >
>> >     |<------------->|
>> >     +-------+------------         +-------+
>> >     +  AAA  + (corrupted)   ...   +  BBB  +
>> >     +-------+------------         +-------+
>> >
>> 
>> AFAIK, queue->dma_alignment is only used in the case of direct-io, 
>> i.e. in
>> blk_rq_map_user/kern(), which are mainly used in IOCTL.
>> If a request's buffer len and/or buffer start addr is not aligned with
>> queue->dma_alignment, bio.c will make a bounce bio such that the 
>> request
>> get a new buffer which starts on a new page. After the bounce bio is
> ended,
>> the data in the bound bio will be copied to the initial buffer.
>> 
>> So in this fix, you are making sure the AAA and BBB are all mapped to 
>> one
>> bounce bio and stay in one bi_vec, so when we do map_sg they come in 
>> one
>> sglist, please correct me if I am wrong.
>> 
>> If my understanding is correct, what is the real use case here - 
>> why/how
>> user starts a request which can generate more than one sglists whose 
>> sizes
>> are all under 4KB? I am just curious.
>> 
>> Thanks,
>> 
>> Can Guo.
> 
> You nearly exactly got what I’m thinking.
> And I think there could be various cases making those situations,
> which are definitely up to user programs. That is the case using
> different memory areas to contain something.
> 

If you want to make AAA and BBB stay in one bi_vec, they should be
continuous in memory (DDR), otherwise they will be put into two
bi_vecs and eventually become two sglist entries. My doubt is that
if user uses different memory areas to contain something as you
said, how can this fix make AAA and BBB stay in one sglist entry?

Thanks,

Can Guo.

> Thanks.
> Kiwoong Kim
>> 
>> > Signed-off-by: Kiwoong Kim <kwmad.kim@samsung.com>
>> > ---
>> >  drivers/scsi/ufs/ufs-exynos.c | 9 +++++++++
>> >  1 file changed, 9 insertions(+)
>> >
>> > diff --git a/drivers/scsi/ufs/ufs-exynos.c
>> > b/drivers/scsi/ufs/ufs-exynos.c index a8770ff..8635d9d 100644
>> > --- a/drivers/scsi/ufs/ufs-exynos.c
>> > +++ b/drivers/scsi/ufs/ufs-exynos.c
>> > @@ -14,6 +14,7 @@
>> >  #include <linux/of_address.h>
>> >  #include <linux/phy/phy.h>
>> >  #include <linux/platform_device.h>
>> > +#include <linux/blkdev.h>
>> >
>> >  #include "ufshcd.h"
>> >  #include "ufshcd-pltfrm.h"
>> > @@ -1193,6 +1194,13 @@ static int exynos_ufs_resume(struct ufs_hba
>> > *hba, enum ufs_pm_op pm_op)
>> >  	return 0;
>> >  }
>> >
>> > +static void exynos_ufs_slave_configure(struct scsi_device *sdev) {
>> > +	struct request_queue *q = sdev->request_queue;
>> > +
>> > +	blk_queue_update_dma_alignment(q, PAGE_SIZE - 1); }
>> > +
>> >  static struct ufs_hba_variant_ops ufs_hba_exynos_ops = {
>> >  	.name				= "exynos_ufs",
>> >  	.init				= exynos_ufs_init,
>> > @@ -1204,6 +1212,7 @@ static struct ufs_hba_variant_ops
>> > ufs_hba_exynos_ops = {
>> >  	.hibern8_notify			= exynos_ufs_hibern8_notify,
>> >  	.suspend			= exynos_ufs_suspend,
>> >  	.resume				= exynos_ufs_resume,
>> > +	.slave_configure		= exynos_ufs_slave_configure,
>> >  };
>> >
>> >  static int exynos_ufs_probe(struct platform_device *pdev)