[PATCH v3 0/2] permit to set block parameters per vendor

[PATCH v3 0/2] permit to set block parameters per vendor

[Kiwoong Kim]

v3 -> v2: rename exynos functions
v1 -> v2: rename the vops and fix some typos

Kiwoong Kim (2):
  ufs: add a vops to configure block parameter
  ufs: ufs-exynos: set dma_alignment to 4095

 drivers/scsi/ufs/ufs-exynos.c | 9 +++++++++
 drivers/scsi/ufs/ufshcd.c     | 2 ++
 drivers/scsi/ufs/ufshcd.h     | 8 ++++++++
 3 files changed, 19 insertions(+)

-- 
2.7.4

[PATCH v3 1/2] ufs: add a vops to configure block parameter

[Kiwoong Kim]

There could be some cases to set block paramters
per host, because of its own dma structure or whatever.

Signed-off-by: Kiwoong Kim <kwmad.kim@samsung.com>
---
 drivers/scsi/ufs/ufshcd.c | 2 ++
 drivers/scsi/ufs/ufshcd.h | 8 ++++++++
 2 files changed, 10 insertions(+)

diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
index 92d433d..5f89b0e 100644
--- a/drivers/scsi/ufs/ufshcd.c
+++ b/drivers/scsi/ufs/ufshcd.c
@@ -4758,6 +4758,8 @@ static int ufshcd_slave_configure(struct scsi_device *sdev)
 
 	ufshcd_crypto_setup_rq_keyslot_manager(hba, q);
 
+	ufshcd_vops_slave_configure(hba, sdev);
+
 	return 0;
 }
 
diff --git a/drivers/scsi/ufs/ufshcd.h b/drivers/scsi/ufs/ufshcd.h
index 61344c4..4bf4fed 100644
--- a/drivers/scsi/ufs/ufshcd.h
+++ b/drivers/scsi/ufs/ufshcd.h
@@ -329,6 +329,7 @@ struct ufs_hba_variant_ops {
 					void *data);
 	int	(*program_key)(struct ufs_hba *hba,
 			       const union ufs_crypto_cfg_entry *cfg, int slot);
+	void	(*slave_configure)(struct scsi_device *sdev);
 };
 
 /* clock gating state  */
@@ -1228,6 +1229,13 @@ static inline void ufshcd_vops_config_scaling_param(struct ufs_hba *hba,
 		hba->vops->config_scaling_param(hba, profile, data);
 }
 
+static inline void ufshcd_vops_slave_configure(struct ufs_hba *hba,
+						    struct scsi_device *sdev)
+{
+	if (hba->vops && hba->vops->slave_configure)
+		hba->vops->slave_configure(sdev);
+}
+
 extern struct ufs_pm_lvl_states ufs_pm_lvl_states[];
 
 /*
-- 
2.7.4

Re: [PATCH v3 1/2] ufs: add a vops to configure block parameter

[Can Guo]

On 2020-12-22 10:21, Kiwoong Kim wrote:
> There could be some cases to set block paramters
> per host, because of its own dma structure or whatever.
> 
> Signed-off-by: Kiwoong Kim <kwmad.kim@samsung.com>
> ---
>  drivers/scsi/ufs/ufshcd.c | 2 ++
>  drivers/scsi/ufs/ufshcd.h | 8 ++++++++
>  2 files changed, 10 insertions(+)
> 
> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> index 92d433d..5f89b0e 100644
> --- a/drivers/scsi/ufs/ufshcd.c
> +++ b/drivers/scsi/ufs/ufshcd.c
> @@ -4758,6 +4758,8 @@ static int ufshcd_slave_configure(struct
> scsi_device *sdev)
> 
>  	ufshcd_crypto_setup_rq_keyslot_manager(hba, q);
> 
> +	ufshcd_vops_slave_configure(hba, sdev);
> +
>  	return 0;
>  }
> 
> diff --git a/drivers/scsi/ufs/ufshcd.h b/drivers/scsi/ufs/ufshcd.h
> index 61344c4..4bf4fed 100644
> --- a/drivers/scsi/ufs/ufshcd.h
> +++ b/drivers/scsi/ufs/ufshcd.h
> @@ -329,6 +329,7 @@ struct ufs_hba_variant_ops {
>  					void *data);
>  	int	(*program_key)(struct ufs_hba *hba,
>  			       const union ufs_crypto_cfg_entry *cfg, int slot);
> +	void	(*slave_configure)(struct scsi_device *sdev);
>  };
> 
>  /* clock gating state  */
> @@ -1228,6 +1229,13 @@ static inline void
> ufshcd_vops_config_scaling_param(struct ufs_hba *hba,
>  		hba->vops->config_scaling_param(hba, profile, data);
>  }
> 
> +static inline void ufshcd_vops_slave_configure(struct ufs_hba *hba,
> +						    struct scsi_device *sdev)
> +{
> +	if (hba->vops && hba->vops->slave_configure)
> +		hba->vops->slave_configure(sdev);
> +}
> +
>  extern struct ufs_pm_lvl_states ufs_pm_lvl_states[];
> 
>  /*

Reviewed-by: Can Guo <cang@codeaurora.org>

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

[Kiwoong Kim]

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. 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.

    |<------------->|
    +-------+------------         +-------+
    +  AAA  + (corrupted)   ...   +  BBB  +
    +-------+------------         +-------+

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)
-- 
2.7.4

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

[Can Guo]

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.

> 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.

> 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)

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

[Kiwoong Kim]

> 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.

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)

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

[Can Guo]

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)

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

[Can Guo]

On 2020-12-22 11:34, Can Guo wrote:
> 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.
>> 
> 

Hi Kiwoong,

Sorry if I made you confused. I think I know your intention here now.
When bio.c makes the bounce bio, it allocates one new page and
add this page to the bounce bio. In this way, only one bi_vec is
needed and only one sglist entry shall be generated. Am I right?

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)

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

[Kiwoong Kim]

> On 2020-12-22 11:34, Can Guo wrote:
> > 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.
> >>
> >
> 
> Hi Kiwoong,
> 
> Sorry if I made you confused. I think I know your intention here now.
> When bio.c makes the bounce bio, it allocates one new page and
> add this page to the bounce bio. In this way, only one bi_vec is
> needed and only one sglist entry shall be generated. Am I right?
> 
> Thanks,
> 
> Can Guo. 

Yes, that's it.


Thanks.
Kiwoong Kim
> >
> >> 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)

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

[Can Guo]

On 2020-12-22 12:39, Kiwoong Kim wrote:
>> On 2020-12-22 11:34, Can Guo wrote:
>> > 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.
>> >>
>> >
>> 
>> Hi Kiwoong,
>> 
>> Sorry if I made you confused. I think I know your intention here now.
>> When bio.c makes the bounce bio, it allocates one new page and
>> add this page to the bounce bio. In this way, only one bi_vec is
>> needed and only one sglist entry shall be generated. Am I right?
>> 
>> Thanks,
>> 
>> Can Guo.
> 
> Yes, that's it.
> 
> 
> Thanks.
> Kiwoong Kim


Other than the typo in commit msg, the change looks good to me. :)

Regards,

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)