[PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[FUJITA Tomonori]

This patch moves blk_queue_max_segment_size to scsi_alloc_queue from
llds. It enables scsi_add_host to tells iommu lld's
dma_max_segment_size. If a low-level driver doesn't specify
dma_max_segment_size, scsi-ml uses 65536 (MAX_SEGMENT_SIZE). So there
are not any functional changes.

Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
---
 drivers/scsi/hosts.c     |    5 +++++
 drivers/scsi/scsi_lib.c  |    1 +
 include/scsi/scsi_host.h |    6 ++++++
 3 files changed, 12 insertions(+), 0 deletions(-)

diff --git a/drivers/scsi/hosts.c b/drivers/scsi/hosts.c
index 96bc312..22877d3 100644
--- a/drivers/scsi/hosts.c
+++ b/drivers/scsi/hosts.c
@@ -365,6 +365,11 @@ struct Scsi_Host *scsi_host_alloc(struct scsi_host_template *sht, int privsize)
 	else
 		shost->dma_boundary = 0xffffffff;
 
+	if (sht->dma_max_segment_size)
+		shost->dma_max_segment_size = sht->dma_max_segment_size;
+	else
+		shost->dma_max_segment_size = 65536;
+
 	rval = scsi_setup_command_freelist(shost);
 	if (rval)
 		goto fail_kfree;
diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index a417a6f..2ec77a9 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -1564,6 +1564,7 @@ struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
 	blk_queue_max_sectors(q, shost->max_sectors);
 	blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
 	blk_queue_segment_boundary(q, shost->dma_boundary);
+	blk_queue_max_segment_size(q, shost->dma_max_segment_size);
 
 	if (!shost->use_clustering)
 		clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
diff --git a/include/scsi/scsi_host.h b/include/scsi/scsi_host.h
index 3b8a6a8..1eb8435 100644
--- a/include/scsi/scsi_host.h
+++ b/include/scsi/scsi_host.h
@@ -383,6 +383,11 @@ struct scsi_host_template {
 	unsigned long dma_boundary;
 
 	/*
+	 * dma scatter gather segment size limit.
+	 */
+	unsigned int dma_max_segment_size;
+
+	/*
 	 * This specifies "machine infinity" for host templates which don't
 	 * limit the transfer size.  Note this limit represents an absolute
 	 * maximum, and may be over the transfer limits allowed for
@@ -571,6 +576,7 @@ struct Scsi_Host {
 	short unsigned int sg_tablesize;
 	short unsigned int max_sectors;
 	unsigned long dma_boundary;
+	unsigned int dma_max_segment_size;
 	/* 
 	 * Used to assign serial numbers to the cmds.
 	 * Protected by the host lock.
-- 
1.5.2.4

Re: [PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[Jens Axboe]

On Wed, Sep 26 2007, FUJITA Tomonori wrote:
> This patch moves blk_queue_max_segment_size to scsi_alloc_queue from
> llds. It enables scsi_add_host to tells iommu lld's
> dma_max_segment_size. If a low-level driver doesn't specify
> dma_max_segment_size, scsi-ml uses 65536 (MAX_SEGMENT_SIZE). So there
> are not any functional changes.
> 
> Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> ---
>  drivers/scsi/hosts.c     |    5 +++++
>  drivers/scsi/scsi_lib.c  |    1 +
>  include/scsi/scsi_host.h |    6 ++++++
>  3 files changed, 12 insertions(+), 0 deletions(-)
> 
> diff --git a/drivers/scsi/hosts.c b/drivers/scsi/hosts.c
> index 96bc312..22877d3 100644
> --- a/drivers/scsi/hosts.c
> +++ b/drivers/scsi/hosts.c
> @@ -365,6 +365,11 @@ struct Scsi_Host *scsi_host_alloc(struct scsi_host_template *sht, int privsize)
>  	else
>  		shost->dma_boundary = 0xffffffff;
>  
> +	if (sht->dma_max_segment_size)
> +		shost->dma_max_segment_size = sht->dma_max_segment_size;
> +	else
> +		shost->dma_max_segment_size = 65536;

Use MAX_SEGMENT_SIZE here, instead of manually entering 64k?

Otherwise everything looks good from a quick look, it's definitely
missing functionality that we have wanted for quite some time!

-- 
Jens Axboe

Re: [PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[Jeff Garzik]

FUJITA Tomonori wrote:
> This patch moves blk_queue_max_segment_size to scsi_alloc_queue from
> llds. It enables scsi_add_host to tells iommu lld's
> dma_max_segment_size. If a low-level driver doesn't specify
> dma_max_segment_size, scsi-ml uses 65536 (MAX_SEGMENT_SIZE). So there
> are not any functional changes.
> 
> Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> ---
>  drivers/scsi/hosts.c     |    5 +++++
>  drivers/scsi/scsi_lib.c  |    1 +
>  include/scsi/scsi_host.h |    6 ++++++
>  3 files changed, 12 insertions(+), 0 deletions(-)

hmmmmm...  All the patches look technically correct, but IMO this really 
should behave more the the dma_mask interface:  platform sets a sane 
dma_mask (usually 0xffffffff), and LLDD calls dma_set_mask() or 
pci_set_dma_mask().

Thus, IMO an LLDD should call dma_set_max_seg(), and then SCSI midlayer 
can obtain that value from struct device.

Just like dma_mask, I think we can avoid a scsi_host_template addition.

I echo Jens sentiment, though, in closing:  thanks for doing this work, 
it's been needed for quite a while.

	Jeff

Re: [PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[FUJITA Tomonori]

On Wed, 26 Sep 2007 06:11:45 -0400
Jeff Garzik <jeff@garzik.org> wrote:

> FUJITA Tomonori wrote:
> > This patch moves blk_queue_max_segment_size to scsi_alloc_queue from
> > llds. It enables scsi_add_host to tells iommu lld's
> > dma_max_segment_size. If a low-level driver doesn't specify
> > dma_max_segment_size, scsi-ml uses 65536 (MAX_SEGMENT_SIZE). So there
> > are not any functional changes.
> > 
> > Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> > ---
> >  drivers/scsi/hosts.c     |    5 +++++
> >  drivers/scsi/scsi_lib.c  |    1 +
> >  include/scsi/scsi_host.h |    6 ++++++
> >  3 files changed, 12 insertions(+), 0 deletions(-)
> 
> hmmmmm...  All the patches look technically correct, but IMO this really 
> should behave more the the dma_mask interface:  platform sets a sane 
> dma_mask (usually 0xffffffff), and LLDD calls dma_set_mask() or 
> pci_set_dma_mask().
>
> Thus, IMO an LLDD should call dma_set_max_seg(), and then SCSI midlayer 
> can obtain that value from struct device.

Yeah, I agreed that max_segment_size should work like dma_mask (that's
why I simply put max_segment_size to device structure).

scsi_debug doesn't use dma but calls blk_queue_max_segment_size (I
guess that it wants large I/Os). If we can remove it (thanks to
chaining sg), scsi-ml gets that value that llds set via
dma_set_max_seg and calls blk_queue_max_segment_size.

Re: [PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[Jeff Garzik]

FUJITA Tomonori wrote:
> On Wed, 26 Sep 2007 06:11:45 -0400
> Jeff Garzik <jeff@garzik.org> wrote:
> 
>> FUJITA Tomonori wrote:
>>> This patch moves blk_queue_max_segment_size to scsi_alloc_queue from
>>> llds. It enables scsi_add_host to tells iommu lld's
>>> dma_max_segment_size. If a low-level driver doesn't specify
>>> dma_max_segment_size, scsi-ml uses 65536 (MAX_SEGMENT_SIZE). So there
>>> are not any functional changes.
>>>
>>> Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
>>> ---
>>>  drivers/scsi/hosts.c     |    5 +++++
>>>  drivers/scsi/scsi_lib.c  |    1 +
>>>  include/scsi/scsi_host.h |    6 ++++++
>>>  3 files changed, 12 insertions(+), 0 deletions(-)
>> hmmmmm...  All the patches look technically correct, but IMO this really 
>> should behave more the the dma_mask interface:  platform sets a sane 
>> dma_mask (usually 0xffffffff), and LLDD calls dma_set_mask() or 
>> pci_set_dma_mask().
>>
>> Thus, IMO an LLDD should call dma_set_max_seg(), and then SCSI midlayer 
>> can obtain that value from struct device.
> 
> Yeah, I agreed that max_segment_size should work like dma_mask (that's
> why I simply put max_segment_size to device structure).

Yep!


> scsi_debug doesn't use dma but calls blk_queue_max_segment_size (I
> guess that it wants large I/Os). If we can remove it (thanks to
> chaining sg), scsi-ml gets that value that llds set via
> dma_set_max_seg and calls blk_queue_max_segment_size.

[/me checks the code]  Actually scsi_debug has its own pseudo-bus and 
struct device, so it sounds like scsi_debug can call dma_set_max_seg() 
just like any other LLDD?

Maybe dev_set_max_seg() is a better name, if people get really picky (I 
don't care).

	Jeff

Re: [PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[FUJITA Tomonori]

On Wed, 26 Sep 2007 10:42:06 -0400
Jeff Garzik <jeff@garzik.org> wrote:

> FUJITA Tomonori wrote:
> > On Wed, 26 Sep 2007 06:11:45 -0400
> > Jeff Garzik <jeff@garzik.org> wrote:
> > 
> >> FUJITA Tomonori wrote:
> >>> This patch moves blk_queue_max_segment_size to scsi_alloc_queue from
> >>> llds. It enables scsi_add_host to tells iommu lld's
> >>> dma_max_segment_size. If a low-level driver doesn't specify
> >>> dma_max_segment_size, scsi-ml uses 65536 (MAX_SEGMENT_SIZE). So there
> >>> are not any functional changes.
> >>>
> >>> Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> >>> ---
> >>>  drivers/scsi/hosts.c     |    5 +++++
> >>>  drivers/scsi/scsi_lib.c  |    1 +
> >>>  include/scsi/scsi_host.h |    6 ++++++
> >>>  3 files changed, 12 insertions(+), 0 deletions(-)
> >> hmmmmm...  All the patches look technically correct, but IMO this really 
> >> should behave more the the dma_mask interface:  platform sets a sane 
> >> dma_mask (usually 0xffffffff), and LLDD calls dma_set_mask() or 
> >> pci_set_dma_mask().
> >>
> >> Thus, IMO an LLDD should call dma_set_max_seg(), and then SCSI midlayer 
> >> can obtain that value from struct device.
> > 
> > Yeah, I agreed that max_segment_size should work like dma_mask (that's
> > why I simply put max_segment_size to device structure).
> 
> Yep!
> 
> 
> > scsi_debug doesn't use dma but calls blk_queue_max_segment_size (I
> > guess that it wants large I/Os). If we can remove it (thanks to
> > chaining sg), scsi-ml gets that value that llds set via
> > dma_set_max_seg and calls blk_queue_max_segment_size.
> 
> [/me checks the code]  Actually scsi_debug has its own pseudo-bus and 
> struct device, so it sounds like scsi_debug can call dma_set_max_seg() 
> just like any other LLDD?

You are right. scsi_debug's pseudo-bus works.

But probabaly, scsi_debug doesn't need to call
blk_queue_max_segment_size now.


> Maybe dev_set_max_seg() is a better name, if people get really picky (I 
> don't care).

How about dma_set_max_seg_size()?


diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index a417a6f..7adadfb 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -1554,6 +1554,7 @@ struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
 					 request_fn_proc *request_fn)
 {
 	struct request_queue *q;
+	struct device *dev = shost->shost_gendev.parent;
 
 	q = blk_init_queue(request_fn, NULL);
 	if (!q)
@@ -1565,6 +1566,9 @@ struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
 	blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
 	blk_queue_segment_boundary(q, shost->dma_boundary);
 
+	if (dev->max_segment_size)
+		blk_queue_max_segment_size(q, dev->max_segment_size);
+
 	if (!shost->use_clustering)
 		clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
 	return q;
diff --git a/include/linux/device.h b/include/linux/device.h
index 3a38d1f..8046b60 100644
--- a/include/linux/device.h
+++ b/include/linux/device.h
@@ -443,6 +443,13 @@ struct device {
 
 	struct dma_coherent_mem	*dma_mem; /* internal for coherent mem
 					     override */
+
+	/*
+	 * a low level driver may set these to teach IOMMU code about
+	 * sg limitations.
+	 */
+	unsigned int max_segment_size;
+
 	/* arch specific additions */
 	struct dev_archdata	archdata;
 
diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
index 2dc21cb..30404b8 100644
--- a/include/linux/dma-mapping.h
+++ b/include/linux/dma-mapping.h
@@ -31,6 +31,11 @@ static inline int valid_dma_direction(int dma_direction)
 		(dma_direction == DMA_FROM_DEVICE));
 }
 
+static inline void dma_set_max_seg_size(struct device *dev, unsigned int size)
+{
+	dev->max_segment_size = size;
+}
+
 #ifdef CONFIG_HAS_DMA
 #include <asm/dma-mapping.h>
 #else

Re: [PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[Jeff Garzik]

FUJITA Tomonori wrote:
> You are right. scsi_debug's pseudo-bus works.
> 
> But probabaly, scsi_debug doesn't need to call
> blk_queue_max_segment_size now.

Either way works for me.  :)


>> Maybe dev_set_max_seg() is a better name, if people get really picky (I 
>> don't care).
> 
> How about dma_set_max_seg_size()?
> 
> 
> diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
> index a417a6f..7adadfb 100644
> --- a/drivers/scsi/scsi_lib.c
> +++ b/drivers/scsi/scsi_lib.c
> @@ -1554,6 +1554,7 @@ struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
>  					 request_fn_proc *request_fn)
>  {
>  	struct request_queue *q;
> +	struct device *dev = shost->shost_gendev.parent;
>  
>  	q = blk_init_queue(request_fn, NULL);
>  	if (!q)
> @@ -1565,6 +1566,9 @@ struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
>  	blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
>  	blk_queue_segment_boundary(q, shost->dma_boundary);
>  
> +	if (dev->max_segment_size)
> +		blk_queue_max_segment_size(q, dev->max_segment_size);
> +
>  	if (!shost->use_clustering)
>  		clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
>  	return q;
> diff --git a/include/linux/device.h b/include/linux/device.h
> index 3a38d1f..8046b60 100644
> --- a/include/linux/device.h
> +++ b/include/linux/device.h
> @@ -443,6 +443,13 @@ struct device {
>  
>  	struct dma_coherent_mem	*dma_mem; /* internal for coherent mem
>  					     override */
> +
> +	/*
> +	 * a low level driver may set these to teach IOMMU code about
> +	 * sg limitations.
> +	 */
> +	unsigned int max_segment_size;
> +
>  	/* arch specific additions */
>  	struct dev_archdata	archdata;
>  
> diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
> index 2dc21cb..30404b8 100644
> --- a/include/linux/dma-mapping.h
> +++ b/include/linux/dma-mapping.h
> @@ -31,6 +31,11 @@ static inline int valid_dma_direction(int dma_direction)
>  		(dma_direction == DMA_FROM_DEVICE));
>  }
>  
> +static inline void dma_set_max_seg_size(struct device *dev, unsigned int size)
> +{
> +	dev->max_segment_size = size;
> +}
> +

Definitely moving in the right direction, IMO.  I would suggest a few 
minor changes to dma_set_max_seg_size(), to permit platforms to override 
the default behavior, and to get people into the habit of checking the 
return value (since it might matter in the future):

1) surround with #ifndef ARCH_HAS_DMA_MAX_SEG

2) change return value to 'int'

3) unconditionally return zero, for now.

Otherwise, approach and code look good to me.

	Jeff

Re: [PATCH 1/5] add dma_max_segment_size option to scsi_host_template

[FUJITA Tomonori]

On Thu, 27 Sep 2007 02:31:14 -0400
Jeff Garzik <jeff@garzik.org> wrote:

> FUJITA Tomonori wrote:
> > You are right. scsi_debug's pseudo-bus works.
> > 
> > But probabaly, scsi_debug doesn't need to call
> > blk_queue_max_segment_size now.
> 
> Either way works for me.  :)
> 
> 
> >> Maybe dev_set_max_seg() is a better name, if people get really picky (I 
> >> don't care).
> > 
> > How about dma_set_max_seg_size()?
> > 
> > 
> > diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
> > index a417a6f..7adadfb 100644
> > --- a/drivers/scsi/scsi_lib.c
> > +++ b/drivers/scsi/scsi_lib.c
> > @@ -1554,6 +1554,7 @@ struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
> >  					 request_fn_proc *request_fn)
> >  {
> >  	struct request_queue *q;
> > +	struct device *dev = shost->shost_gendev.parent;
> >  
> >  	q = blk_init_queue(request_fn, NULL);
> >  	if (!q)
> > @@ -1565,6 +1566,9 @@ struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
> >  	blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
> >  	blk_queue_segment_boundary(q, shost->dma_boundary);
> >  
> > +	if (dev->max_segment_size)
> > +		blk_queue_max_segment_size(q, dev->max_segment_size);
> > +
> >  	if (!shost->use_clustering)
> >  		clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
> >  	return q;
> > diff --git a/include/linux/device.h b/include/linux/device.h
> > index 3a38d1f..8046b60 100644
> > --- a/include/linux/device.h
> > +++ b/include/linux/device.h
> > @@ -443,6 +443,13 @@ struct device {
> >  
> >  	struct dma_coherent_mem	*dma_mem; /* internal for coherent mem
> >  					     override */
> > +
> > +	/*
> > +	 * a low level driver may set these to teach IOMMU code about
> > +	 * sg limitations.
> > +	 */
> > +	unsigned int max_segment_size;
> > +
> >  	/* arch specific additions */
> >  	struct dev_archdata	archdata;
> >  
> > diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
> > index 2dc21cb..30404b8 100644
> > --- a/include/linux/dma-mapping.h
> > +++ b/include/linux/dma-mapping.h
> > @@ -31,6 +31,11 @@ static inline int valid_dma_direction(int dma_direction)
> >  		(dma_direction == DMA_FROM_DEVICE));
> >  }
> >  
> > +static inline void dma_set_max_seg_size(struct device *dev, unsigned int size)
> > +{
> > +	dev->max_segment_size = size;
> > +}
> > +
> 
> Definitely moving in the right direction, IMO.  I would suggest a few 
> minor changes to dma_set_max_seg_size(), to permit platforms to override 
> the default behavior, and to get people into the habit of checking the 
> return value (since it might matter in the future):
> 
> 1) surround with #ifndef ARCH_HAS_DMA_MAX_SEG
> 
> 2) change return value to 'int'
> 
> 3) unconditionally return zero, for now.
> 
> Otherwise, approach and code look good to me.

I will, thanks!

I'll post an updated patchset for -mm kernel soon.