[git-patches] EH, irq-pio merge and NCQ patches

[git-patches] EH, irq-pio merge and NCQ patches

[Tejun Heo]

Hello, Jeff.

This is the updated git-patches of previously posted five
patchsets[1].  Changes are

* SCSI-implement-host_eh_scheduled patch updated as discussed.

* A new patch, which renames ata_down_sata_spd_limit() and friends to
  sata_*(), is added to prep patchset.

* ata_scr_*() renamed to sata_scr_*()

* ehc->flags has been moved to ehi->flags so that it can also be used
  to pass flags from IRQ handler to EH.

* Spurious ata_dev_init() prototype killed from ering patch.

* ata_ering is changed to include #ATA_ERING_SIZE predefined ering
  entries.  As this makes add-dev-ering patch trivial, the patch is
  collapsed into implement-ering.

* sata_sil EH now directly calls ata_bmdma_irq_clear() instead of
  calling ops->irq_clear().

I didn't add DMA engine stopping to ata_bmdma_freeze() in this series.
When BMDMA engine is turned off, all status bits go too.  However, the
status is used by BMDMA EH to report and to recognize host bus error
reprsented as timeout.

To make all of them work, the BMDMA status needs to be stored in
ata_bmdma_freeze() and used in EH, which is currently a bit difficult
to implement for generic BMDMA.  I think we can resolve this later.

The series can be browsed at the following URL.

	http://htj.dyndns.org/git/?p=libata-tj.git;a=shortlog;h=for-jeff

If you're satisfied with the updates, please pull from.

	git://htj.dyndns.org/libata-tj for-jeff

The tree is tagged with the following.

	fj-prep
	fj-eh-fw
	fj-eh
	fj-merge-irq-pio
	fj-ncq

Thanks.

[1] http://article.gmane.org/gmane.linux.ide/9959
[2] http://article.gmane.org/gmane.linux.ide/9984
[3] http://article.gmane.org/gmane.linux.ide/9995
[4] http://article.gmane.org/gmane.linux.ide/10005
[5] http://article.gmane.org/gmane.linux.ide/10011

-- 
tejun

Re: [git-patches] EH, irq-pio merge and NCQ patches

[Tejun Heo]

Tejun Heo wrote:
> Hello, Jeff.
> 
> This is the updated git-patches of previously posted five
> patchsets[1].  Changes are
> 
> * SCSI-implement-host_eh_scheduled patch updated as discussed.
> 
> * A new patch, which renames ata_down_sata_spd_limit() and friends to
>   sata_*(), is added to prep patchset.
> 
> * ata_scr_*() renamed to sata_scr_*()
> 
> * ehc->flags has been moved to ehi->flags so that it can also be used
>   to pass flags from IRQ handler to EH.
> 
> * Spurious ata_dev_init() prototype killed from ering patch.
> 
> * ata_ering is changed to include #ATA_ERING_SIZE predefined ering
>   entries.  As this makes add-dev-ering patch trivial, the patch is
>   collapsed into implement-ering.
> 
> * sata_sil EH now directly calls ata_bmdma_irq_clear() instead of
>   calling ops->irq_clear().

I forgot one.

* AHCI now has proper IRQ-mask manipulating freeze/thaw().

-- 
tejun

Re: [git-patches] EH, irq-pio merge and NCQ patches

[Jeff Garzik]

Tejun Heo wrote:
> I didn't add DMA engine stopping to ata_bmdma_freeze() in this series.
> When BMDMA engine is turned off, all status bits go too.  However, the
> status is used by BMDMA EH to report and to recognize host bus error
> reprsented as timeout.
> 
> To make all of them work, the BMDMA status needs to be stored in
> ata_bmdma_freeze() and used in EH, which is currently a bit difficult
> to implement for generic BMDMA.  I think we can resolve this later.

Agree re storing BMDMA status.  The very next operation needs to be 
DMA-stop, though.  Otherwise shadow register block isn't accessible. 
And you must still check BMDMA status -after- stopping DMA, too (that 
doesn't invalidate your logic quoted above, of course).


> The series can be browsed at the following URL.
> 
> 	http://htj.dyndns.org/git/?p=libata-tj.git;a=shortlog;h=for-jeff
> 
> If you're satisfied with the updates, please pull from.
> 
> 	git://htj.dyndns.org/libata-tj for-jeff

Pulled into #upstream and other branches, and pushed to kernel.org. 
Please check and make sure everything looks OK, after kernel.org 
finishes mirroring to the outside world.

Thanks.

BTW, it would be nice if you would start with branch #sii-irq when you 
are updating the sata_sil interrupt handling path.  And eventually we 
want to push #sii-lbt too.

	Jeff

Re: [git-patches] EH, irq-pio merge and NCQ patches

[Tejun Heo]

Jeff Garzik wrote:
> Tejun Heo wrote:
>> I didn't add DMA engine stopping to ata_bmdma_freeze() in this series.
>> When BMDMA engine is turned off, all status bits go too.  However, the
>> status is used by BMDMA EH to report and to recognize host bus error
>> reprsented as timeout.
>>
>> To make all of them work, the BMDMA status needs to be stored in
>> ata_bmdma_freeze() and used in EH, which is currently a bit difficult
>> to implement for generic BMDMA.  I think we can resolve this later.
> 
> Agree re storing BMDMA status.  The very next operation needs to be 
> DMA-stop, though.  Otherwise shadow register block isn't accessible. And 
> you must still check BMDMA status -after- stopping DMA, too (that 
> doesn't invalidate your logic quoted above, of course).

I see.  I think this can be done by making all BMDMA related stuff into 
a layer between highlevel libata and BMDMA LLDDs.  We can move all IO 
ops into that layer and newer controllers like ahci, sata_sil24 and 
sata_mv don't have to care about them.  Have yet to think about how to 
do it in simple way though, such that changes to BMDMA LLDDs are minimal.

> Pulled into #upstream and other branches, and pushed to kernel.org. 
> Please check and make sure everything looks OK, after kernel.org 
> finishes mirroring to the outside world.

I glanced over them and diffed with my own tree.  All look good & dandy.

> BTW, it would be nice if you would start with branch #sii-irq when you 
> are updating the sata_sil interrupt handling path.

I certainly can, but I'm not sure it's a good idea to do it.  Most 
likely, it will end up in another rewriting of the rewritten irq handler 
(commit-wise, it will be merge-sii-irq then rewrite sii-irq).  The new 
one in hp-LLDD-prep uses BMDMA2 to cut down on IO access and also makes 
use of the fact that sata_sil uses ATA_DMA_INTR bit to indicate IRQ 
pending even when non-DMA command is in progress, so the whole structure 
is somewhat different from the one in sii-irq.

 > And eventually we want to push #sii-lbt too.

Yeap, I looked at that one, too.  But, IIRC, the test result wasn't too 
positive.  Am I remembering correctly?

-- 
tejun

Re: [git-patches] EH, irq-pio merge and NCQ patches

[Jeff Garzik]

Tejun Heo wrote:
> Jeff Garzik wrote:
>> Tejun Heo wrote:
>>> I didn't add DMA engine stopping to ata_bmdma_freeze() in this series.
>>> When BMDMA engine is turned off, all status bits go too.  However, the
>>> status is used by BMDMA EH to report and to recognize host bus error
>>> reprsented as timeout.
>>>
>>> To make all of them work, the BMDMA status needs to be stored in
>>> ata_bmdma_freeze() and used in EH, which is currently a bit difficult
>>> to implement for generic BMDMA.  I think we can resolve this later.
>>
>> Agree re storing BMDMA status.  The very next operation needs to be 
>> DMA-stop, though.  Otherwise shadow register block isn't accessible. 
>> And you must still check BMDMA status -after- stopping DMA, too (that 
>> doesn't invalidate your logic quoted above, of course).
> 
> I see.  I think this can be done by making all BMDMA related stuff into 
> a layer between highlevel libata and BMDMA LLDDs.  We can move all IO 
> ops into that layer and newer controllers like ahci, sata_sil24 and 
> sata_mv don't have to care about them.  Have yet to think about how to 
> do it in simple way though, such that changes to BMDMA LLDDs are minimal.

In an ideal world, drivers would register themselves as BMDMA drivers 
rather than libata drivers.  Move all the BMDMA hooks to a separate 
structure, leaving only the high level ->qc_issue() interface in 
ata_port_operations.


>> BTW, it would be nice if you would start with branch #sii-irq when you 
>> are updating the sata_sil interrupt handling path.
> 
> I certainly can, but I'm not sure it's a good idea to do it.  Most 
> likely, it will end up in another rewriting of the rewritten irq handler 
> (commit-wise, it will be merge-sii-irq then rewrite sii-irq).  The new 
> one in hp-LLDD-prep uses BMDMA2 to cut down on IO access and also makes 
> use of the fact that sata_sil uses ATA_DMA_INTR bit to indicate IRQ 
> pending even when non-DMA command is in progress, so the whole structure 
> is somewhat different from the one in sii-irq.

OK.  I'll drop my branch then.


>  > And eventually we want to push #sii-lbt too.
> 
> Yeap, I looked at that one, too.  But, IIRC, the test result wasn't too 
> positive.  Am I remembering correctly?

My test results were quite positive :)  At least one negative report on 
3114 IIRC.

	Jeff