[PATCHSET] implement and use port_task

[PATCHSET] implement and use port_task

[Tejun Heo]

Hello, Jeff and Albert.

This patchset implements port_task and replace pio_task and
packet_task with it.  port_task isn't very different from pio_task or
packet_task - it's straight forward generalization.  The port_task is
a per-port task avaliable to low level drivers which is managed by
libata core layer.

Differences from pio/packet task.

* Not limited to PIO HSM implementation.  Other drivers can use it.

* Synchronization against EH core layer is done within libata core
  layer, not in PIO HSM implementation (ata_qc_timeout()).

Thanks.

--
tejun

Re: [PATCHSET] implement and use port_task

[Albert Lee]

Hi Jeff,
> 
> This patchset implements port_task and replace pio_task and
> packet_task with it.  port_task isn't very different from pio_task or
> packet_task - it's straight forward generalization.  The port_task is
> a per-port task avaliable to low level drivers which is managed by
> libata core layer.
> 
> Differences from pio/packet task.
> 
> * Not limited to PIO HSM implementation.  Other drivers can use it.
> 
> * Synchronization against EH core layer is done within libata core
>   layer, not in PIO HSM implementation (ata_qc_timeout()).
> 
> Thanks.
> 
> --
> tejun
> 

Something related to the PIO HSM.
After the task renaming by this patchset, the task name is changed from pio_task to port_task.
Should we add polling DMA support to the polling HSM?
It could make the polling HSM more complete.

Polling DMA does no performance gain. However, DMA has crc check,
so it won't be worse than PIO when irq is not working.

Albert

Re: [PATCHSET] implement and use port_task

[Tejun Heo]

Albert Lee wrote:
> Hi Jeff,
> 
>>This patchset implements port_task and replace pio_task and
>>packet_task with it.  port_task isn't very different from pio_task or
>>packet_task - it's straight forward generalization.  The port_task is
>>a per-port task avaliable to low level drivers which is managed by
>>libata core layer.
>>
>>Differences from pio/packet task.
>>
>>* Not limited to PIO HSM implementation.  Other drivers can use it.
>>
>>* Synchronization against EH core layer is done within libata core
>>  layer, not in PIO HSM implementation (ata_qc_timeout()).
>>
>>Thanks.
>>
>>--
>>tejun
>>
> 
> 
> Something related to the PIO HSM.
> After the task renaming by this patchset, the task name is changed from pio_task to port_task.
> Should we add polling DMA support to the polling HSM?
> It could make the polling HSM more complete.
> 
> Polling DMA does no performance gain. However, DMA has crc check,
> so it won't be worse than PIO when irq is not working.
> 

Hmmm... Does irq-pio branch still have polling pio support? I hope we 
can implement polling without impacting normal code path in some generic 
way (pseudo periodic interrupts and each interrupt handler returning 
without doing any harm if nothing has happened should work). I don't 
think complicating HSM with DMA-by-polling is a good idea. It's not like 
all low level drivers can share single implementation of HSM.

Thanks.

-- 
tejun

Re: [PATCHSET] implement and use port_task

[Jeff Garzik]

Albert Lee wrote:
> Something related to the PIO HSM.
> After the task renaming by this patchset, the task name is changed from pio_task to port_task.
> Should we add polling DMA support to the polling HSM?
> It could make the polling HSM more complete.
> 
> Polling DMA does no performance gain. However, DMA has crc check,
> so it won't be worse than PIO when irq is not working.


Polling DMA support with nIEN=1 is theoretically invalid.  You are not 
supposed to touch the ATA shadow register set until after receiving an 
irq event.

On chipsets where nIEN=0 and the chip's IntrMask is also zero, its doable.

I'm satisfied with simply never doing polling DMA, and change our mind 
only if a critical need arises.

	Jeff

Re: [PATCHSET] implement and use port_task

[Albert Lee]

Jeff Garzik wrote:
> 
> 
> Polling DMA support with nIEN=1 is theoretically invalid.  You are not
> supposed to touch the ATA shadow register set until after receiving an
> irq event.
> 
> On chipsets where nIEN=0 and the chip's IntrMask is also zero, its doable.
> 
> I'm satisfied with simply never doing polling DMA, and change our mind
> only if a critical need arises.
> 
> 

Understand. After checking the code, polling DMA looks bad idea.
Thanks for the advice.

BTW, Tejun's proposal of pseudo periodic interrupts reminds me
of some patches to integrate the polling pio with irq-pio.
Will submit the patches for your review later.

Albert

[PATCHSET] implement and use port_task

[Tejun Heo]

Hello, Jeff and Albert.

This patchset implements port_task and replace pio_task and
packet_task with it.  port_task isn't very different from pio_task or
packet_task - it's straight forward generalization.  The port_task is
a per-port task avaliable to low level drivers which is managed by
libata core layer.

Differences from pio/packet task.

* Not limited to PIO HSM implementation.  Other drivers can use it.

* Synchronization against EH core layer is done within libata core
  layer, not in PIO HSM implementation (ata_qc_timeout()).

Thanks.

--
tejun