Re: [PATCH v5 0/7] add virt-dma support for imx-sdma

[Lucas Stach <l.stach@pengutronix.de>]

Hi Robin,

I've tested this whole series with the SDMA being used for SPI, UART
and SSI with no regressions spotted. As this should cover most common
use-cases, I think this series is good to go in.

Tested-by: Lucas Stach <l.stach@pengutronix.de>

Regards,
Lucas

Am Mittwoch, den 20.06.2018, 00:56 +0800 schrieb Robin Gong:
> The legacy sdma driver has below limitations or drawbacks:
>   1. Hardcode the max BDs number as "PAGE_SIZE / sizeof(*)", and alloc
>      one page size for one channel regardless of only few BDs needed
>      most time. But in few cases, the max PAGE_SIZE maybe not enough.
>   2. One SDMA channel can't stop immediatley once channel disabled which
>      means SDMA interrupt may come in after this channel terminated.There
>      are some patches for this corner case such as commit "2746e2c389f9",
>      but not cover non-cyclic.
> 
> The common virt-dma overcomes the above limitations. It can alloc bd
> dynamically and free bd once this tx transfer done. No memory wasted or
> maximum limititation here, only depends on how many memory can be requested
> from kernel. For No.2, such issue can be workaround by checking if there
> is available descript("sdmac->desc") now once the unwanted interrupt
> coming. At last the common virt-dma is easier for sdma driver maintain.
> 
> Change from v4:
>   1. identify lockdep issue which caused by allocate memory with
>      'GFP_KERNEL', change to 'GFP_NOWAIT' instead so that lockdep
>      ignore check. That also make sense since Audio/uart driver may
>      call dma function after spin_lock_irqsave()...
>   2. use dma pool instead for bd description allocated,since audio
>      driver may call dma_terminate_all in irq. Please refer to 7/7.
>   3. remove 7/7 serial patch in v4, since lockdep issued fixed by No.1 
> 
> Change from v3:
>   1. add two uart patches which impacted by this patchset.
>   2. unlock 'vc.lock' before cyclic dma callback and lock again after
>      it because some driver such as uart will call dmaengine_tx_status
>      which will acquire 'vc.lock' again and dead lock comes out.
>   3. remove 'Revert commit' stuff since that patch is not wrong and
>      combine two patch into one patch as Sascha's comment.
> 
> Change from v2:
>   1. include Sascha's patch to make the main patch easier to review.
>      Thanks Sacha.
>   2. remove useless 'desc'/'chan' in struct sdma_channe.
> 
> Change from v1:
>   1. split v1 patch into 5 patches.
>   2. remove some unnecessary condition check.
>   3. remove unnecessary 'pending' list.
> 
> Robin Gong (6):
>   tty: serial: imx: correct dma cookie status
>   dmaengine: imx-sdma: add virt-dma support
>   dmaengine: imx-sdma: remove useless 'lock' and 'enabled' in 'struct
>     sdma_channel'
>   dmaengine: imx-sdma: remove the maximum limitation for bd numbers
>   dmaengine: imx-sdma: add sdma_transfer_init to decrease code overlap
>   dmaengine: imx-sdma: alloclate bd memory from dma pool
> 
> Sascha Hauer (1):
>   dmaengine: imx-sdma: factor out a struct sdma_desc from struct
>     sdma_channel
> 
>  drivers/dma/Kconfig      |   1 +
>  drivers/dma/imx-sdma.c   | 400 +++++++++++++++++++++++++++--------------------
>  drivers/tty/serial/imx.c |   2 +-
>  3 files changed, 235 insertions(+), 168 deletions(-)
>