[PATCH] staging: comedi: s626: Remove pci-dma-compat wrapper APIs.

[Christophe JAILLET]

Le 11/07/2020 à 14:35, Suraj Upadhyay a écrit :
> The legacy API wrappers in include/linux/pci-dma-compat.h
> should go away as it creates unnecessary midlayering
> for include/linux/dma-mapping.h APIs, instead use dma-mapping.h
> APIs directly.
>
> The patch has been generated with the coccinelle script below
> and compile-tested.
>
> [...]
> @@ expression E1, E2, E3; @@
> - pci_alloc_consistent(E1, E2, E3)
> + dma_alloc_coherent(&E1->dev, E2, E3, GFP_ATOMIC)
>
> @@ expression E1, E2, E3; @@
> - pci_zalloc_consistent(E1, E2, E3)
> + dma_alloc_coherent(&E1->dev, E2, E3, GFP_ATOMIC)

This is the tricky part of this kind of cleanup, see below.

GFP_ATOMIC can't be wrong because it is was exactly what was done with 
the pci_ function.
However, most of the time, it can safely be replaced by GFP_KERNEL which 
gives more opportunities to the memory allocator.

> [...]
> diff --git a/drivers/staging/comedi/drivers/s626.c b/drivers/staging/comedi/drivers/s626.c
> index 084a8e7b9fc2..c159416662fd 100644
> --- a/drivers/staging/comedi/drivers/s626.c
> +++ b/drivers/staging/comedi/drivers/s626.c
> @@ -2130,13 +2130,15 @@ static int s626_allocate_dma_buffers(struct comedi_device *dev)
>   	void *addr;
>   	dma_addr_t appdma;
>   
> -	addr = pci_alloc_consistent(pcidev, S626_DMABUF_SIZE, &appdma);
> +	addr = dma_alloc_coherent(&pcidev->dev, S626_DMABUF_SIZE, &appdma,
> +				  GFP_ATOMIC);
>   	if (!addr)
>   		return -ENOMEM;
>   	devpriv->ana_buf.logical_base = addr;
>   	devpriv->ana_buf.physical_base = appdma;
>   
> -	addr = pci_alloc_consistent(pcidev, S626_DMABUF_SIZE, &appdma);
> +	addr = dma_alloc_coherent(&pcidev->dev, S626_DMABUF_SIZE, &appdma,
> +				  GFP_ATOMIC);
>   	if (!addr)
>   		return -ENOMEM;
>   	devpriv->rps_buf.logical_base = addr;
's626_allocate_dma_buffers()' is only called from 's626_auto_attach()'.
In this function, around the call to 's626_allocate_dma_buffers()', you 
can see:
   - a few lines before, a call to 'comedi_alloc_devpriv()'

   - a few lines after, a call to 'comedi_alloc_subdevices()'

These 2 functions make some memory allocation using GFP_KERNEL.

So it is likely that using GFP_KERNEL in your proposal is also valid.


Just my 2c.

CJ