Re: [RFC PATCH 05/13] staging: vchiq_arm: Add 36-bit address support

[Krzysztof Kozlowski <krzk@kernel.org>]

On 03/03/2024 16:10, Maarten Vanraes wrote:
> From: Phil Elwell <phil@raspberrypi.com>
> 
> Conditional on a new compatible string, change the pagelist encoding
> such that the top 24 bits are the pfn, leaving 8 bits for run length
> (-1), giving a 36-bit address range.
> 
> Manage the split between addresses for the VPU and addresses for the
> 40-bit DMA controller with a dedicated DMA device pointer that on non-
> BCM2711 platforms is the same as the main VCHIQ device. This allows
> the VCHIQ node to stay in the usual place in the DT.
> 
> Signed-off-by: Phil Elwell <phil@raspberrypi.com>
> Signed-off-by: Maarten Vanraes <maarten@rmail.be>
> ---


>  
>  	/* Partial cache lines (fragments) require special measures */
> @@ -413,7 +461,7 @@ free_pagelist(struct vchiq_instance *instance, struct vchiq_pagelist_info *pagel
>  	 * NOTE: dma_unmap_sg must be called before the
>  	 * cpu can touch any of the data/pages.
>  	 */
> -	dma_unmap_sg(instance->state->dev, pagelistinfo->scatterlist,
> +	dma_unmap_sg(g_dma_dev, pagelistinfo->scatterlist,
>  		     pagelistinfo->num_pages, pagelistinfo->dma_dir);
>  	pagelistinfo->scatterlist_mapped = 0;
>  
> @@ -468,6 +516,7 @@ free_pagelist(struct vchiq_instance *instance, struct vchiq_pagelist_info *pagel
>  static int vchiq_platform_init(struct platform_device *pdev, struct vchiq_state *state)
>  {
>  	struct device *dev = &pdev->dev;
> +	struct device *dma_dev = NULL;
>  	struct vchiq_drvdata *drvdata = platform_get_drvdata(pdev);
>  	struct rpi_firmware *fw = drvdata->fw;
>  	struct vchiq_slot_zero *vchiq_slot_zero;
> @@ -489,6 +538,24 @@ static int vchiq_platform_init(struct platform_device *pdev, struct vchiq_state
>  	g_cache_line_size = drvdata->cache_line_size;
>  	g_fragments_size = 2 * g_cache_line_size;
>  
> +	if (drvdata->use_36bit_addrs) {
> +		struct device_node *dma_node =
> +			of_find_compatible_node(NULL, NULL, "brcm,bcm2711-dma");

Why do you look up nodes by compatible instead of phandle? That's
fragile. Makes relationship undocumented, probing not ordered and is
completely non-portable between different devices.

You have basically here deferred probe issue.

Best regards,
Krzysztof