ipq806x: do not allocate coherent memory in dma engine hotpath

The available amount of coherent DMA memory is very limited. On Linux
4.4 this issue was worked around by increasing the pool size.

It turns out that using coherent memory here is completely unnecessary.
This change reworks the driver code to use kzalloc+dma_map_single
instead.

Signed-off-by: Felix Fietkau <nbd@nbd.name>

target/linux/ipq806x/patches-4.9/0002-dmaengine-Add-ADM-driver.patch

@@ -49,7 +49,7 @@ Signed-off-by: Thomas Pedersen <twp@codeaurora.org>
 +obj-$(CONFIG_QCOM_ADM) += qcom_adm.o
 --- /dev/null
 +++ b/drivers/dma/qcom/qcom_adm.c
-@@ -0,0 +1,900 @@
+@@ -0,0 +1,914 @@
 +/*
 + * Copyright (c) 2013-2015, The Linux Foundation. All rights reserved.
 + *
@@ -406,6 +406,7 @@ Signed-off-by: Thomas Pedersen <twp@codeaurora.org>
 +	struct adm_device *adev = achan->adev;
 +	struct adm_async_desc *async_desc;
 +	struct scatterlist *sg;
++	dma_addr_t cple_addr;
 +	u32 i, burst;
 +	u32 single_count = 0, box_count = 0, crci = 0;
 +	void *desc;
@@ -454,7 +455,7 @@ Signed-off-by: Thomas Pedersen <twp@codeaurora.org>
 +		}
 +	}
 +
-+	async_desc = kzalloc(sizeof(*async_desc), GFP_NOWAIT);
++	async_desc = kzalloc(sizeof(*async_desc), GFP_ATOMIC);
 +	if (!async_desc)
 +		return ERR_PTR(-ENOMEM);
 +
@@ -467,13 +468,9 @@ Signed-off-by: Thomas Pedersen <twp@codeaurora.org>
 +				box_count * sizeof(struct adm_desc_hw_box) +
 +				sizeof(*cple) + 2 * ADM_DESC_ALIGN;
 +
-+	async_desc->cpl = dma_alloc_writecombine(adev->dev, async_desc->dma_len,
-+				&async_desc->dma_addr, GFP_NOWAIT);
-+
-+	if (!async_desc->cpl) {
-+		kfree(async_desc);
-+		return ERR_PTR(-ENOMEM);
-+	}
++	async_desc->cpl = kzalloc(async_desc->dma_len, GFP_ATOMIC);
++	if (!async_desc->cpl)
++		goto free;
 +
 +	async_desc->adev = adev;
 +
@@ -481,10 +478,6 @@ Signed-off-by: Thomas Pedersen <twp@codeaurora.org>
 +	cple = PTR_ALIGN(async_desc->cpl, ADM_DESC_ALIGN);
 +	desc = PTR_ALIGN(cple + 1, ADM_DESC_ALIGN);
 +
-+	/* init cmd list */
-+	*cple = ADM_CPLE_LP;
-+	*cple |= (desc - async_desc->cpl + async_desc->dma_addr) >> 3;
-+
 +	for_each_sg(sgl, sg, sg_len, i) {
 +		async_desc->length += sg_dma_len(sg);
 +
@@ -496,7 +489,27 @@ Signed-off-by: Thomas Pedersen <twp@codeaurora.org>
 +							   direction);
 +	}
 +
++	async_desc->dma_addr = dma_map_single(adev->dev, async_desc->cpl,
++					      async_desc->dma_len,
++					      DMA_TO_DEVICE);
++	if (dma_mapping_error(adev->dev, async_desc->dma_addr))
++		goto free;
++
++	cple_addr = async_desc->dma_addr + ((void *)cple - async_desc->cpl);
++
++	/* init cmd list */
++	dma_sync_single_for_cpu(adev->dev, cple_addr, sizeof(*cple),
++				DMA_TO_DEVICE);
++	*cple = ADM_CPLE_LP;
++	*cple |= (async_desc->dma_addr + ADM_DESC_ALIGN) >> 3;
++	dma_sync_single_for_device(adev->dev, cple_addr, sizeof(*cple),
++				   DMA_TO_DEVICE);
++
 +	return vchan_tx_prep(&achan->vc, &async_desc->vd, flags);
++
++free:
++	kfree(async_desc);
++	return ERR_PTR(-ENOMEM);
 +}
 +
 +/**
@@ -721,8 +734,9 @@ Signed-off-by: Thomas Pedersen <twp@codeaurora.org>
 +	struct adm_async_desc *async_desc = container_of(vd,
 +			struct adm_async_desc, vd);
 +
-+	dma_free_writecombine(async_desc->adev->dev, async_desc->dma_len,
-+		async_desc->cpl, async_desc->dma_addr);
++	dma_unmap_single(async_desc->adev->dev, async_desc->dma_addr,
++			 async_desc->dma_len, DMA_TO_DEVICE);
++	kfree(async_desc->cpl);
 +	kfree(async_desc);
 +}
 +