Re: [PATCH 4.4 101/105] spi: Fix scatterlist elements size in spi_map_buf

[Ben Hutchings]

On Sun, 2018-07-01 at 18:02 +0200, Greg Kroah-Hartman wrote:
> 4.4-stable review patch.ÂÂIf anyone has any objections, please let me know.
> 
> ------------------
> 
> From: Maxime Chevallier <maxime.chevallier@xxxxxxxxxxx>
> 
> commit ce99319a182fe766be67f96338386f3ec73e321c upstream.
> 
> When SPI transfers can be offloaded using DMA, the SPI core need to
> build a scatterlist to make sure that the buffer to be transferred is
> dma-able.
> 
> This patch fixes the scatterlist entry size computation in the case
> where the maximum acceptable scatterlist entry supported by the DMA
> controller is less than PAGE_SIZE, when the buffer is vmalloced.
> 
> For each entry, the actual size is given by the minimum between the
> desc_len (which is the max buffer size supported by the DMA controller)
> and the remaining buffer length until we cross a page boundary.
> 
> Fixes: 65598c13fd66 ("spi: Fix per-page mapping of unaligned vmalloc-ed buffer")
> Signed-off-by: Maxime Chevallier <maxime.chevallier@xxxxxxxxxxx>
> Signed-off-by: Mark Brown <broonie@xxxxxxxxxx>
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Sudip Mukherjee <sudipm.mukherjee@xxxxxxxxx>
> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> ---
> Âdrivers/spi/spi.c |ÂÂÂ10 ++++++++--
> Â1 file changed, 8 insertions(+), 2 deletions(-)
> 
> --- a/drivers/spi/spi.c
> +++ b/drivers/spi/spi.c
> @@ -707,8 +707,14 @@ static int spi_map_buf(struct spi_master
> Â	for (i = 0; i < sgs; i++) {
> Â
> Â		if (vmalloced_buf) { 
> -			min = min_t(size_t,
> -				ÂÂÂÂlen, desc_len - offset_in_page(buf));
> +			/*
> +			Â* Next scatterlist entry size is the minimum between
> +			Â* the desc_len and the remaining buffer length that
> +			Â* fits in a page.
> +			Â*/
> +			min = min_t(size_t, desc_len,
> +				ÂÂÂÂmin_t(size_t, len,
> +					ÂÂPAGE_SIZE - offset_in_page(buf)));
> Â			vm_page = vmalloc_to_page(buf);
> Â			if (!vm_page) {
> Â				sg_free_table(sgt);

I think this makes sense upstream, but in 4.4 it seems to be a no-op
since desc_len is always set to PAGE_SIZE when vmalloced_buf is true.
This fix depends on (at least) commit df88e91bbfd1 "spi: respect the
maximum segment size of DMA device" which went into 4.6.

Ben.

-- 
Ben Hutchings, Software Developer                Â        Codethink Ltd
https://www.codethink.co.uk/                 Dale House, 35 Dale Street
                                     Manchester, M1 2HF, United Kingdom