spi/pxa2xx: Prevent DMA from transferring too many bytes
Subsystems (3)
PATCH 1/1 spi/pxa2xx: Prevent DMA from transferring too many bytes
In case we are doing DMA transfer and the size of the buffer is not multiple of 4 bytes the driver truncates that to 4-byte boundary and tries to handle remaining bytes using PIO.
Or that is what it tried to do. What actually happens is that it calls ALIGN() to the buffer size which aligns it to the next 4-byte boundary (doesn't truncate). Doing this results 1-3 bytes extra to be transferred. Furthermore we handle remaining bytes using PIO which results one extra byte to be transferred. In worst case the driver transfers 4 extra bytes.
While investigating this it turned out that the DMA hardware doesn't even have such limitation so we can solve this by dropping the code that tries to handle unaligned bytes.
--- drivers/spi/spi-pxa2xx-dma.c | 16 ---------------- 1 file changed, 16 deletions(-)
drivers/spi/spi-pxa2xx-dma.c +0 -16
@@ -29,18 +29,6 @@ static int pxa2xx_spi_map_dma_buffer(struct driver_data *drv_data,struct sg_table *sgt;void *buf, *pbuf; -/* -* Some DMA controllers have problems transferring buffers that are -* not multiple of 4 bytes. So we truncate the transfer so that it -* is suitable for such controllers, and handle the trailing bytes -* manually after the DMA completes. -* -* REVISIT: It would be better if this information could be -* retrieved directly from the DMA device in a similar way than -* ->copy_align etc. is done. -*/ - len =ALIGN (drv_data->len,4 ); -if (dir ==DMA_TO_DEVICE ) { dmadev = drv_data->tx_chan->device->dev; sgt = &drv_data->tx_sgt;
@@ -144,12 +132,8 @@ static void pxa2xx_spi_dma_transfer_complete(struct driver_data *drv_data,if (!error) {pxa2xx_spi_unmap_dma_buffers (drv_data); -/* Handle the last bytes of unaligned transfer */ drv_data->tx += drv_data->tx_map_len; - drv_data->write (drv_data); - drv_data->rx += drv_data->rx_map_len; - drv_data->read (drv_data); msg->actual_length += drv_data->len; msg->state =pxa2xx_spi_next_transfer (drv_data);
Signature · 1 line
2.0.0.rc2
Mark Brown
maintainerQuoted 3 lines · Mika Westerberg
In case we are doing DMA transfer and the size of the buffer is not multiple of 4 bytes the driver truncates that to 4-byte boundary and tries to handle remaining bytes using PIO.
...
Quoted 3 lines
While investigating this it turned out that the DMA hardware doesn't even have such limitation so we can solve this by dropping the code that tries to handle unaligned bytes.
Is this definitely the case for all of the IPs using this driver? It seems like something which might have been present in actual PXA implemenetations but got fixed in later revisons used with x86. Equally well the current code is clearly broken either way so I'm not sure that problems with older systems should be a barrier to merging the patch but it seems better to check.
Quoted 11 lines · Mark Brown
Quoted 3 lines · Mika Westerberg
In case we are doing DMA transfer and the size of the buffer is not multiple of 4 bytes the driver truncates that to 4-byte boundary and tries to handle remaining bytes using PIO.
...
Quoted 3 lines
While investigating this it turned out that the DMA hardware doesn't even have such limitation so we can solve this by dropping the code that tries to handle unaligned bytes.
Quoted 6 lines
Is this definitely the case for all of the IPs using this driver? It seems like something which might have been present in actual PXA implemenetations but got fixed in later revisons used with x86. Equally well the current code is clearly broken either way so I'm not sure that problems with older systems should be a barrier to merging the patch but it seems better to check.
This code came with x86 LPSS implementation originally. The PXA one, which lives in a different file (spi-pxa2xx-pxadma.c) didn't have any such checks AFAIK.
Mark Brown
maintainerQuoted 9 lines · Mika Westerberg
Quoted 6 lines · Mark Brown
Is this definitely the case for all of the IPs using this driver? It seems like something which might have been present in actual PXA implemenetations but got fixed in later revisons used with x86. Equally well the current code is clearly broken either way so I'm not sure that problems with older systems should be a barrier to merging the patch but it seems better to check.
Quoted 3 lines
This code came with x86 LPSS implementation originally. The PXA one, which lives in a different file (
spi-pxa2xx-pxadma.c) didn't have any such checks AFAIK.
OK, that should be fine then. The PXA platforms should be being converted over to use this file as part of the dmaengine transition on that platform so we can't assume it's Intel specific even though it was originally written for x86.
Quoted 10 lines · Mark Brown
Quoted 8 lines · Mika Westerberg
Quoted 6 lines · Mark Brown
Is this definitely the case for all of the IPs using this driver? It seems like something which might have been present in actual PXA implemenetations but got fixed in later revisons used with x86. Equally well the current code is clearly broken either way so I'm not sure that problems with older systems should be a barrier to merging the patch but it seems better to check.
Quoted 8 lines
Quoted 3 lines
This code came with x86 LPSS implementation originally. The PXA one, which lives in a different file (
spi-pxa2xx-pxadma.c) didn't have any such checks AFAIK.OK, that should be fine then. The PXA platforms should be being converted over to use this file as part of the dmaengine transition on that platform so we can't assume it's Intel specific even though it was originally written for x86.
Right and it isn't. However, this code (the one I'm removing with this patch) was developed on early Haswell machines where we encountered this restriction. With the hardware that is shipping the restriction is gone so I don't see much point keeping it there.
Quoted 3 lines · Mika Westerberg
In case we are doing DMA transfer and the size of the buffer is not multiple of 4 bytes the driver truncates that to 4-byte boundary and tries to handle remaining bytes using PIO.
Applied, thanks.