Re: bus_dmamap_sync: bug?

[mlelstv%serpens.de@localhost (Michael van Elst)]

max%m00nbsd.net@localhost (Maxime Villard) writes:

>Seems to me there is a bug in x86's bus_dmamap_sync(). When the DMA buffer
>is an UIO, and we bounce, the 'offset' argument is not taken into account
>properly.

>We should do
>	copy uio+off into bouncebuf+off
>but we do
>	copy uio+0 into bouncebuf+off

>A quick glance shows that other architectures have the same problem, like
>ARM32's _bus_dma_uiomove().


Yes, uio and offset != 0 doesn't work. The function is modelled after
uiomove() which doesn't support offsets into the scatter/gather buffers.

We can either change bus_dma_uiomove to take an extra offset
argument and calculate the beginning of the transfer, or assert
that uio DMA buffers aren't used with an offset != 0.

N.B. the uio->uio_offset field tracks position of the scatter/gather
buffers in a larger object and corresponds to uio->uio_iov[0]->iov_base.
The real uiomove() adjusts it to reflect the completed I/O operation,
for bus_dma_uiomove() it has no meaning.

-- 
-- 
                                Michael van Elst
Internet: mlelstv%serpens.de@localhost
                                "A potential Snark may lurk in every tree."