Hi Stephen,
I like to make a small comment or perhaps suggestions to the
raw IO changes which I hope make it into the Linux stable
(production) kernels a.s.a.p.
Your kernel API does only allow 512 byte memory aligned user space
buffers ... this for technical reasons. Performing in kernel memory
copies are left out by you because of performance I guess.
However for portability and allowing existing applications to take
advantage of raw disk devices I strongly suggest you to alter the
driver and document it.
I.e. now you return an EINVAL (errno 22) whenever the user buffer
address is *not* 512 bytes aligned. Such cripples for example a
simple 'dd if=/dev/raw1 of=....'
Since you already perform a check to detect the misalignment to
return EINVAL you may decide to invoke a kernel internal memcopy
instead. This to a kernel private buffer and correcting the
alignment. Documenting the driver 'man 8 raw' may describe the
details about the performance loss for I/O's on file descriptors
where the user buffer was not aligned. Perhaps even equip the
driver with an ioctl(2) interface allowing it to either trap and
return EINVAL for cases where only correct alignment and I/O
lengths are provided from the user process.
Regards,
-- Wim ten Have.
> Hi,
>
> Sorry to follow so quickly from the last version but there was a stupid
> one-liner fault which managed to sneak into the 2.2 patch I sent out
> yesterday. Get
>
>
> ftp://ftp.uk.linux.org/pub/linux/sct/fs/raw-io/raw-19990728.tar.gz
>
> Change summary:
>
> Raw device support v0.6, 28 July 1999
>
> Changed since v0.5: Removed 2.3 support: it's now in the mainstream
> kernel as of 2.3.12-pre7.
>
> Added diffs for both plain 2.2.5 and RedHat-6.0
> 2.2.5.
>
> Fixed a one-liner compile problem which affected
> the 2.2 patch as distributed.
>
> --Stephen
>
