Re: [PATCH] bsg: iovec support

[Douglas Gilbert <dougg@torque.net>]

FUJITA Tomonori wrote:
> From: Pete Wyckoff <pw@osc.edu>
> Subject: [PATCH] bsg: iovec support
> Date: Thu, 1 Mar 2007 17:29:08 -0500
> 
>> Support vectored IO as in SGv3.  The iovec structure uses explicit
>> sizes to avoid the need for compat conversion.
>>
>> Signed-off-by: Pete Wyckoff <pw@osc.edu>
>> ---
>>
>> My application definitely can take advantage of scatter/gather IO,
>> which is supported in sgv3 but not in the bsg implementation of sgv4.
>> I understand Tomo's concerns about code bloat and the need for
>> 32/64 compat translations, but this will make things much easier on
>> users of bsg who read or write out of multiple buffers in a single
>> SCSI operation.
> 
> (snip)
> 
>> + * Vector of address/length pairs, used when dout_iovec_count (or din_)
>> + * is non-zero.  In that case, dout_xferp is a list of struct sg_io_v4_vec
>> + * and dout_iovec_count is the number of entries in that list.  dout_xfer_len
>> + * is the total length of the list.  Note the use of u64 instead of a
>> + * native pointer to avoid compat issues, and padding to avoid structure
>> + * alignment problems.
>> + */
>> +struct sg_io_v4_vec {
>> +	__u64 iov_base;
>> +	__u32 iov_len;
>> +	__u32 __pad1;
>> +};
> 
> I don't think that it's a good idea to add a new scatter/gather
> structure and export it to user space.

User space scatter gather is not a new feature.
It is defined and works in sg v3.

It was also partially defined in sg v4 and dropped
out in the bsg implementation. I agree with Pete that
it should be put back.

Pete is also suggesting (shown above) a revised sg_io_vec
structure that uses a uint64_t for the pointer to simplify
32, 64 bit thunking.

> bsg can support scatter/gather IO with ioctl (SG_IO) easily (I mean,
> without adding ugly compat code to bsg.c). I guess that SG_IO doesn't
> work for you because it works synchronously. However, all system calls
> might work asynchronously in the future.

User space scatter gather is completely decoupled from
in-kernel scatter gather lists built for HBA DMA
engines. Same technique but at different levels.

Someone might user space scatter gather to efficiently
fetch several OSD objects implemented in a block device
as adjacent blocks

Doug Gilbert