Re: selective block polling and preadv2/pwritev2 revisited

quoted

quoted

quoted

quoted

"Sagi" == Sagi Grimberg [off-list ref] writes:

quoted

The idea that was floated aroud a few times was to have a flag where
the first half of the vectors would be the data, and the second half
the metadata.

Sagi> This means that the user would need to pass iovec entries of 8
Sagi> bytes correct? Seems like a waste for large IOs (sorry for
Sagi> diverging from the subject)

The assumption was that there would be a 1:1 mapping between the number
of data buffers and the metadata ditto. But nothing says that a data
iovec entry is limited in size to a single sector.

The other option to have a single iovec at the end representing the
metadata for all data buffers. I think there are valid use cases for
either approach and we may end up having to support both via a flag.

-- 
Martin K. Petersen	Oracle Linux Engineering

Hi Martin,

Sagi> This means that the user would need to pass iovec entries of 8
Sagi> bytes correct? Seems like a waste for large IOs (sorry for
Sagi> diverging from the subject)

The assumption was that there would be a 1:1 mapping between the number
of data buffers and the metadata ditto. But nothing says that a data
iovec entry is limited in size to a single sector.

Yea... I meant 1:1, I got confused on the 8 bytes comment...

The other option to have a single iovec at the end representing the
metadata for all data buffers. I think there are valid use cases for
either approach and we may end up having to support both via a flag.

Either approach presents limitation, but I guess user-space can deal
with it...