On 03/02/2011 03:29 PM, Pádraig Brady wrote:
The O_DIRECT -> normal switch done in iwrite() for the
last write, should only happen once.
If it happens more than that, we can print a warning,
suggesting the use of iflag=fullblock.
Yes, that sounds like a better approach.
Thanks for thinking this through.
I looked into O_CIO and it appears that it does
imply O_DIRECT, but O_DIRECT does not have the
problems on AIX (which has O_CIO) that I guess it
has on GNU/Linux. Quoting from
<http://www-03.ibm.com/systems/resources/systems_p_os_aix_whitepapers_db_perf_aix.pdf>
The use of Direct I/O requires that certain
alignment and length restrictions be met by the
application’s I/O requests. Table 1 lists these
requirements for JFS2. Failure to meet these
requirements causes reads and writes to be done
using normal cached I/O, but after the data is
transferred to the application buffer, the cached
copy is discarded.
which suggests that iwrite's fcntl in iwrite is
not needed on AIX. I don't know if this is worth
optimizing, but perhaps it's worth a comment as
to why we worry about O_DIRECT but not O_CIO in
iwrite.
