On 02/03/11 17:55, Paul Eggert wrote: > On 03/02/2011 04:53 AM, Pádraig Brady wrote: >> + /* Enable 'fullblock' with 'direct' or 'cio' as we don't want to >> + write partial blocks to output, which would disable O_DIRECT. An >> + alternative would be to enable C_TWOBUFS to accumulate full output >> + blocks. However that wouldn't work when a count is specified, and >> + is also less efficient. */ >> + if (output_flags & (O_DIRECT | O_CIO)) >> + input_flags |= O_FULLBLOCK; > > I'm afraid this patch feels wrong somehow. It's conflating four > issues: counting, efficiency, disabling O_DIRECT, and disabling O_CIO. > Some thoughts: > > 1. Offhand I don't see why O_CIO has anything to do with blocking; > why is it mentioned here? > > 2. If C_TWOBUFS is already in effect, then dd needn't set > O_FULLBLOCK, since C_TWOBUFS already prevents the disabling of > O_DIRECT. > > 3. The counting issue is independent of oflag=direct or oflag=cio: > any solution for counting should work regardless of oflag settings. > > The more I think about it, the more I suspect that this > O_FULLBLOCK-inferring code should be omitted. It adds little benefit, as > it covers a rare combination of flags that is likely to be used only > by experts, who should know the gotchas in this area anyway. And the > extra complexity in documentation will penalize everybody who reads > it, even the non-experts.
Hmm I actually thought of something that may be better. 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. cheers, Pádraig.
