On Wed, Mar 2, 2011 at 4:34 AM, Andrei Warkentin [off-list ref] wrote:

Before the generic improvements are made to the block layer, I think
there is some value
in implementing the (simpler) ones in mmc block code, as well as
expose an mmc block quirk interface by which its easy to add complex
workarounds. Some things will never be able to completely stay above
mmc block code, for example, when splitting up smaller accesses, you
need to be careful on the Toshiba card, since the 4th consecutive 8KB
block results in the entire 32KB getting pushed ?into the bigger 4MB
buffer. On our platform, there are a lot of accesses in the 16KB-32KB
range which benefit from the splitting. Data collected showed
splitting more than 32KB to have adverse effect on performance (I
guess that sort of makes sense, after all, why else would the
controller treat 4 consecutive 8KB accesses as a larger access and
treat it accordingly?) On the other hand, that data was collected on
code that used reliable write for every portion of the split access,
so I'm going to have to get some new data...

Just want to correct myself - any consecutive write that exceeds 8K
goes into the 4MB buffer.
Also, according to vendor, there is no performance penalty for using
reliable write.
This is why in the patch set, for splitting larger requests (to
improve lifetime by reducing the number of AU write/erase cycles) I
perform a reliable write for each split block set.