Given that errors are rare (at least, they had BETTER be rare!), why not
just re-issue the requests, but (somehow) prevent them from being merged.
You _don't_ need to optimize for the error case, as long as you handle it
properly.
That would be the right answer, except that end_request() doesn't know
which blocks were read in correctly, and which blocks failed. The
device driver needs to give an indication of how far in the read or
write request it got before it found a failure, and then have
end_request reissue those blocks that haven't yet been read.
The "quick hack" alternative which wouldn't require making changes to
all of the block device drivers would be to change end_request() that
when there's an error in a multi-block request, it breaks apart the
request into individual requests, and then retries them *all*, since it
doesn't know which block failed. This would really slow down reads in
the face of errors, but at least it's correct.
However, if the plan is to rewrite the entire block device interface
anyway, we might as well do it right instead of putting in a kludge.
- Ted