Simon Kagstrom wrote: > On Wed, 19 Aug 2009 16:06:12 +0800 > "Liu Dave-R63238" <[email protected]> wrote: > > >>> Well, that's what the other patch I sent does (the link above), but >>> there were wishes then to handle this above the driver layer, hence >>> this patch :-) >>> >>> I'm fine with either way, but if there are other drivers with >>> alignment requirements, I'd prefer this variant. >>> >> I believe some Freescale's CPM/QE controllers need care the >> alignement requirement. >> >> If you don't care the performance, you can malloc memory to >> get the aligned buffer pool in the initiaize phase, >> then use the buffer pool to resolve your alignment issue, of course >> It needs a copy, but it avoids to malloc/free buffer in run-time. >> > > The malloc is only done once - the first time a non-aligned send is > done. After that it's just reused until the alignment requirement > exceeds the current allocation (if another interface is set > as current). > > So I think the performance impact is very small. > > // Simon > _______________________________________________ > U-Boot mailing list > [email protected] > http://lists.denx.de/mailman/listinfo/u-boot > My preference is for the drivers to handle alignment. I like Dave's idea of creating a buffer (or multiple buffers) in initialization(), then memcpy'ing to to it in the send() routine. You could make it smart to only memcpy if the frame is not aligned as needed.
Your original patch created a buffer of size 9000, which I commented was probably bigger than necessary. I didn't see a response to that (which doesn't mean you didn't send one, but I get a lot of e-mail :) regards, Ben _______________________________________________ U-Boot mailing list [email protected] http://lists.denx.de/mailman/listinfo/u-boot
