On 30 Mar 2015, scottwood@freescale.com wrote:
On Mon, 2015-03-30 at 22:40 +0200, Stefan Agner wrote:
However, if removing the caching in the driver would lead to a performance drop, I would rather prefer to keep it...
What is special about this controller, that caching makes sense here but not on other controllers? If it makes sense everywhere, then the upper layer is the place to do it.
While I observed some performance differences, but that is an aside. I think the mxc driver is similar in sub-page (or partial page) support. Before doing a sub-page write, it reads a full page and then copies the sub-page update to the buffer to be reprogrammed. This works fine if each and every sub-page has separate OOB data for software ECC. The hardware ECC of this vf610_nfc controller doesn't support this.
At least that is my understanding of this [mxc_nand] code,
case NAND_CMD_SEQIN: if (column >= mtd->writesize) { /* * before sending SEQIN command for partial write, * we need read one page out. FSL NFC does not support * partial write. It always sends out 512+ecc+512+ecc * for large page nand flash. But for small page nand * flash, it does support SPARE ONLY operation. */ if (host->pagesize_2k) { /* call ourself to read a page */ mxc_nand_command(mtd, NAND_CMD_READ0, 0, page_addr); }
in 'drivers/mtd/nand/mxc_nand.c'. So, originally this was mainly for sub-pages (an optimization but also functionality).
So the 'caching' is just to preserve read data before a partial write as the vf610 requires a full page (like the mxc). The same 'caching' was keep for 'write followed by read' when the ECC is successful.
I realize that the upper layers may not like this so I mentioned it; I guess that is proof that the patch is getting a good review and I am not causing problems for Stefan ;)
I also agree with Stefan that setting the SECSZ in the 2nd patch will result in less data transfer (and maybe less current/power and system noise). However, this will be inside the NFC controller. Talking over the AHB to the NFC controller is quite slow on the Vybrid. Certainly it seems a little more elegant to tell the controller to transfer nothing (and that we [programmers] expect nothing as a sort of documentation)?
I think it would be worthwhile to benchmark without the cache. Or maybe Stefan already has some numbers? Upper layers doing partial pages will definitely benefit with the 'cache'; we would also need more DDR memory as the NFC controller memory is being used as a scratch buffer.
Fwiw, Bill Pringlemeir.