[RFC,v3 3/5] spi: add Mediatek SPI Nand controller driver
xiangsheng.hou
xiangsheng.hou at mediatek.com
Fri Nov 12 00:40:24 PST 2021
Hi Miquel,
On Tue, 2021-11-09 at 12:46 +0100, Miquel Raynal wrote:
> Hi Xiangsheng,
>
> xiangsheng.hou at mediatek.com wrote on Fri, 22 Oct 2021 10:40:19 +0800:
>
> > This version the SPI driver cowork with MTK pipelined
> > HW ECC engine.
> >
> > Signed-off-by: Xiangsheng Hou <xiangsheng.hou at mediatek.com>
> > ---
> > +
> > +static int mtk_snfi_ecc_init(struct nand_device *nand)
> > +{
> > + struct nand_ecc_props *reqs = &nand->ecc.requirements;
> > + struct nand_ecc_props *user = &nand->ecc.user_conf;
> > + struct nand_ecc_props *conf = &nand->ecc.ctx.conf;
> > + struct mtk_snfi *snfi = mtk_nand_to_spi(nand);
> > + struct mtk_ecc_engine *eng;
> > + u32 spare, idx;
> > + int ret;
> > +
> > + eng = kzalloc(sizeof(*eng), GFP_KERNEL);
> > + if (!eng)
> > + return -ENOMEM;
> > +
> > + nand->ecc.ctx.priv = eng;
> > + nand->ecc.engine->priv = eng;
> > +
> > + /* Configure the correction depending on the NAND device
> > topology */
> > + if (user->step_size && user->strength) {
> > + conf->step_size = user->step_size;
> > + conf->strength = user->strength;
> > + } else if (reqs->step_size && reqs->strength) {
> > + conf->step_size = reqs->step_size;
> > + conf->strength = reqs->strength;
> > + }
> > +
> > + /*
> > + * align eccstrength and eccsize.
> > + * The MTK HW ECC engine only support 512 and 1024 eccsize.
> > + */
> > + if (conf->step_size < 1024) {
> > + if (nand->memorg.pagesize > 512 &&
> > + snfi->caps->max_sector_size > 512) {
> > + conf->step_size = 1024;
> > + conf->strength <<= 1;
> > + } else {
> > + conf->step_size = 512;
> > + }
> > + } else {
> > + conf->step_size = 1024;
> > + }
> > +
> > + ret = mtk_snfi_set_spare_per_sector(nand, snfi, &spare, &idx);
> > +
> > + /* These will be used by the snfi driver */
> > + snfi->ecc.page_size = nand->memorg.pagesize;
> > + snfi->ecc.spare_per_sector = spare;
> > + snfi->ecc.spare_idx = idx;
> > + snfi->ecc.sectors = nand->memorg.pagesize / conf->step_size;
> > +
> > + /* These will be used by HW ECC engine */
> > + eng->oob_per_sector = spare;
> > + eng->nsteps = snfi->ecc.sectors;
>
> I believe most of this function should move into mtk_ecc.c.
I'm also confused about this when coding.
Obviously, most of the code logic belong to the ecc driver.
However, some ecc related parameter have to config at the snfi
controller register, such as sector size, available oob bytes for each
sector used to calculate ecc level. The are all attribute defined at
the snfi controller register.
How about I move these code logic, sector size and useable spare size
for each sector which belog to the snfi controller attribute to the ecc
driver, parse and config when mtk_snfi_ecc_prepare_io_req in the snfi
driver?
>
> > +
> > + return ret;
> > +}
> > +
> > +static int mtk_snfi_ecc_init_ctx(struct nand_device *nand)
> > +{
> > + struct nand_ecc_engine_ops *ops = mtk_ecc_get_pipelined_ops();
> > + int ret;
> > +
> > + ret = mtk_snfi_ecc_init(nand);
> > + if (ret) {
> > + pr_info("mtk snfi ecc init fail!\n");
> > + return ret;
> > + }
> > +
> > + return ops->init_ctx(nand);
> > +}
> > +
> > +static void mtk_snfi_ecc_cleanup_ctx(struct nand_device *nand)
> > +{
> > + struct nand_ecc_engine_ops *ops = mtk_ecc_get_pipelined_ops();
> > +
> > + ops->cleanup_ctx(nand);
> > +}
> > +
> > +static int mtk_snfi_prepare_for_ecc(struct nand_device *nand,
> > + struct mtk_snfi *snfi)
> > +{
> > + struct nand_ecc_props *conf = &nand->ecc.ctx.conf;
> > + struct mtk_ecc_engine *eng = nand->ecc.ctx.priv;
> > + u32 val;
> > +
> > + switch (nand->memorg.pagesize) {
> > + case 512:
> > + val = PAGEFMT_512_2K | PAGEFMT_SEC_SEL_512;
> > + break;
> > + case KB(2):
> > + if (conf->step_size == 512)
> > + val = PAGEFMT_2K_4K | PAGEFMT_SEC_SEL_512;
> > + else
> > + val = PAGEFMT_512_2K;
> > + break;
> > + case KB(4):
> > + if (conf->step_size == 512)
> > + val = PAGEFMT_4K_8K | PAGEFMT_SEC_SEL_512;
> > + else
> > + val = PAGEFMT_2K_4K;
> > + break;
> > + case KB(8):
> > + if (conf->step_size == 512)
> > + val = PAGEFMT_8K_16K | PAGEFMT_SEC_SEL_512;
> > + else
> > + val = PAGEFMT_4K_8K;
> > + break;
> > + case KB(16):
> > + val = PAGEFMT_8K_16K;
> > + break;
> > + default:
> > + dev_err(snfi->dev, "invalid page len: %d\n",
> > + nand->memorg.pagesize);
> > + return -EINVAL;
> > + }
> > +
> > + snfi->fdm_size = eng->fdm_size;
> > + snfi->fdm_ecc_size = eng->fdm_ecc_size;
> > +
> > + val |= snfi->ecc.spare_idx << PAGEFMT_SPARE_SHIFT;
> > + val |= snfi->fdm_size << PAGEFMT_FDM_SHIFT;
> > + val |= snfi->fdm_ecc_size << PAGEFMT_FDM_ECC_SHIFT;
> > + writel(val, snfi->regs + NFI_PAGEFMT);
> > +
> > + return 0;
> > +}
> > +
> > +static int mtk_snfi_ecc_prepare_io_req(struct nand_device *nand,
> > + struct nand_page_io_req
> > *req)
> > +{
> > + struct nand_ecc_engine_ops *ops = mtk_ecc_get_pipelined_ops();
> > + struct mtk_snfi *snfi = mtk_nand_to_spi(nand);
> > + int ret;
> > +
> > + snfi->ecc.enabled = (req->mode != MTD_OPS_RAW);
>
> Shouldn't you check ecc.enabled before calling prepare_for_ecc ?
The funcion name may make you confused.
Actually, the prepare_for_ecc function did not include any logic about
ecc enable or not.Only config pagesize/sparesize/sector size to snfi
controller register.
I will modify the function name mtk_snfi_prepare_for_ecc,
mtk_snfi_config for example.
>
> > + ret = mtk_snfi_prepare_for_ecc(nand, snfi);
> > + if (ret)
> > + return ret;
> > +
> > + return ops->prepare_io_req(nand, req);
> > +}
> > +
> > +static int mtk_snfi_ecc_finish_io_req(struct nand_device *nand,
> > + struct nand_page_io_req
> > *req)
> > +{
> > + struct nand_ecc_engine_ops *ops = mtk_ecc_get_pipelined_ops();
> > + struct mtk_ecc_engine *eng = nand->ecc.ctx.priv;
> > + struct mtk_snfi *snfi = mtk_nand_to_spi(nand);
> > +
> > + if (snfi->ecc.enabled) {
>
> I am currently looking at a better way of keeping this
> information, while being safer against concurrent accesses. So
> far parallel operations are not supported so this is safe, but
> we might improve the core in a little while and I don't want
> this to be an issue. My next iteration on the Macronix engine will
> solve this.
>
Look forward to the next interation.
Thanks
Xiangsheng Hou
More information about the linux-mtd
mailing list