From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from mail.bootlin.com ([62.4.15.54]) by bombadil.infradead.org with esmtp (Exim 4.90_1 #2 (Red Hat Linux)) id 1fIr55-0004F7-Q5 for linux-mtd@lists.infradead.org; Wed, 16 May 2018 07:42:54 +0000 Date: Wed, 16 May 2018 09:42:27 +0200 From: Miquel Raynal To: Chris Moore Cc: Boris Brezillon , "Wan, Jane (Nokia - US/Sunnyvale)" , "dwmw2@infradead.org" , "computersforpeace@gmail.com" , "richard@nod.at" , "marek.vasut@gmail.com" , "yamada.masahiro@socionext.com" , "prabhakar.kushwaha@nxp.com" , "shawnguo@kernel.org" , "jagdish.gediya@nxp.com" , "shreeya.patel23498@gmail.com" , "Bos, Ties (Nokia - US/Sunnyvale)" , "linux-mtd@lists.infradead.org" , "linux-kernel@vger.kernel.org" Subject: Re: [PATCH v7] mtd: rawnand: use bit-wise majority to recover the contents of ONFI parameter Message-ID: <20180516094227.14132e74@xps13> In-Reply-To: References: <20180515093429.34902670@bbrezillon> MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: quoted-printable List-Id: Linux MTD discussion mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Hi Chris, > >>> +static void nand_bit_wise_majority(const void **srcbufs, > >>> + unsigned int nsrcbufs, > >>> + void *dstbuf, > >>> + unsigned int bufsize) > >>> +{ > >>> + int i, j, k; > >>> + > >>> + for (i =3D 0; i < bufsize; i++) { > >>> + u8 cnt, val; > >>> + > >>> + val =3D 0; > >>> + for (j =3D 0; j < 8; j++) { > >>> + cnt =3D 0; > >>> + for (k =3D 0; k < nsrcbufs; k++) { > >>> + const u8 *srcbuf =3D srcbufs[k]; > >>> + > >>> + if (srcbuf[i] & BIT(j)) > >>> + cnt++; > >>> + } > >>> + if (cnt > nsrcbufs / 2) > >>> + val |=3D BIT(j); > >>> + } > >>> + ((u8 *)dstbuf)[i] =3D val; > >>> + } > >>> +} > >>> + > >>> +/* > >>> * Check if the NAND chip is ONFI compliant, returns 1 if it is, 0= otherwise. > >>> */ > >>> static int nand_flash_detect_onfi(struct nand_chip *chip) > >>> @@ -5102,7 +5131,7 @@ static int nand_flash_detect_onfi(struct nand_c= hip *chip) > >>> return 0; =20 > >>> >>> /* ONFI chip: allocate a buffer to hold its parameter page= */ =20 > >>> - p =3D kzalloc(sizeof(*p), GFP_KERNEL); > >>> + p =3D kzalloc((sizeof(*p) * 3), GFP_KERNEL); > >>> if (!p) > >>> return -ENOMEM; =20 > >>> >>> @@ -5113,21 +5142,32 @@ static int nand_flash_detect_onfi(stru= ct nand_chip *chip) =20 > >>> } =20 > >>> >>> for (i =3D 0; i < 3; i++) { =20 > >>> - ret =3D nand_read_data_op(chip, p, sizeof(*p), true); > >>> + ret =3D nand_read_data_op(chip, &p[i], sizeof(*p), true); > >>> if (ret) { > >>> ret =3D 0; > >>> goto free_onfi_param_page; > >>> } =20 > >>> >>> - if (onfi_crc16(ONFI_CRC_BASE, (uint8_t *)p, 254) =3D=3D =20 > >>> + if (onfi_crc16(ONFI_CRC_BASE, (u8 *)&p[i], 254) =3D=3D > >>> le16_to_cpu(p->crc)) { > >>> + if (i) > >>> + memcpy(p, &p[i], sizeof(*p)); > >>> break; > >>> } > >>> } =20 > >>> >>> if (i =3D=3D 3) { =20 > >>> - pr_err("Could not find valid ONFI parameter page; aborting\n"); > >>> - goto free_onfi_param_page; > >>> + const void *srcbufs[3] =3D {p, p + 1, p + 2}; > >>> + > >>> + pr_warn("Could not find a valid ONFI parameter page, trying bit-wi= se majority to recover it\n"); > >>> + nand_bit_wise_majority(srcbufs, ARRAY_SIZE(srcbufs), p, > >>> + sizeof(*p)); > >>> + > >>> + if (onfi_crc16(ONFI_CRC_BASE, (u8 *)p, 254) !=3D > >>> + le16_to_cpu(p->crc)) { > >>> + pr_err("ONFI parameter recovery failed, aborting\n"); > >>> + goto free_onfi_param_page; > >>> + } > >>> } =20 > >>> >>> /* Check version */ =20 > >> This version is still hard coded for a three sample bitwise majority v= ote. > >> So why not use the method which I suggested previously for v2 and which > >> I repeat below? =20 > > Because I want the nand_bit_wise_majority() function to work with > > nsrcbufs > 3 (the ONFI spec says there's at least 3 copy of the param > > page, but NAND vendor can decide to put more). Also, if the X copies of > > the PARAM are corrupted (which is rather unlikely), that means we > > already spent quite a lot of time reading the different copies and > > calculating the CRC, so I think we don't care about perf optimizations > > when doing bit-wise majority. > > =20 > >> The three sample bitwise majority can be implemented without bit level > >> manipulation using the identity: > >> majority3(a, b, c) =3D (a & b) | (a & c) | (b & c) > >> This can be factorized slightly to (a & (b | c)) | (b & c) > >> This enables the operation to be performed 8, 16, 32 or even 64 bits at > >> a time depending on the hardware. > >> > >> This method is not only faster and but also more compact. > >> =20 >=20 > I do understand that the ONFI specifications permit more than 3 copies. > However elsewhere the proposed code shows no intention of handling other = cases. > The constant 3 is hard coded in the following lines extracted from the pr= oposed code: > ... > +=C2=A0=C2=A0=C2=A0 p =3D kzalloc((sizeof(*p) * 3), GFP_KERNEL); > ... > =C2=A0=C2=A0=C2=A0=C2=A0 for (i =3D 0; i < 3; i++) { > ... > =C2=A0=C2=A0=C2=A0=C2=A0 if (i =3D=3D 3) { > ... > +=C2=A0=C2=A0=C2=A0 =C2=A0=C2=A0=C2=A0 const void *srcbufs[3] =3D {p, p += 1, p + 2}; >=20 > Moreover the last of these is difficult to generalize. Indeed, this is something to improve. I think Boris' request was to prepare changes like this one, to avoid the situation where the code does not scale (like this 'p, p + 1, p + 2'). Thanks, Miqu=C3=A8l