From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752565AbeEPHmj convert rfc822-to-8bit (ORCPT ); Wed, 16 May 2018 03:42:39 -0400 Received: from mail.bootlin.com ([62.4.15.54]:57806 "EHLO mail.bootlin.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751449AbeEPHma (ORCPT ); Wed, 16 May 2018 03:42:30 -0400 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> Organization: Bootlin X-Mailer: Claws Mail 3.15.0-dirty (GTK+ 2.24.31; x86_64-pc-linux-gnu) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8BIT Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org 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 = 0; i < bufsize; i++) { > >>> + u8 cnt, val; > >>> + > >>> + val = 0; > >>> + for (j = 0; j < 8; j++) { > >>> + cnt = 0; > >>> + for (k = 0; k < nsrcbufs; k++) { > >>> + const u8 *srcbuf = srcbufs[k]; > >>> + > >>> + if (srcbuf[i] & BIT(j)) > >>> + cnt++; > >>> + } > >>> + if (cnt > nsrcbufs / 2) > >>> + val |= BIT(j); > >>> + } > >>> + ((u8 *)dstbuf)[i] = 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_chip *chip) > >>> return 0; > >>> >>> /* ONFI chip: allocate a buffer to hold its parameter page */ > >>> - p = kzalloc(sizeof(*p), GFP_KERNEL); > >>> + p = kzalloc((sizeof(*p) * 3), GFP_KERNEL); > >>> if (!p) > >>> return -ENOMEM; > >>> >>> @@ -5113,21 +5142,32 @@ static int nand_flash_detect_onfi(struct nand_chip *chip) > >>> } > >>> >>> for (i = 0; i < 3; i++) { > >>> - ret = nand_read_data_op(chip, p, sizeof(*p), true); > >>> + ret = nand_read_data_op(chip, &p[i], sizeof(*p), true); > >>> if (ret) { > >>> ret = 0; > >>> goto free_onfi_param_page; > >>> } > >>> >>> - if (onfi_crc16(ONFI_CRC_BASE, (uint8_t *)p, 254) == > >>> + if (onfi_crc16(ONFI_CRC_BASE, (u8 *)&p[i], 254) == > >>> le16_to_cpu(p->crc)) { > >>> + if (i) > >>> + memcpy(p, &p[i], sizeof(*p)); > >>> break; > >>> } > >>> } > >>> >>> if (i == 3) { > >>> - pr_err("Could not find valid ONFI parameter page; aborting\n"); > >>> - goto free_onfi_param_page; > >>> + const void *srcbufs[3] = {p, p + 1, p + 2}; > >>> + > >>> + pr_warn("Could not find a valid ONFI parameter page, trying bit-wise 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) != > >>> + le16_to_cpu(p->crc)) { > >>> + pr_err("ONFI parameter recovery failed, aborting\n"); > >>> + goto free_onfi_param_page; > >>> + } > >>> } > >>> >>> /* Check version */ > >> This version is still hard coded for a three sample bitwise majority vote. > >> So why not use the method which I suggested previously for v2 and which > >> I repeat below? > > 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. > > > >> The three sample bitwise majority can be implemented without bit level > >> manipulation using the identity: > >> majority3(a, b, c) = (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. > >> > > 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 proposed code: > ... > +    p = kzalloc((sizeof(*p) * 3), GFP_KERNEL); > ... >      for (i = 0; i < 3; i++) { > ... >      if (i == 3) { > ... > +        const void *srcbufs[3] = {p, p + 1, p + 2}; > > 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, Miquèl