From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-13.8 required=3.0 tests=BAYES_00,DKIM_INVALID, DKIM_SIGNED,INCLUDES_CR_TRAILER,INCLUDES_PATCH,MAILING_LIST_MULTI, SPF_HELO_NONE,SPF_PASS autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 0C7E3C07E95 for ; Mon, 19 Jul 2021 18:39:11 +0000 (UTC) Received: from lists.sourceforge.net (lists.sourceforge.net [216.105.38.7]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by mail.kernel.org (Postfix) with ESMTPS id ACD4860FF4 for ; Mon, 19 Jul 2021 18:39:10 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org ACD4860FF4 Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=kernel.org Authentication-Results: mail.kernel.org; spf=pass smtp.mailfrom=linux-f2fs-devel-bounces@lists.sourceforge.net Received: from [127.0.0.1] (helo=sfs-ml-1.v29.lw.sourceforge.com) by sfs-ml-1.v29.lw.sourceforge.com with esmtp (Exim 4.90_1) (envelope-from ) id 1m5YAX-0005hG-KQ; Mon, 19 Jul 2021 18:39:09 +0000 Received: from [172.30.20.202] (helo=mx.sourceforge.net) by sfs-ml-1.v29.lw.sourceforge.com with esmtps (TLSv1.2:ECDHE-RSA-AES256-GCM-SHA384:256) (Exim 4.90_1) (envelope-from ) id 1m5YAV-0005gv-EC for linux-f2fs-devel@lists.sourceforge.net; Mon, 19 Jul 2021 18:39:07 +0000 DKIM-Signature: v=1; a=rsa-sha256; q=dns/txt; c=relaxed/relaxed; d=sourceforge.net; s=x; h=In-Reply-To:Content-Type:MIME-Version:References: Message-ID:Subject:Cc:To:From:Date:Sender:Reply-To:Content-Transfer-Encoding: Content-ID:Content-Description:Resent-Date:Resent-From:Resent-Sender: Resent-To:Resent-Cc:Resent-Message-ID:List-Id:List-Help:List-Unsubscribe: List-Subscribe:List-Post:List-Owner:List-Archive; bh=wb9JJpUH/uuc81h3kuoSoHHUZhb2DMDXAddqFch/OzA=; b=fwDfNohkmXz6hm9qouKB6CsB/J EEDLOtQgqh4vkE0rnLUy4muFj2TZaf9f2J943KnKlAfxU5l/V3xiEedNpnUfNdR9eMo848Y8aTj04 gE8CCJwn6Up+/OdKYRqlEXkhPucpDZvwzT/lWdnkSXk8IjKUFFszxrcDw1W1mwmUn60k=; DKIM-Signature: v=1; a=rsa-sha256; q=dns/txt; c=relaxed/relaxed; d=sf.net; s=x ; h=In-Reply-To:Content-Type:MIME-Version:References:Message-ID:Subject:Cc:To :From:Date:Sender:Reply-To:Content-Transfer-Encoding:Content-ID: Content-Description:Resent-Date:Resent-From:Resent-Sender:Resent-To:Resent-Cc :Resent-Message-ID:List-Id:List-Help:List-Unsubscribe:List-Subscribe: List-Post:List-Owner:List-Archive; bh=wb9JJpUH/uuc81h3kuoSoHHUZhb2DMDXAddqFch/OzA=; b=DImTI2n9PQVwz8IzrlgFvrILTB mvWOWWjBxm1FC+iJ0nYrAewubr4hQrotoShJpoADGwKk64qCeOgv1tTfrbittRxlQU/nFMttbWLrC KDUTmggEFvIr/lNqezmwWiP5eFYXz1nAAkm0wp9o6Ac4l5PI+qzOpYVsP7xdxZnv0BCU=; Received: from mail.kernel.org ([198.145.29.99]) by sfi-mx-2.v28.lw.sourceforge.com with esmtps (TLSv1.2:ECDHE-RSA-AES256-GCM-SHA384:256) (Exim 4.92.3) id 1m5YAJ-0001Bi-IW for linux-f2fs-devel@lists.sourceforge.net; Mon, 19 Jul 2021 18:39:07 +0000 Received: by mail.kernel.org (Postfix) with ESMTPSA id 0FD7461001; Mon, 19 Jul 2021 18:38:50 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=kernel.org; s=k20201202; t=1626719930; bh=XAFEEvD3nR4KfB4EUkhX7aCbFMc/mjRTf1uxFVpvuCY=; h=Date:From:To:Cc:Subject:References:In-Reply-To:From; b=qmcBF93MmkGg1pjgrYtgaC3K1iGGXdgNiriEIV2sZZP87nshMF7ZviuU8zFB50kk7 rKUuY7/BDxhWuCxjWwlsaIxSJrjLmbeUtMOy1PQiZ36mJOEjx3+oitP1yF5CnyoSUN nGXDqco7AJhzG6TV5L11cgJXYJfJhYVKi/W1LfCCa+zZcAIjGxRoaBsUxBEvEFcPT+ u1Npn7gyd658OpKeoSu76JSj32bzaKcfTu3lfWtduMicOa5p3/GR0gV+gZKof8okXX f1//7H/9Gxnu3+FB4Le8fn0MR6opAXci0mnKEksivTxfVh9Hn8P5qaUNV0sCtyB9Ry 5dtn9qPynBiaQ== Date: Mon, 19 Jul 2021 11:38:48 -0700 From: Jaegeuk Kim To: Chao Yu Message-ID: References: <648a96f7-2c83-e9ed-0cbd-4ee8e4797724@kernel.org> <55e069f7-662d-630c-1201-d0163b38bc17@kernel.org> <8f8d5645-9860-3e16-a09d-1a988ca6be72@kernel.org> MIME-Version: 1.0 Content-Disposition: inline In-Reply-To: X-Headers-End: 1m5YAJ-0001Bi-IW Subject: Re: [f2fs-dev] [PATCH v2 RFC] f2fs: fix to force keeping write barrier for strict fsync mode X-BeenThere: linux-f2fs-devel@lists.sourceforge.net X-Mailman-Version: 2.1.21 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: linux-kernel@vger.kernel.org, linux-f2fs-devel@lists.sourceforge.net Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Errors-To: linux-f2fs-devel-bounces@lists.sourceforge.net On 07/14, Chao Yu wrote: > On 2021/7/14 10:19, Jaegeuk Kim wrote: > > On 07/14, Chao Yu wrote: > > > On 2021/7/14 7:34, Jaegeuk Kim wrote: > > > > On 07/13, Chao Yu wrote: > > > > > On 2021/7/8 1:48, Jaegeuk Kim wrote: > > > > > > On 07/02, Chao Yu wrote: > > > > > > > On 2021/7/2 9:32, Jaegeuk Kim wrote: > > > > > > > > On 07/02, Chao Yu wrote: > > > > > > > > > On 2021/7/2 1:10, Jaegeuk Kim wrote: > > > > > > > > > > On 06/01, Chao Yu wrote: > > > > > > > > > > > [1] https://www.mail-archive.com/linux-f2fs-devel@lists.sourceforge.net/msg15126.html > > > > > > > > > > > > > > > > > > > > > > As [1] reported, if lower device doesn't support write barrier, in below > > > > > > > > > > > case: > > > > > > > > > > > > > > > > > > > > > > - write page #0; persist > > > > > > > > > > > - overwrite page #0 > > > > > > > > > > > - fsync > > > > > > > > > > > - write data page #0 OPU into device's cache > > > > > > > > > > > - write inode page into device's cache > > > > > > > > > > > - issue flush > > > > > > > > > > > > > > > > > > > > Well, we have preflush for node writes, so I don't think this is the case. > > > > > > > > > > > > > > > > > > > > fio.op_flags |= REQ_PREFLUSH | REQ_FUA; > > > > > > > > > > > > > > > > > > This is only used for atomic write case, right? > > > > > > > > > > > > > > > > > > I mean the common case which is called from f2fs_issue_flush() in > > > > > > > > > f2fs_do_sync_file(). > > > > > > > > > > > > > > > > How about adding PREFLUSH when writing node blocks aligned to the above set? > > > > > > > > > > > > > > You mean implementation like v1 as below? > > > > > > > > > > > > > > https://lore.kernel.org/linux-f2fs-devel/20200120100045.70210-1-yuchao0@huawei.com/ > > > > > > > > > > > > Yea, I think so. :P > > > > > > > > > > I prefer v2, we may have several schemes to improve performance with v2, e.g. > > > > > - use inplace IO to avoid newly added preflush > > > > > - use flush_merge option to avoid redundant preflush > > > > > - if lower device supports barrier IO, we can avoid newly added preflush > > > > > > > > Doesn't v2 give one more flush than v1? Why do you want to take worse one and > > > > > > FUA implies an extra preflush command or similar mechanism in lower device to keep data > > > in bio being persistent before this command's completion. > > > > > > Also if lower device doesn't support FUA natively, block layer turns it into an empty > > > PREFLUSH command. > > > > > > So, it's hard to say which one will win the benchmark game, maybe we need some > > > performance data before making the choice, but you know, it depends on device's > > > character. > > > > I was looking at # of bios. > > > > > > > > > try to improve back? Not clear the benefit on v2. > > > > > > Well, if user suffer and complain performance regression with v1, any plan to improve it? > > > > > > I just thought about plan B/C/D for no matter v1 or v2. > > > > I assumed you wanted v2 since it might be used for B/C/D improvements. But, it > > seems it wasn't. My point is to save one bio, but piggyback the flag to the > > device driver. > > I doubt the conclusion...but it needs to get some data to prove it. > > I think the right way is merging v1 now to fix the bug firstly, and let me do > the comparison on them a little bit later to see whether we need another > implementation... thoughts? Chao, could you please post v1 with an updated description? > > Thanks, > > > > > > > > > Thanks, > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > And please see do_checkpoint(), we call f2fs_flush_device_cache() and > > > > > > > > > commit_checkpoint() separately to keep persistence order of CP datas. > > > > > > > > > > > > > > > > > > See commit 46706d5917f4 ("f2fs: flush cp pack except cp pack 2 page at first") > > > > > > > > > for details. > > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > If SPO is triggered during flush command, inode page can be persisted > > > > > > > > > > > before data page #0, so that after recovery, inode page can be recovered > > > > > > > > > > > with new physical block address of data page #0, however there may > > > > > > > > > > > contains dummy data in new physical block address. > > > > > > > > > > > > > > > > > > > > > > Then what user will see is: after overwrite & fsync + SPO, old data in > > > > > > > > > > > file was corrupted, if any user do care about such case, we can suggest > > > > > > > > > > > user to use STRICT fsync mode, in this mode, we will force to trigger > > > > > > > > > > > preflush command to persist data in device cache in prior to node > > > > > > > > > > > writeback, it avoids potential data corruption during fsync(). > > > > > > > > > > > > > > > > > > > > > > Signed-off-by: Chao Yu > > > > > > > > > > > --- > > > > > > > > > > > v2: > > > > > > > > > > > - fix this by adding additional preflush command rather than using > > > > > > > > > > > atomic write flow. > > > > > > > > > > > fs/f2fs/file.c | 14 ++++++++++++++ > > > > > > > > > > > 1 file changed, 14 insertions(+) > > > > > > > > > > > > > > > > > > > > > > diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c > > > > > > > > > > > index 7d5311d54f63..238ca2a733ac 100644 > > > > > > > > > > > --- a/fs/f2fs/file.c > > > > > > > > > > > +++ b/fs/f2fs/file.c > > > > > > > > > > > @@ -301,6 +301,20 @@ static int f2fs_do_sync_file(struct file *file, loff_t start, loff_t end, > > > > > > > > > > > f2fs_exist_written_data(sbi, ino, UPDATE_INO)) > > > > > > > > > > > goto flush_out; > > > > > > > > > > > goto out; > > > > > > > > > > > + } else { > > > > > > > > > > > + /* > > > > > > > > > > > + * for OPU case, during fsync(), node can be persisted before > > > > > > > > > > > + * data when lower device doesn't support write barrier, result > > > > > > > > > > > + * in data corruption after SPO. > > > > > > > > > > > + * So for strict fsync mode, force to trigger preflush to keep > > > > > > > > > > > + * data/node write order to avoid potential data corruption. > > > > > > > > > > > + */ > > > > > > > > > > > + if (F2FS_OPTION(sbi).fsync_mode == FSYNC_MODE_STRICT && > > > > > > > > > > > + !atomic) { > > > > > > > > > > > + ret = f2fs_issue_flush(sbi, inode->i_ino); > > > > > > > > > > > + if (ret) > > > > > > > > > > > + goto out; > > > > > > > > > > > + } > > > > > > > > > > > } > > > > > > > > > > > go_write: > > > > > > > > > > > /* > > > > > > > > > > > -- > > > > > > > > > > > 2.29.2 _______________________________________________ Linux-f2fs-devel mailing list Linux-f2fs-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel