From patchwork Wed Feb 9 06:00:46 2022 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Jingbo Xu X-Patchwork-Id: 12739655 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id 8BDEAC433F5 for ; Wed, 9 Feb 2022 06:06:05 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S234740AbiBIGFM (ORCPT ); Wed, 9 Feb 2022 01:05:12 -0500 Received: from gmail-smtp-in.l.google.com ([23.128.96.19]:51864 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S234266AbiBIGBL (ORCPT ); Wed, 9 Feb 2022 01:01:11 -0500 Received: from out30-56.freemail.mail.aliyun.com (out30-56.freemail.mail.aliyun.com [115.124.30.56]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 9B97DC050CC0; Tue, 8 Feb 2022 22:01:13 -0800 (PST) X-Alimail-AntiSpam: AC=PASS;BC=-1|-1;BR=01201311R171e4;CH=green;DM=||false|;DS=||;FP=0|-1|-1|-1|0|-1|-1|-1;HT=e01e04357;MF=jefflexu@linux.alibaba.com;NM=1;PH=DS;RN=15;SR=0;TI=SMTPD_---0V3zaQQM_1644386468; Received: from localhost(mailfrom:jefflexu@linux.alibaba.com fp:SMTPD_---0V3zaQQM_1644386468) by smtp.aliyun-inc.com(127.0.0.1); Wed, 09 Feb 2022 14:01:09 +0800 From: Jeffle Xu To: dhowells@redhat.com, linux-cachefs@redhat.com, xiang@kernel.org, chao@kernel.org, linux-erofs@lists.ozlabs.org Cc: torvalds@linux-foundation.org, gregkh@linuxfoundation.org, willy@infradead.org, linux-fsdevel@vger.kernel.org, joseph.qi@linux.alibaba.com, bo.liu@linux.alibaba.com, tao.peng@linux.alibaba.com, gerry@linux.alibaba.com, eguan@linux.alibaba.com, linux-kernel@vger.kernel.org Subject: [PATCH v3 00/22] fscache,erofs: fscache-based demand-read semantics Date: Wed, 9 Feb 2022 14:00:46 +0800 Message-Id: <20220209060108.43051-1-jefflexu@linux.alibaba.com> X-Mailer: git-send-email 2.27.0 MIME-Version: 1.0 Precedence: bulk List-ID: X-Mailing-List: linux-fsdevel@vger.kernel.org changes since v2: - fscache,erofs: Now erofs uses fscache_read() directly instead of netfs library to read data from cache, to avoid the potential conflict with the following netfs library refactoring [1] (patch 12) (David Howells) - erofs: Implement fscache-based readahead. The current implementation is quite rough and is synchronous though. Need to be improved in the following iteration. - cachefiles_ondemand: use xarray instead of IDR managing pending read requests (patch 5) (Matthew Wilcox) - I also upload this patch set at: https://github.com/lostjeffle/linux/commits/jingbo/dev-erofs-fscache [1] https://lore.kernel.org/all/2946d871-b9e1-cf29-6d39-bcab30f2854f@linux.alibaba.com/t/#mfbb2053476760d8fac723c57dad529192a5084c6 RFC: https://lore.kernel.org/all/YbRL2glGzjfZkVbH@B-P7TQMD6M-0146.local/t/ v1: https://lore.kernel.org/lkml/47831875-4bdd-8398-9f2d-0466b31a4382@linux.alibaba.com/T/ v2: https://lore.kernel.org/all/2946d871-b9e1-cf29-6d39-bcab30f2854f@linux.alibaba.com/t/ [Background] ============ Nydus is a remote container snapthotter specially optimised for container images distribution over network. It has recently been accepted as a sub-project of containerd[1]. Nydus is an excellent container image acceleration solution, since it only pulls data from remote when it's really needed, a.k.a. on-demand reading. erofs (Enhanced Read-Only File System) is a filesystem specially optimised for read-only scenarios. (Documentation/filesystem/erofs.rst) Recently we are focusing on erofs in container images distribution scenario [2], trying to combine it with nydus. In this case, erofs can be mounted from one bootstrap file (metadata) with (optional) multiple data blob files (data) stored on another local filesystem. (All these files are actually image files in erofs disk format.) To accelerate the container startup (fetching container image from remote and then start the container), we do hope that the bootstrap blob file could support demand read. That is, erofs can be mounted and accessed even when the bootstrap/data blob files have not been fully downloaded. That means we have to manage the cache state of the bootstrap/data blob files (if cache hit, read directly from the local cache; if cache miss, fetch the data somehow). It would be painful and may be dumb for erofs to implement the cache management itself. Thus we prefer fscache/cachefiles to do the cache management. Besides, the demand-read feature shall be general and it can benefit other using scenarios if it can be implemented in fscache level. [1] https://d7y.io/en-us/blog/containerd_accepted_nydus-snapshotter.html [2] https://sched.co/pcdL [Overall Design] ================ The upper fs uses a backing file on the local fs as the local cache (exactly the "cachefiles" way), and relies on fscache to detect if data is ready or not (cache hit/miss). Since currently fscache detects cache hit/miss by detecting the hole of the backing files, our demand-read mechanism also relies on the hole detecting. 1. initial phase On the first beginning, the user daemon will touch the backing files (bootstrap/data blob files) under corresponding directory (under /cache///) in advance. These backing files are completely sparse files (with zero disk usage). Since these backing files are all read-only and the file size is known prior mounting, user daemon will set corresponding file size and thus create all these sparse backing files in advance. 2. cache miss When a file range (of bootstrap/data blob file) is accessed for the first time, a cache miss will be triggered and then .issue_op() will be called to fetch the data somehow. In the demand-read case, we relies on a user daemon to fetch the data from local/remote. In this case, .issue_op() just packages the file range into a message and informs the user daemon. User daemon needs to poll and wait on the devnode (/dev/cachefiles_demand). Once awaken, the user daemon will read the devnode to get the file range information, and then fetch the data corresponding to the file range somehow, e.g. download from remote through network. Once data ready, the user daemon will write the fetched data into the backing file and then inform cachefiles backend by writing to the devnode. Cachefiles backend getting blocked on the previous .issue_op() calling will be awaken then. By then the data has been ready in the backing file, and the upper fs will reinitiate a read request from the backing file. 3. cache hit Once data is already ready in the backing file, upper fs will read from the backing file directly. [Advantage of fscache-based demand-read] ======================================== 1. Asynchronous Prefetch In current mechanism, fscache is responsible for cache state management, while the data plane (fetch data from local/remote on cache miss) is done on the user daemon side. If data has already been ready in the backing file, the upper fs (e.g. erofs) will read from the backing file directly and won't be trapped to user space anymore. Thus the user daemon could fetch data (from remote) asynchronously on the background, and thus accelerate the backing file accessing in some degree. 2. Support massive blob files Besides this mechanism supports a large amount of backing files, and thus can benefit the densely employed scenario. In our using scenario, one container image can correspond to one bootstrap file (required) and multiple data blob files (optional). For example, one container image for node.js will corresponds to ~20 files in total. In densely employed environment, there could be as many as hundreds of containers and thus thousands of backing files on one machine. [Test] ========== You could start a quick test by https://github.com/lostjeffle/demand-read-cachefilesd Jeffle Xu (22): fscache: export fscache_end_operation() fscache: add a method to support on-demand read semantics cachefiles: extract generic function for daemon methods cachefiles: detect backing file size in on-demand read mode cachefiles: introduce new devnode for on-demand read mode erofs: use meta buffers for erofs_read_superblock() erofs: export erofs_map_blocks() erofs: add mode checking helper erofs: register global fscache volume erofs: add cookie context helper functions erofs: add anonymous inode managing page cache of blob file erofs: add erofs_fscache_read_page() helper erofs: register cookie context for bootstrap blob erofs: implement fscache-based metadata read erofs: implement fscache-based data read for non-inline layout erofs: implement fscache-based data read for inline layout erofs: register cookie context for data blobs erofs: implement fscache-based data read for data blobs erofs: implement fscache-based data readahead for hole erofs: implement fscache-based data readahead for non-inline layout erofs: implement fscache-based data readahead for inline layout erofs: add 'uuid' mount option Documentation/filesystems/netfs_library.rst | 18 + fs/cachefiles/Kconfig | 13 + fs/cachefiles/daemon.c | 243 +++++++++-- fs/cachefiles/internal.h | 12 + fs/cachefiles/io.c | 60 +++ fs/cachefiles/main.c | 27 ++ fs/cachefiles/namei.c | 60 ++- fs/erofs/Makefile | 3 +- fs/erofs/data.c | 18 +- fs/erofs/fscache.c | 451 ++++++++++++++++++++ fs/erofs/inode.c | 6 +- fs/erofs/internal.h | 30 ++ fs/erofs/super.c | 106 ++++- fs/fscache/internal.h | 11 - fs/nfs/fscache.c | 8 - include/linux/fscache.h | 39 ++ include/linux/netfs.h | 4 + include/uapi/linux/cachefiles_ondemand.h | 14 + 18 files changed, 1050 insertions(+), 73 deletions(-) create mode 100644 fs/erofs/fscache.c create mode 100644 include/uapi/linux/cachefiles_ondemand.h