Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <74ebf1cd-ce00-1aa2-8258-1bd336dc8470@squashfs.org.uk>
Date:   Tue, 16 May 2023 20:17:49 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.11.0
Subject: Re: [PATCH v3] squashfs: cache partial compressed blocks
To:     Vincent Whitchurch <vincent.whitchurch@axis.com>,
        Andrew Morton <akpm@linux-foundation.org>
Cc:     hch@lst.de, linux-kernel@vger.kernel.org,
        squashfs-devel@lists.sourceforge.net, kernel@axis.com
References: <20230510-squashfs-cache-v3-1-9f56ffd43f03@axis.com>
Content-Language: en-GB
From:   Phillip Lougher <phillip@squashfs.org.uk>
In-Reply-To: <20230510-squashfs-cache-v3-1-9f56ffd43f03@axis.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 16/05/2023 09:22, Vincent Whitchurch wrote:
> Before commit 93e72b3c612adcaca1 ("squashfs: migrate from ll_rw_block
> usage to BIO"), compressed blocks read by squashfs were cached in the
> page cache, but that is not the case after that commit.  That has lead
> to squashfs having to re-read a lot of sectors from disk/flash.
>
> For example, the first sectors of every metadata block need to be read
> twice from the disk.  Once partially to read the length, and a
> second time to read the block itself.  Also, in linear reads of large
> files, the last sectors of one data block are re-read from disk when
> reading the next data block, since the compressed blocks are of variable
> sizes and not aligned to device blocks.  This extra I/O results in a
> degrade in read performance of, for example, ~16% in one scenario on my
> ARM platform using squashfs with dm-verity and NAND.
>
> Since the decompressed data is cached in the page cache or squashfs'
> internal metadata and fragment caches, caching _all_ compressed pages
> would lead to a lot of double caching and is undesirable.  But make the
> code cache any disk blocks which were only partially requested, since
> these are the ones likely to include data which is needed by other file
> system blocks.  This restores read performance in my test scenario.
>
> The compressed block caching is only applied when the disk block size is
> equal to the page size, to avoid having to deal with caching sub-page
> reads.
>
> Signed-off-by: Vincent Whitchurch <vincent.whitchurch@axis.com>

Reviewed-by: Phillip Lougher <phillip@squashfs.org.uk>