[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <89e27e25-ec2b-337e-4bf0-705912eea62d@squashfs.org.uk>
Date: Tue, 16 Nov 2021 20:50:32 +0000
From: Phillip Lougher <phillip@...ashfs.org.uk>
To: Zheng Liang <zhengliang6@...wei.com>, linux-kernel@...r.kernel.org,
Andrew Morton <akpm@...ux-foundation.org>
Cc: yi.zhang@...wei.com, houtao1@...wei.com, miaoxie@...wei.com
Subject: Re: [PATCH] squashfs: provides backing_dev_info in order to disable
read-ahead
On 16/11/2021 11:31, Zheng Liang wrote:
> the commit c1f6925e1091("mm: put readahead pages in cache earlier")
> causes the read performance of squashfs to deteriorate.Through testing,
> we find that the performance will be back by closing the readahead of
> squashfs. So we want to learn the way of ubifs, provides backing_dev_info
> and disable read-ahead.
>
> --------------------------------------------------------------------
> We tested the following data by fio.
> squashfs image blocksize=128K
> test command:
> fio --name basic --bs=? --filename="/mnt/test_file" --rw=? --iodepth=1 --ioengine=psync --runtime=200 --time_based
>
> turn on squashfs readahead in 5.10 kernel
> bs(k) read/randread MB/s
> 4 randread 271
> 128 randread 231
> 1024 randread 246
> 4 read 310
> 128 read 245
> 1024 read 247
>
> turn off squashfs readahead in 5.10 kernel
> bs(k) read/randread MB/s
> 4 randread 293
> 128 randread 330
> 1024 randread 363
> 4 read 338
> 128 read 360
> 1024 read 365
>
> turn on squashfs readahead and revert the
> commit c1f6925e1091("mm: put readahead
> pages in cache earlier") in 5.10 kernel
> bs(k) read/randread MB/s
> 4 randread 289
> 128 randread 306
> 1024 randread 335
> 4 read 337
> 128 read 336
> 1024 read 338
>
> Signed-off-by: Zheng Liang <zhengliang6@...wei.com>
Signed-off-by: Phillip Lougher <phillip@...ashfs.org.uk>
Looks good to me.
Phillip
--
Squashfs maintainer.
> ---
> fs/squashfs/super.c | 33 +++++++++++++++++++++++++++++++++
> 1 file changed, 33 insertions(+)
>
> diff --git a/fs/squashfs/super.c b/fs/squashfs/super.c
> index bb44ff4c5cc6..7a4c865fc8e6 100644
> --- a/fs/squashfs/super.c
> +++ b/fs/squashfs/super.c
> @@ -29,6 +29,7 @@
> #include <linux/module.h>
> #include <linux/magic.h>
> #include <linux/xattr.h>
> +#include <linux/backing-dev.h>
>
> #include "squashfs_fs.h"
> #include "squashfs_fs_sb.h"
> @@ -112,6 +113,24 @@ static const struct squashfs_decompressor *supported_squashfs_filesystem(
> return decompressor;
> }
>
> +static int squashfs_bdi_init(struct super_block *sb)
> +{
> + int err;
> + unsigned int major = MAJOR(sb->s_dev);
> + unsigned int minor = MINOR(sb->s_dev);
> +
> + bdi_put(sb->s_bdi);
> + sb->s_bdi = &noop_backing_dev_info;
> +
> + err = super_setup_bdi_name(sb, "squashfs_%u_%u", major, minor);
> + if (err)
> + return err;
> +
> + sb->s_bdi->ra_pages = 0;
> + sb->s_bdi->io_pages = 0;
> +
> + return 0;
> +}
>
> static int squashfs_fill_super(struct super_block *sb, struct fs_context *fc)
> {
> @@ -127,6 +146,20 @@ static int squashfs_fill_super(struct super_block *sb, struct fs_context *fc)
>
> TRACE("Entered squashfs_fill_superblock\n");
>
> + /*
> + * squashfs provides 'backing_dev_info' in order to disable read-ahead. For
> + * squashfs, I/O is not deferred, it is done immediately in readpage,
> + * which means the user would always have to wait their own I/O. So the effect
> + * of readahead is very weak for squashfs. squashfs_bdi_init will set
> + * sb->s_bdi->ra_pages and sb->s_bdi->io_pages to 0 and close readahead for
> + * squashfs.
> + */
> + err = squashfs_bdi_init(sb);
> + if (err) {
> + errorf(fc, "squashfs init bdi failed");
> + return err;
> + }
> +
> sb->s_fs_info = kzalloc(sizeof(*msblk), GFP_KERNEL);
> if (sb->s_fs_info == NULL) {
> ERROR("Failed to allocate squashfs_sb_info\n");
>
Powered by blists - more mailing lists