Re: [PATCH 2/2] erofs: add sysfs node to control sync decompression strategy

[Huang Jianan]

Hi Xiang,

在 2021/11/9 11:16, Gao Xiang 写道:
> On Tue, Nov 09, 2021 at 10:54:45AM +0800, Huang Jianan via Linux-erofs wrote:
> > Although readpage is a synchronous path, there will be no additional
> > kworker scheduling overhead in non-atomic contexts. So we can add a
> > sysfs node to allow disable sync decompression.
> >
> > Signed-off-by: Huang Jianan <huangjianan@xxxxxxxx>
> > ---
> >   fs/erofs/internal.h | 2 +-
> >   fs/erofs/super.c    | 2 +-
> >   fs/erofs/sysfs.c    | 6 ++++++
> >   fs/erofs/zdata.c    | 9 ++++-----
> >   4 files changed, 12 insertions(+), 7 deletions(-)
> >
> > diff --git a/fs/erofs/internal.h b/fs/erofs/internal.h
> > index d0cd712dc222..1ab96878009d 100644
> > --- a/fs/erofs/internal.h
> > +++ b/fs/erofs/internal.h
> > @@ -60,7 +60,7 @@ struct erofs_mount_opts {
> >   #ifdef CONFIG_EROFS_FS_ZIP
> >   	/* current strategy of how to use managed cache */
> >   	unsigned char cache_strategy;
> > -	/* strategy of sync decompression (false - auto, true - force on) */
> > +	/* strategy of sync decompression (false - disable, true - force on) */
> Please leave false - auto.
Agree.
> >   	bool readahead_sync_decompress;
> >   
> >   	/* threshold for decompression synchronously */
> > diff --git a/fs/erofs/super.c b/fs/erofs/super.c
> > index abc1da5d1719..26585d865503 100644
> > --- a/fs/erofs/super.c
> > +++ b/fs/erofs/super.c
> > @@ -423,7 +423,7 @@ static void erofs_default_options(struct erofs_fs_context *ctx)
> >   #ifdef CONFIG_EROFS_FS_ZIP
> >   	ctx->opt.cache_strategy = EROFS_ZIP_CACHE_READAROUND;
> >   	ctx->opt.max_sync_decompress_pages = 3;
> > -	ctx->opt.readahead_sync_decompress = false;
> > +	ctx->opt.readahead_sync_decompress = true;
> Please leave readahead_sync_decompress = false 'auto' by default.
>
> I don't like .readpage() applies async decompression since it's
> actually a sync way, otherwise, it will cause more scheduling
> overhead, see:
> https://lore.kernel.org/r/20201016160443.18685-1-willy@xxxxxxxxxxxxx
> https://lore.kernel.org/r/20201102184312.25926-16-willy@xxxxxxxxxxxxx

Maybe consider adding a disbale strategy? And then we will have:
0 - auto, 1 - force on, 2 - disable
I will try this in the next version.

for .readpage(), If the decompression can be done in a non-atomic context,
there will be no additional scheduling overhead? In this case, we can 
turn off
the syncdecompression.

Thanks,
Jianan

> Thanks,
> Gao Xiang