Re: [PATCH] erofs: support adjust lz4 history window size

[Gao Xiang <hsiangkao@redhat.com>]

On Tue, Feb 23, 2021 at 10:03:59AM +0800, Huang Jianan wrote:
> Hi Xiang,
> 
> On 2021/2/22 12:44, Gao Xiang wrote:
> > Hi Jianan,
> > 
> > On Thu, Feb 18, 2021 at 08:00:49PM +0800, Huang Jianan via Linux-erofs wrote:
> > > From: huangjianan <huangjianan@oppo.com>
> > > 
> > > lz4 uses LZ4_DISTANCE_MAX to record history preservation. When
> > > using rolling decompression, a block with a higher compression
> > > ratio will cause a larger memory allocation (up to 64k). It may
> > > cause a large resource burden in extreme cases on devices with
> > > small memory and a large number of concurrent IOs. So appropriately
> > > reducing this value can improve performance.
> > > 
> > > Decreasing this value will reduce the compression ratio (except
> > > when input_size <LZ4_DISTANCE_MAX). But considering that erofs
> > > currently only supports 4k output, reducing this value will not
> > > significantly reduce the compression benefits.
> > > 
> > > Signed-off-by: Huang Jianan <huangjianan@oppo.com>
> > > Signed-off-by: Guo Weichao <guoweichao@oppo.com>
> > > ---
> > >   fs/erofs/decompressor.c | 13 +++++++++----
> > >   fs/erofs/erofs_fs.h     |  3 ++-
> > >   fs/erofs/internal.h     |  3 +++
> > >   fs/erofs/super.c        |  3 +++
> > >   4 files changed, 17 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/fs/erofs/decompressor.c b/fs/erofs/decompressor.c
> > > index 1cb1ffd10569..94ae56b3ff71 100644
> > > --- a/fs/erofs/decompressor.c
> > > +++ b/fs/erofs/decompressor.c
> > > @@ -36,22 +36,27 @@ static int z_erofs_lz4_prepare_destpages(struct z_erofs_decompress_req *rq,
> > >   	struct page *availables[LZ4_MAX_DISTANCE_PAGES] = { NULL };
> > >   	unsigned long bounced[DIV_ROUND_UP(LZ4_MAX_DISTANCE_PAGES,
> > >   					   BITS_PER_LONG)] = { 0 };
> > > +	unsigned int lz4_distance_pages = LZ4_MAX_DISTANCE_PAGES;
> > >   	void *kaddr = NULL;
> > >   	unsigned int i, j, top;
> > > +	if (EROFS_SB(rq->sb)->compr_alg)
> > > +		lz4_distance_pages = DIV_ROUND_UP(EROFS_SB(rq->sb)->compr_alg,
> > > +						  PAGE_SIZE) + 1;
> > > +
> > Thanks for your patch, I agree that will reduce runtime memory
> > footpoint. and keep max sliding window ondisk in bytes (rather
> > than in blocks) is better., but could we calculate lz4_distance_pages
> > ahead when reading super_block?
> Thanks for suggestion, i will update it soon.
> > Also, in the next cycle, I'd like to introduce a bitmap for available
> > algorithms (maximum 16-bit) for the next LZMA algorithm, and for each
> > available algorithm introduces an on-disk variable-array like below:
> > bitmap(16-bit)    2       1       0
> >                  ...     LZMA    LZ4
> > __le16		compr_opt_off;      /* get the opt array start offset
> >                                         (I think also in 4-byte) */
> > 
> > compr alg 0 (lz4)	__le16	alg_opt_size;
> > 	/* next opt off = roundup(off + alg_opt_size, 4); */
> > 			__le16	lz4_max_distance;
> > 
> > /* 4-byte aligned */
> > compr alg x (if available)	u8	alg_opt_size;
> > 				...
> > 
> > ...
> > 
> > When reading sb, first, it scans the whole bitmap, and get all the
> > available algorithms in the image at once. And then read such compr
> > opts one-by-one.
> > 
> > Do you have some interest and extra time to implement it? :) That
> > makes me work less since I'm debugging mbpcluster compression now...
> 
> Sounds good, I will try to do this part of the work.

Yeah, but it seems to be part of the next LZMA algorithm patchset (with
a new brand new INCOMPET feature). I think we could introduce a __le16
lz4_max_distance field from sb reserved for now as a simple backporting
solution (since we only use < 64kb sliding window, so the image would
be forward compatibility with old kernels. 0 means 64kb sliding window,
otherwise it will < 64kb.)

And with the new INCOMPAT_COMPR_OPT feature, lz4_max_distance field
will be turned into compr_opt_off instead. And variable-array will be
used then.

So could you revise the patchset as above? Thanks!

Thanks,
Gao Xiang

> 
> Thanks,
> 
> Jianan
> 
> > Thanks,
> > Gao Xiang
> > 
>