[PATCH v2 3/3] erofs: introduce readmore decompression strategy

[Gao Xiang]

On Sun, Oct 17, 2021 at 11:34:22PM +0800, Chao Yu wrote:
> On 2021/10/9 4:08, Gao Xiang wrote:
> > From: Gao Xiang <hsiangkao at linux.alibaba.com>
> > 
> > Previously, the readahead window was strictly followed by EROFS
> > decompression strategy in order to minimize extra memory footprint.
> > However, it could become inefficient if just reading the partial
> > requested data for much big LZ4 pclusters and the upcoming LZMA
> > implementation.
> > 
> > Let's try to request the leading data in a pcluster without
> > triggering memory reclaiming instead for the LZ4 approach first
> > to boost up 100% randread of large big pclusters, and it has no real
> > impact on low memory scenarios.
> > 
> > It also introduces a way to expand read lengths in order to decompress
> > the whole pcluster, which is useful for LZMA since the algorithm
> > itself is relatively slow and causes CPU bound, but LZ4 is not.
> > 
> > Signed-off-by: Gao Xiang <hsiangkao at linux.alibaba.com>
> > ---
> >   fs/erofs/internal.h | 13 ++++++
> >   fs/erofs/zdata.c    | 99 ++++++++++++++++++++++++++++++++++++---------
> >   2 files changed, 93 insertions(+), 19 deletions(-)
> > 
> > diff --git a/fs/erofs/internal.h b/fs/erofs/internal.h
> > index 48bfc6eb2b02..7f96265ccbdb 100644
> > --- a/fs/erofs/internal.h
> > +++ b/fs/erofs/internal.h
> > @@ -307,6 +307,19 @@ static inline unsigned int erofs_inode_datalayout(unsigned int value)
> >   			      EROFS_I_DATALAYOUT_BITS);
> >   }
> > +/*
> > + * Different from grab_cache_page_nowait(), reclaiming is never triggered
> > + * when allocating new pages.
> > + */
> > +static inline
> > +struct page *erofs_grab_cache_page_nowait(struct address_space *mapping,
> > +					  pgoff_t index)
> > +{
> > +	return pagecache_get_page(mapping, index,
> > +			FGP_LOCK|FGP_CREAT|FGP_NOFS|FGP_NOWAIT,
> > +			readahead_gfp_mask(mapping) & ~__GFP_RECLAIM);
> > +}
> > +
> >   extern const struct super_operations erofs_sops;
> >   extern const struct address_space_operations erofs_raw_access_aops;
> > diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
> > index 5c34ef66677f..febb018e10a7 100644
> > --- a/fs/erofs/zdata.c
> > +++ b/fs/erofs/zdata.c
> > @@ -1377,6 +1377,72 @@ static void z_erofs_runqueue(struct super_block *sb,
> >   	z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
> >   }
> > +/*
> > + * Since partial uptodate is still unimplemented for now, we have to use
> > + * approximate readmore strategies as a start.
> > + */
> > +static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
> > +				      struct readahead_control *rac,
> > +				      erofs_off_t end,
> > +				      struct list_head *pagepool,
> > +				      bool backmost)
> > +{
> > +	struct inode *inode = f->inode;
> > +	struct erofs_map_blocks *map = &f->map;
> > +	erofs_off_t cur;
> > +	int err;
> > +
> > +	if (backmost) {
> > +		map->m_la = end;
> > +		/* TODO: pass in EROFS_GET_BLOCKS_READMORE for LZMA later */
> > +		err = z_erofs_map_blocks_iter(inode, map, 0);
> > +		if (err)
> > +			return;
> > +
> > +		/* expend ra for the trailing edge if readahead */
> > +		if (rac) {
> > +			loff_t newstart = readahead_pos(rac);
> > +
> > +			cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
> > +			readahead_expand(rac, newstart, cur - newstart);
> > +			return;
> > +		}
> > +		end = round_up(end, PAGE_SIZE);
> > +	} else {
> > +		end = round_up(map->m_la, PAGE_SIZE);
> > +
> > +		if (!map->m_llen)
> > +			return;
> > +	}
> > +
> > +	cur = map->m_la + map->m_llen - 1;
> > +	while (cur >= end) {
> > +		pgoff_t index = cur >> PAGE_SHIFT;
> > +		struct page *page;
> > +
> > +		page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
> > +		if (!page)
> > +			goto skip;
> > +
> > +		if (PageUptodate(page)) {
> > +			unlock_page(page);
> > +			put_page(page);
> > +			goto skip;
> > +		}
> > +
> > +		err = z_erofs_do_read_page(f, page, pagepool);
> > +		if (err)
> > +			erofs_err(inode->i_sb,
> > +				  "readmore error at page %lu @ nid %llu",
> > +				  index, EROFS_I(inode)->nid);
> > +		put_page(page);
> > +skip:
> > +		if (cur < PAGE_SIZE)
> > +			break;
> > +		cur = (index << PAGE_SHIFT) - 1;
> 
> Looks a little bit weird to readahead backward, any special reason here?

Due to the do_read_page implementation, since I'd like to avoid
to get the exact full extent length (FIEMAP-likewise) inside
do_read_page but only request the needed range, so it should be
all in a backward way. Also the submission chain can be then in
a forward way.

If the question was asked why we should read backward, as I said in the
commit message, big pclusters matter since we could read in more leading
data at once.

Thanks,
Gao Xiang

> 
> Thanks,