RE: [f2fs-dev] [PATCH V2 2/2 RESEND] f2fs: read contiguous sit entrypages by merging for mount performance

[Jaegeuk Kim]

Hi,

2013-11-21 (ë), 11:18 +0800, Chao Yu:
> Hi,
> 
> > -----Original Message-----
> > From: Jaegeuk Kim [mailto:jaegeuk.kim@xxxxxxxxxxx]
> > Sent: Thursday, November 21, 2013 9:32 AM
> > To: Chao Yu
> > Cc: linux-fsdevel@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx; èå
> > Subject: Re: [f2fs-dev] [PATCH V2 2/2 RESEND] f2fs: read contiguous sit entry pages by merging for mount performance
> > 
> > Hi,
> > 
> > It seems that ra_sit_pages() is too tightly coupled with
> > build_sit_entries().
> 
> This code could be improved.
> 
> > Is there another way not to use *is_order?
> 
> Previously the code is like this:
>  -build_sit_entries()
>  next_setp:
> 	for(start = 0; start < TOTAL_SEGS(sbi); start++)
> 		/*step#1 readahead all sit entries blocks*/
> 			if(start % SIT_ENTRY_PER_BLOCK == 0) {
> 				blk_addr = current_sit_addr(sbi, start);
> 				/* grab and submit_read_page */
> 			}
> 			if(start == TOTAL_SEGS(sbi) - 1)
> 				f2fs_submit_read_bio();
> 			continue;
> 		/*step#2 fill sit entries info*/
> 		/*step#3 cover sit entries with journal*/
> 
> But I think its weakness is that it will cost lots of memory to read 
> ahead all sit entry pages when f2fs mount, and also it's serious waste 
> that we read them again after these pages are released by VM when 
> out of memory.
> 
> > 
> > The ra_sit_pages() tries to read consecutive sit pages as many as
> > possible.
> > So then, what about just checking whether its block address is
> > contiguous or not?
> > 
> > Something like this:
> >  -ra_sit_pages()
> > 	blkno = start;
> > 	while (blkno < sit_i->sit_blocks) {
> > 		blk_addr = current_sit_addr(sbi, blkno);
> > 		if (blkno != start && prev_blk_addr + 1 != blk_addr)
> > 			break;
> > 
> > 		/* grab and submit_read_page */
> > 
> > 		prev_blk_addr = blk_addr;
> > 		blkno++;
> > 	}
> 
> Agreed, this method could remove *order.
> Shouldn't we add nrpages for readahead policy as VM?

Aha, agreed.
We need nrpages to avoid too many reads on sit blocks.

But, still it needs to change the nrpages in its caller.
In your patch, it was sit_i->sit_blocks that is total # of sit blocks.
I think 128 or 256 is quite reasonable number.

Anyway, how about implementing ra_sit_pages() with a blk_plug likewise
ra_node_pages()?

-- 
Jaegeuk Kim
Samsung

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/