Re: [PATCH] md/raid10: share pages between read and write bio's during recovery

[NeilBrown <neilb@suse.de>]

On Thu,  9 Jun 2011 03:04:33 +0900 Namhyung Kim <namhyung@gmail.com> wrote:

> When performing a recovery, only first 2 slots in r10_bio are in use,
> for read and write respectively. However all of pages in the write bio
> are never used and just replaced to read bio's when the read completes.
> 
> Get rid of those unused pages and share read pages properly.
> 
> Signed-off-by: Namhyung Kim <namhyung@gmail.com>
> ---
>  drivers/md/raid10.c |   13 ++++++++++---
>  1 files changed, 10 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
> index a53779ffdf89..621594981339 100644
> --- a/drivers/md/raid10.c
> +++ b/drivers/md/raid10.c
> @@ -116,6 +116,13 @@ static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
>  			goto out_free_bio;
>  		r10_bio->devs[j].bio = bio;
>  	}
> +
> +	/*
> +	 * We can share bv_page's during the recovery
> +	 */
> +	if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
> +		nalloc--;
> +
>  	/*
>  	 * Allocate RESYNC_PAGES data pages and attach them
>  	 * where needed.
> @@ -1363,16 +1370,16 @@ static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
>  	int i, d;
>  	struct bio *bio, *wbio;
>  
> -
> -	/* move the pages across to the second bio
> +	/*
> +	 * share the pages with the first bio
>  	 * and submit the write request
>  	 */
>  	bio = r10_bio->devs[0].bio;
>  	wbio = r10_bio->devs[1].bio;
>  	for (i=0; i < wbio->bi_vcnt; i++) {
>  		struct page *p = bio->bi_io_vec[i].bv_page;
> -		bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
>  		wbio->bi_io_vec[i].bv_page = p;
> +		get_page(p);
>  	}
>  	d = r10_bio->devs[1].devnum;
>  


Thanks.   Interesting idea, but I don't think this code is safe.

We end up calling bio_add_page on with an uninitialised 'page' (in
sync_request()).  This could result on BIOVEC_PHYS_MERGEABLE doing funny
things.

It would be OK to set up the two links to the one page in r10buf_pool_alloc,
so recovery_request_write doesn't need to do anything with pages.
That would probably be even more safe than the current code.

Thanks,
NeilBrown