Re: [PATCH] block: Explicitly handle discard/write same segments

[Hugh Dickins <hughd@google.com>]

On Tue, 4 Feb 2014, Kent Overstreet wrote:

> Immutable biovecs changed the way biovecs are interpreted - drivers no
> longer use bi_vcnt, they have to go by bi_iter.bi_size (to allow for
> using part of an existing segment without modifying it).
> 
> This breaks with discards and write_same bios, since for those bi_size
> has nothing to do with segments in the biovec. So for now, we need a
> fairly gross hack - we fortunately know that there will never be more
> than one segment for the entire request, so we can special case
> discard/write_same.
> 
> Signed-off-by: Kent Overstreet <kmo@daterainc.com>
> ---
> On Fri, Jan 31, 2014 at 09:17:25AM -0800, Hugh Dickins wrote:
> > On Thu, 16 Jan 2014, Kent Overstreet wrote:
> > > 
> > > Ok, I reread the code and figured it out - the analagous change also has to be
> > > made in __blk_segment_map_sg(). I'll mail out a patch for this tomorrow after
> > > I've stared at the code more and had less beer.
> > 
> > I'd been hoping for a patch to try, but now your changes have hit Linus's
> > tree: so today we have discard broken there too, crashing as originally
> > reported on the NULL struct page pointer in __blk_recalc_rq_segments()'s
> > page_to_pfn(bv.bv_page).
> > 
> > How to reproduce it?  I hope you'll find easier ways, but I get it with
> > swapping to SSD (remember "swapon -d" to enable discard).  I'm just doing
> > what I've done for years, running a pair of make -j20 kbuilds to tmpfs in
> > limited RAM (I use mem=700M with 1.5G of swap: but that would be far too
> > little RAM for a general config of current tree), to get plenty of fairly
> > chaotic swapping but good forward progress nonetheless (if the sizes are
> > too small, then it'll just thrash abysmally or be OOM-killed).
> > 
> > But please do send me a patch and I'll give it a try - thanks.
> 
> Hugh - can you give this patch a try? Passes my tests but I was never
> able to reproduce your crash, unfortunately.

Thanks a lot, Kent: I'm glad to report, this is working fine for me.

Hugh

> 
>  block/blk-merge.c | 91 +++++++++++++++++++++++++++++++++++++------------------
>  1 file changed, 62 insertions(+), 29 deletions(-)
> 
> diff --git a/block/blk-merge.c b/block/blk-merge.c
> index 8f8adaa954..6c583f9c5b 100644
> --- a/block/blk-merge.c
> +++ b/block/blk-merge.c
> @@ -21,6 +21,16 @@ static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
>  	if (!bio)
>  		return 0;
>  
> +	/*
> +	 * This should probably be returning 0, but blk_add_request_payload()
> +	 * (Christoph!!!!)
> +	 */
> +	if (bio->bi_rw & REQ_DISCARD)
> +		return 1;
> +
> +	if (bio->bi_rw & REQ_WRITE_SAME)
> +		return 1;
> +
>  	fbio = bio;
>  	cluster = blk_queue_cluster(q);
>  	seg_size = 0;
> @@ -161,30 +171,60 @@ new_segment:
>  	*bvprv = *bvec;
>  }
>  
> -/*
> - * map a request to scatterlist, return number of sg entries setup. Caller
> - * must make sure sg can hold rq->nr_phys_segments entries
> - */
> -int blk_rq_map_sg(struct request_queue *q, struct request *rq,
> -		  struct scatterlist *sglist)
> +static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
> +			     struct scatterlist *sglist,
> +			     struct scatterlist **sg)
>  {
>  	struct bio_vec bvec, bvprv = { NULL };
> -	struct req_iterator iter;
> -	struct scatterlist *sg;
> +	struct bvec_iter iter;
>  	int nsegs, cluster;
>  
>  	nsegs = 0;
>  	cluster = blk_queue_cluster(q);
>  
> -	/*
> -	 * for each bio in rq
> -	 */
> -	sg = NULL;
> -	rq_for_each_segment(bvec, rq, iter) {
> -		__blk_segment_map_sg(q, &bvec, sglist, &bvprv, &sg,
> -				     &nsegs, &cluster);
> -	} /* segments in rq */
> +	if (bio->bi_rw & REQ_DISCARD) {
> +		/*
> +		 * This is a hack - drivers should be neither modifying the
> +		 * biovec, nor relying on bi_vcnt - but because of
> +		 * blk_add_request_payload(), a discard bio may or may not have
> +		 * a payload we need to set up here (thank you Christoph) and
> +		 * bi_vcnt is really the only way of telling if we need to.
> +		 */
> +
> +		if (bio->bi_vcnt)
> +			goto single_segment;
> +
> +		return 0;
> +	}
> +
> +	if (bio->bi_rw & REQ_WRITE_SAME) {
> +single_segment:
> +		*sg = sglist;
> +		bvec = bio_iovec(bio);
> +		sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
> +		return 1;
> +	}
> +
> +	for_each_bio(bio)
> +		bio_for_each_segment(bvec, bio, iter)
> +			__blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
> +					     &nsegs, &cluster);
>  
> +	return nsegs;
> +}
> +
> +/*
> + * map a request to scatterlist, return number of sg entries setup. Caller
> + * must make sure sg can hold rq->nr_phys_segments entries
> + */
> +int blk_rq_map_sg(struct request_queue *q, struct request *rq,
> +		  struct scatterlist *sglist)
> +{
> +	struct scatterlist *sg = NULL;
> +	int nsegs = 0;
> +
> +	if (rq->bio)
> +		nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
>  
>  	if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
>  	    (blk_rq_bytes(rq) & q->dma_pad_mask)) {
> @@ -230,20 +270,13 @@ EXPORT_SYMBOL(blk_rq_map_sg);
>  int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
>  		   struct scatterlist *sglist)
>  {
> -	struct bio_vec bvec, bvprv = { NULL };
> -	struct scatterlist *sg;
> -	int nsegs, cluster;
> -	struct bvec_iter iter;
> -
> -	nsegs = 0;
> -	cluster = blk_queue_cluster(q);
> -
> -	sg = NULL;
> -	bio_for_each_segment(bvec, bio, iter) {
> -		__blk_segment_map_sg(q, &bvec, sglist, &bvprv, &sg,
> -				     &nsegs, &cluster);
> -	} /* segments in bio */
> +	struct scatterlist *sg = NULL;
> +	int nsegs;
> +	struct bio *next = bio->bi_next;
> +	bio->bi_next = NULL;
>  
> +	nsegs = __blk_bios_map_sg(q, bio, sglist, &sg);
> +	bio->bi_next = next;
>  	if (sg)
>  		sg_mark_end(sg);
>  
> -- 
> 1.9.rc1