block: avoid scatterlist offsets > PAGE_SIZE

While we generally allow scatterlists to have offsets larger than page
size for an entry, and other subsystems like the crypto code make use of
that, the block layer isn't quite ready for that.  Flip the switch back
to avoid them for now, and revisit that decision early in a merge window
once the known offenders are fixed.

Fixes: 8a96a0e4 ("block: rewrite blk_bvec_map_sg to avoid a nth_page call")
Reviewed-by: NMing Lei <ming.lei@redhat.com>
Tested-by: NGuenter Roeck <linux@roeck-us.net>
Reported-by: NGuenter Roeck <linux@roeck-us.net>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

block/blk-merge.c

@@ -474,9 +474,21 @@ static unsigned blk_bvec_map_sg(struct request_queue *q,
 	while (nbytes > 0) {
 		unsigned offset = bvec->bv_offset + total;
 		unsigned len = min(get_max_segment_size(q, offset), nbytes);
+		struct page *page = bvec->bv_page;
+
+		/*
+		 * Unfortunately a fair number of drivers barf on scatterlists
+		 * that have an offset larger than PAGE_SIZE, despite other
+		 * subsystems dealing with that invariant just fine.  For now
+		 * stick to the legacy format where we never present those from
+		 * the block layer, but the code below should be removed once
+		 * these offenders (mostly MMC/SD drivers) are fixed.
+		 */
+		page += (offset >> PAGE_SHIFT);
+		offset &= ~PAGE_MASK;
 
 		*sg = blk_next_sg(sg, sglist);
-		sg_set_page(*sg, bvec->bv_page, len, offset);
+		sg_set_page(*sg, page, len, offset);
 
 		total += len;
 		nbytes -= len;