Re: [PATCH v3] block: make sure big bio is splitted into at most 256 bvecs

On Mon, Aug 15, 2016 at 11:23:28AM -0700, Christoph Hellwig wrote:

On Mon, Aug 15, 2016 at 11:11:22PM +0800, Ming Lei wrote:

quoted

After arbitrary bio size is supported, the incoming bio may
be very big. We have to split the bio into small bios so that
each holds at most BIO_MAX_PAGES bvecs for safety reason, such
as bio_clone().

I still think working around a rough driver submitting too large
I/O is a bad thing until we've done a full audit of all consuming
bios through ->make_request, and we've enabled it for the common
path as well.

bcache originally had workaround code to split too-large bios when it first went
upstream - that was dropped only after the patches to make
generic_make_request() handle arbitrary size bios went in. So to do what you're
suggesting would mean reverting that bcache patch and bringing that code back,
which from my perspective would be a step in the wrong direction. I just want to
get this over and done with.

re: interactions with other drivers - bio_clone() has already been changed to
only clone biovecs that are live for current bi_iter, so there shouldn't be any
safety issues. A driver would have to be intentionally doing its own open coded
bio cloning that clones all of bi_io_vec, not just the active ones - but if
they're doing that, they're already broken because a driver isn't allowed to
look at bi_vcnt if it isn't a bio that it owns - bi_vcnt is 0 on bios that don't
own their biovec (i.e. that were created by bio_clone_fast).

And the cloning and bi_vcnt usage stuff I audited very thoroughly back when I
was working on immutable biovecs and such back in the day, and I had to do a
fair amount of cleanup/refactoring before that stuff could go in.

quoted

 	bool do_split = true;
 	struct bio *new = NULL;
 	const unsigned max_sectors = get_max_io_size(q, bio);
+	unsigned bvecs = 0;
+
+	*no_merge = true;
 
 	bio_for_each_segment(bv, bio, iter) {
 		/*
+		 * With arbitrary bio size, the incoming bio may be very
+		 * big. We have to split the bio into small bios so that
+		 * each holds at most BIO_MAX_PAGES bvecs because
+		 * bio_clone() can fail to allocate big bvecs.
+		 *
+		 * It should have been better to apply the limit per
+		 * request queue in which bio_clone() is involved,
+		 * instead of globally. The biggest blocker is
+		 * bio_clone() in bio bounce.
+		 *
+		 * If bio is splitted by this reason, we should allow
+		 * to continue bios merging.
+		 *
+		 * TODO: deal with bio bounce's bio_clone() gracefully
+		 * and convert the global limit into per-queue limit.
+		 */
+		if (bvecs++ >= BIO_MAX_PAGES) {
+			*no_merge = false;
+			goto split;
+		}

That being said this simple if check here is simple enough that it's
probably fine.  But I see no need to uglify the whole code path
with that no_merge flag.  Please drop if for now, and if we start
caring for this path in common code we should just move the
REQ_NOMERGE setting into the actual blk_bio_*_split helpers.

Agreed about the no_merge thing.