On Thu, May 24, 2012 at 03:52:02PM -0400, Vivek Goyal wrote:

On Wed, May 23, 2012 at 05:02:45PM -0700, Kent Overstreet wrote:

[..]

quoted

@@ -234,6 +234,13 @@ void bio_free(struct bio *bio, struct bio_set *bs)
 {
 	void *p;
 
+	if (!bs) {
+		if (bio_integrity(bio))
+			bio_integrity_free(bio, fs_bio_set);
+		kfree(bio);
+		return;
+	}
+

Ok, this seems to be the code which will take care of freeing kmalloced
bio. I think putting little comment about the explicit assumption is not
a bad idea.

Yeah, it's changing the semantics of bio_free(). I'll document that.

Somehow we need to integrate two patches so that we don't have memory leak
in bisection and reading code becomes easier.

I don't think there's any memory leak issues with this patch... there
are various annoyances with the dm code, though.

Also then what's the need of bio_reset() in previous patch. That seems to
be independent from getting rid of pkt_bio_destructor(). I would think
that keep we can split the patch and keep bio_reset() logic in a separate
patch. In fact I am not even sure that for one driver we should introduce
bio_reset() in generic block layer. So to me we should get rid of bio_reset()
and let all the gory details remain in driver.

Well, it would be possible to kill bi_destructor without introducing
bio_reset() - but that'd mean the kill bi_destructor patch would have to
muck around in the pktcdvd code. IMO introducing bio_reset() makes the
rest of the patch series much cleaner.