On Wed, Aug 08, 2012 at 11:05:17PM -0700, Tejun Heo wrote:

Hello,

On Wed, Aug 08, 2012 at 05:21:54PM -0700, Kent Overstreet wrote:

quoted

quoted

What's wrong with good ol' NULL?

If it's NULL, we can't distinguish between bios where that field wasn't
set (i.e. bios that were statically allocated somewhere) from bios that
were allocated by bio_kmalloc().

It's just there to make debugging easier - if bi_cnt goes to 0 on a bio
where it shouldn't we'll catch it at the BUG_ON() in bio_free() instead
of kfreeing a bad pointer.

I fail to see how that improves anything.  slab will complain clearly
if it gets passed in a pointer to static area.  The benefit is
imaginery.  If there's no bioset, it's NULL.  Let's please keep things
usual.

But if it's a pointer to heap allocated memory, but the bio was embedded
in another struct? I've seen a fair number of instances of that (md, off
the top of my head).

If you're sure that in a normal config the slab allocator is going to
complain right away and not corrupt itself, fine. But I've been bitten
way too hard by bugs that could've been caught right away by a simple
assert and instead I had to spend hours backtracking, and the block
layer is _rife_ with that kind of thing.