Re: [PATCH 2/2] block: Avoid deadlocks with bio allocation bystacking drivers

From: Kent Overstreet
Date: Mon Sep 10 2012 - 16:24:33 EST


On Mon, Sep 10, 2012 at 10:22:10AM -0700, Tejun Heo wrote:
> Hello, Kent.
>
> On Sun, Sep 09, 2012 at 05:28:10PM -0700, Kent Overstreet wrote:
> > > > + while ((bio = bio_list_pop(current->bio_list)))
> > > > + bio_list_add(bio->bi_pool == bs ? &punt : &nopunt, bio);
> > > > +
> > > > + *current->bio_list = nopunt;
> > >
> > > Why this is necessary needs explanation and it's done in rather
> > > unusual way. I suppose the weirdness is from bio_list API
> > > restriction?
> >
> > It's because bio_lists are singly linked, so deleting an entry from the
> > middle of the list would be a real pain - just much cleaner/simpler to
> > do it this way.
>
> Yeah, I wonder how benefical that singly linked list is. Eh well...

Well, this is the first time I can think of that it's come up, and IMO
this is no less clean a way of writing it... just a bit unusual in C,
it feels more functional to me instead of imperative.

> > > Wouldn't the following be better?
> > >
> > > p = mempool_alloc(bs->bi_pool, gfp_mask);
> > > if (unlikely(!p) && gfp_mask != saved_gfp) {
> > > punt_bios_to_rescuer(bs);
> > > p = mempool_alloc(bs->bi_pool, saved_gfp);
> > > }
> >
> > That'd require duplicating the error handling in two different places -
> > once for the initial allocation, once for the bvec allocation. And I
> > really hate that writing code that does
> >
> > alloc_something()
> > if (fail) {
> > alloc_something_again()
> > }
> >
> > it just screams ugly to me.
>
> I don't know. That at least represents what's going on and goto'ing
> back and forth is hardly pretty. Sometimes the code gets much uglier
> / unwieldy and we have to live with gotos. Here, that doesn't seem to
> be the case.

I think this is really more personal preference than anything, but:

Setting gfp_mask = saved_gfp after calling punt_bio_to_rescuer() is
really the correct thing to do, and makes the code clearer IMO: once
we've run punt_bio_to_rescuer() we don't need to mask out GFP_WAIT (not
until the next time a bio is submitted, really).

This matters a bit for the bvl allocation too, if we call
punt_bio_to_rescuer() for the bio allocation no point doing it again.

So to be rigorously correct, your way would have to be

p = mempool_alloc(bs->bio_pool, gfp_mask);
if (!p && gfp_mask != saved_gfp) {
punt_bios_to_rescuer(bs);
gfp_mask = saved_gfp;
p = mempool_alloc(bs->bio_pool, gfp_mask);
}

And at that point, why duplicate that line of code? It doesn't matter that
much, but IMO a goto retry better labels what's actually going on (it's
something that's not uncommon in the kernel and if I see a retry label
in a function I pretty immediately have an idea of what's going on).

So we could do

retry:
p = mempool_alloc(bs->bio_pool, gfp_mask);
if (!p && gfp_mask != saved_gfp) {
punt_bios_to_rescuer(bs);
gfp_mask = saved_gfp;
goto retry;
}

(side note: not that it really matters here, but gcc will inline the
bvec_alloc_bs() call if it's not duplicated, I've never seen it
consolidate duplicated code and /then/ inline based off that)

This does have the advantage that we're not freeing and reallocating the
bio like Vivek pointed out, but I'm not a huge fan of having the
punting/retry logic in the main code path.

I don't care that much though. I'd prefer not to have the actual
allocations duplicated, but it's starting to feel like bikeshedding to
me.

> > +static void punt_bios_to_rescuer(struct bio_set *bs)
> > +{
> > + struct bio_list punt, nopunt;
> > + struct bio *bio;
> > +
> > + /*
> > + * Don't want to punt all bios on current->bio_list; if there was a bio
> > + * on there for a stacking driver higher up in the stack, processing it
> > + * could require allocating bios from this bio_set, and we don't want to
> > + * do that from our own rescuer.
>
> Hmmm... isn't it more like we "must" process only the bios which are
> from this bio_set to have any kind of forward-progress guarantee? The
> above sounds like it's just something undesirable.

Yeah, that'd be better, I'll change it.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/