Re: bio_chain: proposed solution for bio_alloc failure and large IO simplification

[Jens Axboe]

On Sat, Jun 15 2002, Adam J. Richter wrote:
> Jens Axboe wrote:
> >The I/O path allocations all use GFP_NOIO (or GFP_NOFS), which all have
> >__GFP_WAIT set. So the bio allocations will try normal allocation first,
> >then fall back to the bio pool. If the bio pool is also empty, we will
> >block waiting for entries to be freed there. So there never will be a
> >failure.
> 
> 	I did not realize that allocation with __GFP_WAIT was guaranteed
> to _never_ fail.

See the mempool implementation. Even before bio used mempool, it used
the exact same logic to make sure it never fails. Basically the
heuristic in both cases is:

repeat:
	bio = normal_alloc
	if (bio)
		return bio

	bio = get_from_pool
	if (bio)
		return bio

	if (gfp_mask & __GFP_WAIT)
		start disk i/o
		goto repeat;

> 	Even so, if __GFP_WAIT never fails, then it can deadlock (for
> example, some other device driver has a memory leak).  Under a
> scheme like bio_chain (provided that it is changed not to call a
> memory allocator that can deadlock), the only way you deadlock is
> if there really is deadlock bug in the lower layers that process
> the underlying request.

This whole dead lock debate has been done to death before, I suggest you
find the mempool discussions in the lkml archives from the earlier 2.5
series. Basically we maintain deadlock free allocation although we never
fail allocs by saying that a single bio allocation is short lived (or
at least not held indefinetely). That plus a reserve of XX bios makes
sure that someone will always return a bio to the pool sooner or later
and at least the get_from_pool alloc above will succeed sooner or later
even if vm pressure is ridicilous.

-- 
Jens Axboe