Re: [PATCH 1/2 v2] bcachefs: do not use PF_MEMALLOC_NORECLAIM

[Michal Hocko <mhocko@suse.com>]

On Sun 01-09-24 21:35:30, Kent Overstreet wrote:
[...]
> But I am saying that kmalloc(__GFP_NOFAIL) _should_ fail and return NULL
> in the case of bugs, because that's going to be an improvement w.r.t.
> system robustness, in exactly the same way we don't use BUG_ON() if it's
> something that we can't guarantee won't happen in the wild - we WARN()
> and try to handle the error as best we can.

We have discussed that in a different email thread. And I have to say
that I am not convinced that returning NULL makes a broken code much
better. Why? Because we can expect that broken NOFAIL users will not have a
error checking path. Even valid NOFAIL users will not have one because
they _know_ they do not have a different than retry for ever recovery
path. 

That means that an unexpected NULL return either means OOPS or a subtle
silent error - e.g. memory corruption. The former is a actually a saner
recovery model because the execution is stopped before more harm can be
done. I suspect most of those buggy users will simply OOPS but
systematically checking for latter is a lot of work and needs to be
constantly because code evolves...

I have tried to argue that if allocator cannot or refuse to satisfy
GFP_NOFAIL request because it is trying to use unsupported allocation
mode or size then we should terminate the allocation context. That would
make the API more predictable and therefore safer to use.

This is not what the allocator does today though. Atomic NOFAIL
allocations fail same as kvmalloc requests which are clearly overflows.
Especially the later could become a risk if they are reachable from the
userspace with controlable allocation size.

-- 
Michal Hocko
SUSE Labs